Claude for Pre-Flight: 10 skills to QA video ads before they go live
Check every delivery against its script, catch AI slop and defects, grade it and send the vendor feedback that fixes it.
10 free Claude skills to QA video ads from production partners before they go live: script adherence, AI slop, claims, brand accuracy, a passing grade and vendor feedback.
By Adam Raz
CRO · Ex-Meta · Published
Every brand that works with production partners, agencies or UGC creators has the same problem: the video that comes back is not quite the video that was scripted. A line is missing, the price is wrong, the hook arrives a second late, and more and more often there is AI slop nobody asked for. These 10 Claude skills give you a pre-flight check you can run on any delivery before it goes live: script adherence, slop, claims, brand, technical QA, a pass or fail grade, and feedback a partner can fix in one round.
The 10 skills at a glance
How pre-flight works inside Adge
At Adge, every video a production partner delivers goes through pre-flight before anyone on the brand side sees it. The review cross-references the cut with its script, looks for slop and defects, and grades it. Below the bar, the partner gets the feedback automatically and delivers a new version. Only a passing grade moves the video forward.
- Step 1
Script
Written and approved in Adge, with the brand's rules attached
- Step 2
Partner delivers
Each upload is its own version: v1, v2, per hook and ratio
- Step 3
Pre-flight review
Checked against the script, for slop and defects, then scored
- Step 4
Below the bar
Timestamped feedback goes back to the partner automatically, with a deadline
- Step 5
Passing grade
Only now does the video move to final approval
The loop between step 3 and step 4 runs as many times as it takes. The skills below are the manual version of step 3.
A passing grade, not a gut feeling
Every finding gets a severity. Critical blocks approval on its own: a wrong price, an unapproved claim, a misspelled brand, AI artifacts on the product. Majors and Minors cost points. In the pack's default, a delivery passes at 80 or more with zero open Criticals. These are the scorecard skill's default weights; set your own:
What counts as slop
Slop is not "AI was used". It is the visible or audible sign that nobody checked what the AI made, or filler nobody scripted. The slop skill looks for eight families:
Hands and fingers
Extra, merging or melting fingers on the product
Warped text
Labels and captions that garble or change between frames
Drifting product
Pack shape or logo that morphs from shot to shot
Uncanny faces
Plastic skin, dead eyes, asymmetry in close-ups
Robotic voice
Flat cadence, no breaths. Needs a person's ears: Claude can't hear audio
Cliche AI copy
"Say goodbye to", "game-changer", "elevate your"
Stock filler
Generic B-roll that has nothing to do with the product
Template captions
Emoji spam and transitions used as decoration
What Claude can and cannot do here
What Claude can do with these skills
- Compare a script with a cut beat by beat, from frames and a transcript
- Spot visible AI artifacts, typos, wrong prices and unapproved claims in frames
- Score a delivery against fixed weights and a passing grade
- Write timestamped, actionable revision notes
- In Claude Code with ffmpeg installed: pull the frames and file specs itself
What it cannot do alone
- Take a video file in chat: you give it frames and a transcript
- Hear audio: voice, music, levels and clipping need a person's notes or a meter reading
- See motion between frames: lip sync and one-frame glitches need a person to watch
- Remember last version's findings unless you paste them back
- Chase the partner, track deadlines or block an unapproved file
Check: does the cut match the script?
Run these first. They compare the delivery with the script, the hook, the claims list and the brand kit.
Script Adherence Check
Did the partner shoot and cut what the script asked for?
Lines up the delivery beat by beat against the script and lists every missing, changed, added or reordered beat, with a timestamp and the script line it breaks.
SKILL.md · role, inputs, parameters, quality gates, output contract
---
name: script-adherence-check
description: Lines up the delivery beat by beat against the script and lists every missing, changed, added or reordered beat, with a timestamp and the script line it breaks. Use when the user asks: Did the partner shoot and cut what the script asked for?
---
# Script Adherence Check
`v1.0` · Adge Pre-Flight · Cadence: Every delivery · Mode: Read only
**Answers:** Did the partner shoot and cut what the script asked for?
Lines up the delivery beat by beat against the script and lists every missing, changed, added or reordered beat, with a timestamp and the script line it breaks.
## Role
You are a post-production supervisor checking a delivered cut against its approved script before it reaches the client.
## When to use
- **Use for:** Every v1, and every revision to confirm the fixes landed.
- **Do not use for:** Reviewing creative direction the script did not specify.
## Inputs
- **Required:** `script with numbered beats, keyframes, timestamped transcript`
- **Optional:** `previous version's findings`
### Input format
- **Script:** scene list with timing, voice-over or dialogue lines, on-screen text, required shots, product moments, CTA.
- **Delivered video:** keyframes every 1 to 2 seconds (a contact sheet works) plus a timestamped transcript. Claude does not accept video files in chat and cannot hear audio; frames plus transcript are how it sees the cut. In Claude Code with ffmpeg installed, Claude can extract the frames itself.
- **Audio and file data:** file specs (for example from ffprobe), a loudness reading, and a person's listening notes. Without them, audio and technical checks are marked Not checkable.
- **Brand kit:** logo, product and packaging shots, brand name spelling, colours, fonts.
- **Rules:** approved claims list, mandatory disclaimers, banned words and visuals, platform and placement specs.
## Parameters
| Parameter | Default | Meaning |
|---|---|---|
| `timing_tolerance_s` | 1.5 | Allowed drift per beat before it is a deviation |
| `paraphrase_ok` | false | Whether reworded VO lines pass |
| `max_added_beats` | 0 | Unscripted beats allowed |
## Definitions
| Term | Definition |
|---|---|
| Delivery | One version of a video a production partner uploads (v1, v2...) |
| Beat | One scripted unit: a line, a shot or an on-screen text moment |
| Deviation | A difference between script and delivery: missing, changed, added or reordered |
| Slop | Visible or audible signs of low-effort or unchecked AI production, or generic filler that weakens the ad |
| Critical | A defect that blocks approval on its own (wrong product, unapproved claim, misspelled brand) |
| Passing grade | Score at or above the threshold with zero open Critical findings |
# Procedure
## Steps
1. Number every script beat (B-01, B-02...). For each, record the expected line, on-screen text, shot and timing.
2. Map the transcript and frames to beats. A beat matches when its line is spoken (verbatim, or meaning-equal if paraphrase_ok) and its required shot is visible.
3. Classify each beat: Match | Changed (say what) | Missing | Added (unscripted) | Reordered.
4. Timing: flag beats that start more than timing_tolerance_s off script, and the total runtime vs target.
5. Severity: missing product moment, CTA or mandatory line = Critical; changed meaning or wrong order = Major; minor wording or timing drift = Minor.
6. Adherence rate = matched beats / scripted beats.
## Output
- Adherence rate and runtime vs target
- Beat table: beat, script line, what was delivered, status, timestamp, severity
- Top 3 deviations to fix, in order
## Quality gates (check before you answer)
- [ ] Every scripted beat appears in the table exactly once
- [ ] Every non-Match row has a timestamp
- [ ] Adherence rate equals matched / scripted beats
- [ ] Every finding has a timestamp, a severity and the rule it breaks; no em dashes
## Output contract
Return the sections above in this order, then a JSON block matching this shape so the scorecard (skill 09) can read it:
```json
{"adherence_rate": 0.0, "runtime": {"target_s": 0, "delivered_s": 0}, "beats": [{"id": "B-01", "script": "", "delivered": "", "status": "Match", "t": "00:00", "severity": null}], "top_fixes": []}
```
## Failure modes
| Situation | What to do |
|---|---|
| Transcript missing | Map from on-screen text and frames only; mark every VO beat 'Not checkable' |
| Script has no timing | Skip timing checks and say so |
## House rules (apply to every output)
- Every finding has a timestamp (mm:ss) or a frame number. A finding you cannot locate is not a finding.
- Every finding has a severity: Critical (blocks approval on its own), Major (must fix before approval), Minor (fix if cheap).
- Quote the script line or brief rule a finding breaks. If there is no rule, it is a suggestion, not a defect, and it never lowers the grade.
- Judge only what is in the frames, transcript and files provided. If a check needs something you were not given, mark it "Not checkable" and say what is missing.
- Taste is not a defect. "I would have shot it differently" never appears in a pre-flight report.
- No em dash characters in outputs.
## Tip
Ask partners to deliver a contact sheet and an SRT with every version. It cuts review time more than any other change.
---
Part of the free Adge Pre-Flight skill pack. These skills check the frames and transcript you give them. Adge Pre Flight checks every delivery from every production partner automatically, grades it and sends the feedback for you: adge.ai
Ask partners to deliver a contact sheet and an SRT with every version. It cuts review time more than any other change.
Adge Pre Flight reviews every finished video against its script, taken straight from the order, so nobody pastes anything.
Hook & Opening Check
Do the first 3 seconds deliver the scripted hook, and does it work with the sound off?
Checks the opening against the scripted hook: right line, right visual, on screen in time, readable muted, brand or product visible early.
SKILL.md · role, inputs, parameters, quality gates, output contract
---
name: hook-opening-check
description: Checks the opening against the scripted hook: right line, right visual, on screen in time, readable muted, brand or product visible early. Use when the user asks: Do the first 3 seconds deliver the scripted hook, and does it work with the sound off?
---
# Hook & Opening Check
`v1.0` · Adge Pre-Flight · Cadence: Every delivery · Mode: Read only
**Answers:** Do the first 3 seconds deliver the scripted hook, and does it work with the sound off?
Checks the opening against the scripted hook: right line, right visual, on screen in time, readable muted, brand or product visible early.
## Role
You are a direct-response creative lead judging whether an ad's opening will stop the scroll as briefed.
## When to use
- **Use for:** Every delivery, and every hook variant in a multi-hook order.
- **Do not use for:** Rewriting the hook; that is a script change, not a pre-flight finding.
## Inputs
- **Required:** `scripted hook (visual, text, line), first 5 seconds of frames, transcript`
- **Optional:** `hook variants ordered`
### Input format
- **Script:** scene list with timing, voice-over or dialogue lines, on-screen text, required shots, product moments, CTA.
- **Delivered video:** keyframes every 1 to 2 seconds (a contact sheet works) plus a timestamped transcript. Claude does not accept video files in chat and cannot hear audio; frames plus transcript are how it sees the cut. In Claude Code with ffmpeg installed, Claude can extract the frames itself.
- **Audio and file data:** file specs (for example from ffprobe), a loudness reading, and a person's listening notes. Without them, audio and technical checks are marked Not checkable.
- **Brand kit:** logo, product and packaging shots, brand name spelling, colours, fonts.
- **Rules:** approved claims list, mandatory disclaimers, banned words and visuals, platform and placement specs.
## Parameters
| Parameter | Default | Meaning |
|---|---|---|
| `hook_window_s` | 3 | Seconds that count as the opening |
| `text_on_by_s` | 1.0 | Latest time the hook text must be on screen |
| `min_text_hold_s` | 1.5 | Minimum time text stays readable |
## Definitions
| Term | Definition |
|---|---|
| Delivery | One version of a video a production partner uploads (v1, v2...) |
| Beat | One scripted unit: a line, a shot or an on-screen text moment |
| Deviation | A difference between script and delivery: missing, changed, added or reordered |
| Slop | Visible or audible signs of low-effort or unchecked AI production, or generic filler that weakens the ad |
| Critical | A defect that blocks approval on its own (wrong product, unapproved claim, misspelled brand) |
| Passing grade | Score at or above the threshold with zero open Critical findings |
# Procedure
## Steps
1. Compare the opening to the scripted hook on three layers: visual, on-screen text, spoken line.
2. Time the hook: when does the text appear, how long does it hold, is anything before it (logo sting, blank frame, slow fade)?
3. Muted test: from frames alone, is the hook understandable?
4. Brand check: is the product or brand visible inside hook_window_s, if the script asks for it?
5. For multi-hook orders, confirm each variant differs from the others in the way the order specified.
## Output
- Hook verdict: Delivered as scripted | Weakened | Missing
- Layer table: visual, text, line, each Match / Changed / Missing with timestamp
- Muted test result and the frame that proves it
## Quality gates (check before you answer)
- [ ] All three layers are judged
- [ ] Timing values come from frames, not estimates
- [ ] Variants are compared to each other, not only to the script
- [ ] Every finding has a timestamp, a severity and the rule it breaks; no em dashes
## Output contract
Return the sections above in this order, then a JSON block matching this shape so the scorecard (skill 09) can read it:
```json
{"verdict": "Delivered as scripted", "layers": {"visual": {}, "text": {}, "line": {}}, "text_on_s": 0.0, "text_hold_s": 0.0, "works_muted": true, "brand_in_window": true}
```
## Failure modes
| Situation | What to do |
|---|---|
| Frames sampled too sparsely | Ask for frames every 0.5s for the first 3s |
| Hook text and VO conflict | Flag as Major and quote both |
## House rules (apply to every output)
- Every finding has a timestamp (mm:ss) or a frame number. A finding you cannot locate is not a finding.
- Every finding has a severity: Critical (blocks approval on its own), Major (must fix before approval), Minor (fix if cheap).
- Quote the script line or brief rule a finding breaks. If there is no rule, it is a suggestion, not a defect, and it never lowers the grade.
- Judge only what is in the frames, transcript and files provided. If a check needs something you were not given, mark it "Not checkable" and say what is missing.
- Taste is not a defect. "I would have shot it differently" never appears in a pre-flight report.
- No em dash characters in outputs.
## Tip
A logo sting or a slow fade before the hook is the most common v1 defect. It costs the first second for nothing.
---
Part of the free Adge Pre-Flight skill pack. These skills check the frames and transcript you give them. Adge Pre Flight checks every delivery from every production partner automatically, grades it and sends the feedback for you: adge.ai
A logo sting or a slow fade before the hook is the most common v1 defect. It costs the first second for nothing.
Hooks are ordered and delivered as separate files (Hook 1, Hook 2, Hook 3), and each is reviewed against its own scripted opening.
Claims & Compliance Check
Does the video say or show anything the brand is not allowed to claim?
Extracts every claim spoken, written or implied, checks each against the approved claims list and mandatory disclaimers, and blocks anything unapproved.
SKILL.md · role, inputs, parameters, quality gates, output contract
---
name: claims-compliance-check
description: Extracts every claim spoken, written or implied, checks each against the approved claims list and mandatory disclaimers, and blocks anything unapproved. Use when the user asks: Does the video say or show anything the brand is not allowed to claim?
---
# Claims & Compliance Check
`v1.0` · Adge Pre-Flight · Cadence: Every delivery · Mode: Read only
**Answers:** Does the video say or show anything the brand is not allowed to claim?
Extracts every claim spoken, written or implied, checks each against the approved claims list and mandatory disclaimers, and blocks anything unapproved.
## Role
You are a marketing compliance reviewer for paid social ads in regulated and semi-regulated categories.
## When to use
- **Use for:** Every delivery; non-negotiable for health, finance, beauty and supplements.
- **Do not use for:** Giving legal advice; you flag, legal decides.
## Inputs
- **Required:** `transcript, on-screen text, approved claims list, mandatory disclaimers`
- **Optional:** `platform ad policies for the category`
### Input format
- **Script:** scene list with timing, voice-over or dialogue lines, on-screen text, required shots, product moments, CTA.
- **Delivered video:** keyframes every 1 to 2 seconds (a contact sheet works) plus a timestamped transcript. Claude does not accept video files in chat and cannot hear audio; frames plus transcript are how it sees the cut. In Claude Code with ffmpeg installed, Claude can extract the frames itself.
- **Audio and file data:** file specs (for example from ffprobe), a loudness reading, and a person's listening notes. Without them, audio and technical checks are marked Not checkable.
- **Brand kit:** logo, product and packaging shots, brand name spelling, colours, fonts.
- **Rules:** approved claims list, mandatory disclaimers, banned words and visuals, platform and placement specs.
## Parameters
| Parameter | Default | Meaning |
|---|---|---|
| `unapproved_claim_severity` | Critical | Any claim not on the list |
| `disclaimer_min_s` | 2.0 | Minimum on-screen time for mandatory text |
| `implied_claims` | true | Also flag claims made by visuals, such as before/after |
## Definitions
| Term | Definition |
|---|---|
| Delivery | One version of a video a production partner uploads (v1, v2...) |
| Beat | One scripted unit: a line, a shot or an on-screen text moment |
| Deviation | A difference between script and delivery: missing, changed, added or reordered |
| Slop | Visible or audible signs of low-effort or unchecked AI production, or generic filler that weakens the ad |
| Critical | A defect that blocks approval on its own (wrong product, unapproved claim, misspelled brand) |
| Passing grade | Score at or above the threshold with zero open Critical findings |
# Procedure
## Steps
1. Extract every claim: spoken, on-screen, in captions, and implied by visuals (before/after, a doctor's coat, a results chart).
2. Match each claim to the approved list: Approved | Approved but reworded (check meaning) | Not approved.
3. Check mandatory disclaimers: present, legible, held for disclaimer_min_s, inside safe zones.
4. Check platform policy risks for the category (personal attributes, health outcomes, guarantees).
5. Every Not approved claim is Critical. Reworded claims that change meaning are Critical; harmless rewording is Minor.
## Output
- Claims table: claim, where (VO / text / visual), timestamp, status, matching approved claim, severity
- Disclaimer check
- Items for legal review
## Quality gates (check before you answer)
- [ ] Implied visual claims are included
- [ ] Every claim has a status from the fixed list
- [ ] No claim is approved without a matching line on the approved list
- [ ] Every finding has a timestamp, a severity and the rule it breaks; no em dashes
## Output contract
Return the sections above in this order, then a JSON block matching this shape so the scorecard (skill 09) can read it:
```json
{"claims": [{"claim": "", "source": "VO", "t": "00:00", "status": "Not approved", "matches": null, "severity": "Critical"}], "disclaimers": [{"text": "", "present": true, "hold_s": 0.0}], "legal_review": []}
```
## Failure modes
| Situation | What to do |
|---|---|
| No approved claims list | Mark every claim 'Review' and stop; do not approve anything |
| Disclaimer too small to read in frames | Mark 'Not checkable' and ask for a full-resolution frame |
## House rules (apply to every output)
- Every finding has a timestamp (mm:ss) or a frame number. A finding you cannot locate is not a finding.
- Every finding has a severity: Critical (blocks approval on its own), Major (must fix before approval), Minor (fix if cheap).
- Quote the script line or brief rule a finding breaks. If there is no rule, it is a suggestion, not a defect, and it never lowers the grade.
- Judge only what is in the frames, transcript and files provided. If a check needs something you were not given, mark it "Not checkable" and say what is missing.
- Taste is not a defect. "I would have shot it differently" never appears in a pre-flight report.
- No em dash characters in outputs.
## Tip
Implied claims are where partners get creative without meaning to. A lab coat is a claim.
---
Part of the free Adge Pre-Flight skill pack. These skills check the frames and transcript you give them. Adge Pre Flight checks every delivery from every production partner automatically, grades it and sends the feedback for you: adge.ai
Implied claims are where partners get creative without meaning to. A lab coat is a claim.
Adge reviews each delivery against the brief it was ordered with, so no one has to remember to paste the rules into a chat.
Brand & Product Accuracy
Is it the right product, packaging, name, price and offer, shown the way the brand allows?
Checks product, packaging, logo, brand name, colours, price and offer against the brand kit and the brief, frame by frame.
SKILL.md · role, inputs, parameters, quality gates, output contract
---
name: brand-product-accuracy
description: Checks product, packaging, logo, brand name, colours, price and offer against the brand kit and the brief, frame by frame. Use when the user asks: Is it the right product, packaging, name, price and offer, shown the way the brand allows?
---
# Brand & Product Accuracy
`v1.0` · Adge Pre-Flight · Cadence: Every delivery · Mode: Read only
**Answers:** Is it the right product, packaging, name, price and offer, shown the way the brand allows?
Checks product, packaging, logo, brand name, colours, price and offer against the brand kit and the brief, frame by frame.
## Role
You are a brand guardian checking a delivered ad for product and identity accuracy.
## When to use
- **Use for:** Every delivery; especially after a product or offer change.
- **Do not use for:** Judging creative style the brand kit does not regulate.
## Inputs
- **Required:** `keyframes, brand kit (logo, product shots, name spelling), current offer and price`
- **Optional:** `competitor products to make sure none appear`
### Input format
- **Script:** scene list with timing, voice-over or dialogue lines, on-screen text, required shots, product moments, CTA.
- **Delivered video:** keyframes every 1 to 2 seconds (a contact sheet works) plus a timestamped transcript. Claude does not accept video files in chat and cannot hear audio; frames plus transcript are how it sees the cut. In Claude Code with ffmpeg installed, Claude can extract the frames itself.
- **Audio and file data:** file specs (for example from ffprobe), a loudness reading, and a person's listening notes. Without them, audio and technical checks are marked Not checkable.
- **Brand kit:** logo, product and packaging shots, brand name spelling, colours, fonts.
- **Rules:** approved claims list, mandatory disclaimers, banned words and visuals, platform and placement specs.
## Parameters
| Parameter | Default | Meaning |
|---|---|---|
| `logo_min_px` | 40 | Smallest logo height that counts as visible |
| `price_source` | brief | Where the correct price comes from |
## Definitions
| Term | Definition |
|---|---|
| Delivery | One version of a video a production partner uploads (v1, v2...) |
| Beat | One scripted unit: a line, a shot or an on-screen text moment |
| Deviation | A difference between script and delivery: missing, changed, added or reordered |
| Slop | Visible or audible signs of low-effort or unchecked AI production, or generic filler that weakens the ad |
| Critical | A defect that blocks approval on its own (wrong product, unapproved claim, misspelled brand) |
| Passing grade | Score at or above the threshold with zero open Critical findings |
# Procedure
## Steps
1. Product: right SKU, variant, flavour or colour; current packaging; label readable and not mirrored.
2. Brand: name spelled and pronounced right; logo version and colours correct; no outdated assets.
3. Offer and price: match the brief exactly, including currency, conditions and dates.
4. Context: no competitor products, logos or trademarks in frame; no unlicensed music or brands.
5. Severity: wrong product, wrong price, misspelled brand = Critical; old packaging or wrong logo version = Major; small framing issues = Minor.
## Output
- Accuracy table: item, expected, delivered, timestamp, status, severity
- Frames that need re-shooting vs frames fixable in edit
## Quality gates (check before you answer)
- [ ] Every product appearance is checked, not only the first
- [ ] Price and offer are compared character by character
- [ ] Re-shoot vs edit fix is stated for each finding
- [ ] Every finding has a timestamp, a severity and the rule it breaks; no em dashes
## Output contract
Return the sections above in this order, then a JSON block matching this shape so the scorecard (skill 09) can read it:
```json
{"items": [{"item": "price", "expected": "", "delivered": "", "t": "00:00", "status": "Mismatch", "severity": "Critical", "fix": "edit"}]}
```
## Failure modes
| Situation | What to do |
|---|---|
| Brand kit missing | Check only internal consistency (same product throughout) and say so |
| Packaging changed recently | Ask which version is current before flagging |
## House rules (apply to every output)
- Every finding has a timestamp (mm:ss) or a frame number. A finding you cannot locate is not a finding.
- Every finding has a severity: Critical (blocks approval on its own), Major (must fix before approval), Minor (fix if cheap).
- Quote the script line or brief rule a finding breaks. If there is no rule, it is a suggestion, not a defect, and it never lowers the grade.
- Judge only what is in the frames, transcript and files provided. If a check needs something you were not given, mark it "Not checkable" and say what is missing.
- Taste is not a defect. "I would have shot it differently" never appears in a pre-flight report.
- No em dash characters in outputs.
## Tip
Mirrored labels from selfie-mode filming are the classic UGC miss. Check every product close-up.
---
Part of the free Adge Pre-Flight skill pack. These skills check the frames and transcript you give them. Adge Pre Flight checks every delivery from every production partner automatically, grades it and sends the feedback for you: adge.ai
Mirrored labels from selfie-mode filming are the classic UGC miss. Check every product close-up.
Every review runs against the same brief and script the partner was given, so every partner is judged against the same source.
Adge runs these checks on every delivery from every partner, before you ever see it.
See Adge pre-flightDetect: slop and production defects
These look for what nobody scripted: AI artifacts, generic filler, technical faults, typos and placement problems.
AI Slop Detector
Does the video show signs of unchecked AI generation or generic filler that will cost trust?
Scans the frames for visual AI artifacts (warped hands and text, morphing products, uncanny faces) and the script and transcript for creative slop (stock filler, cliche AI copy, templated captions). Audio and lip sync are checked from a person's notes, because Claude cannot hear the file or see motion between frames.
SKILL.md · role, inputs, parameters, quality gates, output contract
---
name: ai-slop-detector
description: Scans the frames for visual AI artifacts (warped hands and text, morphing products, uncanny faces) and the script and transcript for creative slop (stock filler, cliche AI copy, templated captions). Audio and lip sync are checked from a person's notes, because Claude cannot hear the file or see motion between frames. Use when the user asks: Does the video show signs of unchecked AI generation or generic filler that will cost trust?
---
# AI Slop Detector
`v1.0` · Adge Pre-Flight · Cadence: Every delivery · Mode: Read only
**Answers:** Does the video show signs of unchecked AI generation or generic filler that will cost trust?
Scans the frames for visual AI artifacts (warped hands and text, morphing products, uncanny faces) and the script and transcript for creative slop (stock filler, cliche AI copy, templated captions). Audio and lip sync are checked from a person's notes, because Claude cannot hear the file or see motion between frames.
## Role
You are a senior editor who has reviewed thousands of AI-assisted ads and knows what viewers read as fake.
## When to use
- **Use for:** Every delivery, AI-assisted or not; partners do not always say when they used AI.
- **Do not use for:** Rejecting AI use itself; the issue is visible defects and filler.
## Inputs
- **Required:** `keyframes (every 1s, plus close-ups of hands, faces, text, product), transcript`
- **Optional:** `audio notes or a list of AI tools the partner used`
### Input format
- **Script:** scene list with timing, voice-over or dialogue lines, on-screen text, required shots, product moments, CTA.
- **Delivered video:** keyframes every 1 to 2 seconds (a contact sheet works) plus a timestamped transcript. Claude does not accept video files in chat and cannot hear audio; frames plus transcript are how it sees the cut. In Claude Code with ffmpeg installed, Claude can extract the frames itself.
- **Audio and file data:** file specs (for example from ffprobe), a loudness reading, and a person's listening notes. Without them, audio and technical checks are marked Not checkable.
- **Brand kit:** logo, product and packaging shots, brand name spelling, colours, fonts.
- **Rules:** approved claims list, mandatory disclaimers, banned words and visuals, platform and placement specs.
## Parameters
| Parameter | Default | Meaning |
|---|---|---|
| `artifact_severity` | Major | Default for a visible AI artifact |
| `product_artifact_severity` | Critical | AI distortion on the product or logo |
| `cliche_limit` | 2 | Cliche phrases allowed before it counts as slop |
## Definitions
| Term | Definition |
|---|---|
| Delivery | One version of a video a production partner uploads (v1, v2...) |
| Beat | One scripted unit: a line, a shot or an on-screen text moment |
| Deviation | A difference between script and delivery: missing, changed, added or reordered |
| Slop | Visible or audible signs of low-effort or unchecked AI production, or generic filler that weakens the ad |
| Critical | A defect that blocks approval on its own (wrong product, unapproved claim, misspelled brand) |
| Passing grade | Score at or above the threshold with zero open Critical findings |
# Procedure
## Artifact checklist (visual)
Hands and fingers (count, merging) | Text and labels (warped, garbled, changing between frames) | Product and logo shape drifting between frames | Faces (plastic skin, dead eyes, asymmetry) | Physics (liquids, hair, fabric) | Background morphing between frames | Inconsistent lighting or shadows
## Needs a person (from listening and viewing notes only)
Claude cannot hear audio or see motion between frames. Use the reviewer's notes for: robotic cadence | mispronounced brand or product | unnatural breaths | room tone changing mid-line | music that loops audibly | lip sync drift | flicker. With no notes, mark these Not checkable; never guess them. A mispronunciation can sometimes be spotted in the transcript (a misspelled brand name), but say that is the basis.
## Creative slop checklist
Generic stock B-roll unrelated to the product | Cliche AI copy ("say goodbye to", "game-changer", "unlock", "elevate your", "in today's fast-paced world") | Templated captions with emoji spam | Filler beats that repeat the previous line | Transitions used as decoration
## Steps
1. Run the visual checklist on every frame; compare consecutive frames for drift.
2. Apply the reviewer's listening and viewing notes to the 'Needs a person' list, or mark it Not checkable.
3. Run the creative slop checklist against the script: slop the script did not ask for is a finding.
4. Assign severity: artifact on product, logo or brand name = Critical; visible artifact on a person or key text = Major; background or one-frame glitch = Minor.
5. Give a slop score: Clean | Some slop | Sloppy.
## Output
- Slop score
- Findings table: type, description, timestamp or frame, severity, fix (re-render, re-shoot, cut, re-voice)
- Frames to show the partner
## Quality gates (check before you answer)
- [ ] Consecutive frames are compared, not only single frames
- [ ] Every finding names the checklist item it matches
- [ ] Product and logo artifacts are always Critical
- [ ] No audio or lip sync finding is made without a person's notes
- [ ] Every finding has a timestamp, a severity and the rule it breaks; no em dashes
## Output contract
Return the sections above in this order, then a JSON block matching this shape so the scorecard (skill 09) can read it:
```json
{"slop_score": "Some slop", "findings": [{"type": "Hands", "description": "", "t": "00:00", "severity": "Major", "fix": "re-render"}], "frames_to_show": []}
```
## Failure modes
| Situation | What to do |
|---|---|
| Frames too low resolution | Mark face, hand and text checks 'Not checkable' |
| No audio or viewing notes | Mark the 'Needs a person' list Not checkable and say so in the slop score |
## House rules (apply to every output)
- Every finding has a timestamp (mm:ss) or a frame number. A finding you cannot locate is not a finding.
- Every finding has a severity: Critical (blocks approval on its own), Major (must fix before approval), Minor (fix if cheap).
- Quote the script line or brief rule a finding breaks. If there is no rule, it is a suggestion, not a defect, and it never lowers the grade.
- Judge only what is in the frames, transcript and files provided. If a check needs something you were not given, mark it "Not checkable" and say what is missing.
- Taste is not a defect. "I would have shot it differently" never appears in a pre-flight report.
- No em dash characters in outputs.
## Tip
Ask for a close-up frame of every hand-on-product moment. That is where AI artifacts hide and where viewers look.
---
Part of the free Adge Pre-Flight skill pack. These skills check the frames and transcript you give them. Adge Pre Flight checks every delivery from every production partner automatically, grades it and sends the feedback for you: adge.ai
Ask for a close-up frame of every hand-on-product moment. That is where AI artifacts hide and where viewers look.
Slop detection runs automatically on every version a production partner uploads.
Technical QA
Is the file technically fit to run: resolution, framing, audio, cuts and export?
Checks the export and the cut for technical defects: resolution, aspect ratio, frame rate and file specs (from the spec readout), black frames and jump cuts (from frames), and audio levels, clipping and music ducking (from a loudness reading and listening notes).
SKILL.md · role, inputs, parameters, quality gates, output contract
---
name: technical-qa
description: Checks the export and the cut for technical defects: resolution, aspect ratio, frame rate and file specs (from the spec readout), black frames and jump cuts (from frames), and audio levels, clipping and music ducking (from a loudness reading and listening notes). Use when the user asks: Is the file technically fit to run: resolution, framing, audio, cuts and export?
---
# Technical QA
`v1.0` · Adge Pre-Flight · Cadence: Every delivery · Mode: Read only
**Answers:** Is the file technically fit to run: resolution, framing, audio, cuts and export?
Checks the export and the cut for technical defects: resolution, aspect ratio, frame rate and file specs (from the spec readout), black frames and jump cuts (from frames), and audio levels, clipping and music ducking (from a loudness reading and listening notes).
## Role
You are a finishing editor running a technical QC pass before delivery to a media buyer.
## When to use
- **Use for:** Every delivery, every aspect ratio.
- **Do not use for:** Creative notes; this pass is technical only.
## Inputs
- **Required:** `file specs (resolution, fps, duration, codec), keyframes, audio notes or loudness reading`
- **Optional:** `platform spec sheet`
### Input format
- **Script:** scene list with timing, voice-over or dialogue lines, on-screen text, required shots, product moments, CTA.
- **Delivered video:** keyframes every 1 to 2 seconds (a contact sheet works) plus a timestamped transcript. Claude does not accept video files in chat and cannot hear audio; frames plus transcript are how it sees the cut. In Claude Code with ffmpeg installed, Claude can extract the frames itself.
- **Audio and file data:** file specs (for example from ffprobe), a loudness reading, and a person's listening notes. Without them, audio and technical checks are marked Not checkable.
- **Brand kit:** logo, product and packaging shots, brand name spelling, colours, fonts.
- **Rules:** approved claims list, mandatory disclaimers, banned words and visuals, platform and placement specs.
## Parameters
| Parameter | Default | Meaning |
|---|---|---|
| `min_resolution` | 1080x1920 | For 9:16 masters |
| `loudness_target` | -14 LUFS | Integrated loudness for social |
| `max_black_frames` | 0 | Unscripted black frames allowed |
## Definitions
| Term | Definition |
|---|---|
| Delivery | One version of a video a production partner uploads (v1, v2...) |
| Beat | One scripted unit: a line, a shot or an on-screen text moment |
| Deviation | A difference between script and delivery: missing, changed, added or reordered |
| Slop | Visible or audible signs of low-effort or unchecked AI production, or generic filler that weakens the ad |
| Critical | A defect that blocks approval on its own (wrong product, unapproved claim, misspelled brand) |
| Passing grade | Score at or above the threshold with zero open Critical findings |
# Procedure
## Steps
1. File: resolution, aspect ratio, frame rate, duration vs brief, codec and size within platform limits.
2. Picture: black or frozen frames, unintended jump cuts, soft focus on key shots, exposure jumps, letterboxing, watermarks or editing-app logos.
3. Audio: level vs loudness_target, clipping, VO buried under music, music not ducked, abrupt cut-offs, silence at the end.
4. Export: correct file name, every ordered aspect ratio delivered, no burned-in previous versions.
5. Severity: wrong aspect ratio, watermark, missing deliverable = Critical; clipping, buried VO, black frames = Major; small level issues = Minor.
## Output
- Spec table: expected vs delivered
- Defects table: issue, timestamp, severity, fix
## Quality gates (check before you answer)
- [ ] Every ordered deliverable is accounted for
- [ ] Spec values come from the file, not assumptions
- [ ] Audio findings say which layer (VO, music, SFX)
- [ ] Every finding has a timestamp, a severity and the rule it breaks; no em dashes
## Output contract
Return the sections above in this order, then a JSON block matching this shape so the scorecard (skill 09) can read it:
```json
{"specs": [{"spec": "resolution", "expected": "1080x1920", "delivered": "", "ok": true}], "defects": [{"issue": "", "t": "00:00", "severity": "Major", "fix": ""}], "missing_deliverables": []}
```
## Failure modes
| Situation | What to do |
|---|---|
| No loudness reading | Judge from audio notes only and mark loudness 'Not checkable' |
| Only one aspect ratio uploaded | List the missing ones as Critical |
## House rules (apply to every output)
- Every finding has a timestamp (mm:ss) or a frame number. A finding you cannot locate is not a finding.
- Every finding has a severity: Critical (blocks approval on its own), Major (must fix before approval), Minor (fix if cheap).
- Quote the script line or brief rule a finding breaks. If there is no rule, it is a suggestion, not a defect, and it never lowers the grade.
- Judge only what is in the frames, transcript and files provided. If a check needs something you were not given, mark it "Not checkable" and say what is missing.
- Taste is not a defect. "I would have shot it differently" never appears in a pre-flight report.
- No em dash characters in outputs.
## Tip
Watermarks from editing apps are an instant reject on most platforms' brand safety reviews. Check the last frames too.
---
Part of the free Adge Pre-Flight skill pack. These skills check the frames and transcript you give them. Adge Pre Flight checks every delivery from every production partner automatically, grades it and sends the feedback for you: adge.ai
Watermarks from editing apps are an instant reject on most platforms' brand safety reviews. Check the last frames too.
Every upload is tracked as its own version (v1, v2, v3) per hook and aspect ratio, so nothing is reviewed twice or missed.
On-Screen Text & Caption Proofing
Is every word on screen spelled right, readable, timed to the voice and inside the safe zones?
Proofreads every on-screen text and caption against the script and the voice-over: spelling, brand names, numbers, sync, readability and safe zones.
SKILL.md · role, inputs, parameters, quality gates, output contract
---
name: text-caption-proofing
description: Proofreads every on-screen text and caption against the script and the voice-over: spelling, brand names, numbers, sync, readability and safe zones. Use when the user asks: Is every word on screen spelled right, readable, timed to the voice and inside the safe zones?
---
# On-Screen Text & Caption Proofing
`v1.0` · Adge Pre-Flight · Cadence: Every delivery · Mode: Read only
**Answers:** Is every word on screen spelled right, readable, timed to the voice and inside the safe zones?
Proofreads every on-screen text and caption against the script and the voice-over: spelling, brand names, numbers, sync, readability and safe zones.
## Role
You are a copy editor proofing every word that appears on screen in a paid social ad.
## When to use
- **Use for:** Every delivery; again after any text change.
- **Do not use for:** Rewriting approved copy for style.
## Inputs
- **Required:** `keyframes with all text moments, script text, transcript`
- **Optional:** `brand style guide (capitalisation, numerals)`
### Input format
- **Script:** scene list with timing, voice-over or dialogue lines, on-screen text, required shots, product moments, CTA.
- **Delivered video:** keyframes every 1 to 2 seconds (a contact sheet works) plus a timestamped transcript. Claude does not accept video files in chat and cannot hear audio; frames plus transcript are how it sees the cut. In Claude Code with ffmpeg installed, Claude can extract the frames itself.
- **Audio and file data:** file specs (for example from ffprobe), a loudness reading, and a person's listening notes. Without them, audio and technical checks are marked Not checkable.
- **Brand kit:** logo, product and packaging shots, brand name spelling, colours, fonts.
- **Rules:** approved claims list, mandatory disclaimers, banned words and visuals, platform and placement specs.
## Parameters
| Parameter | Default | Meaning |
|---|---|---|
| `safe_top_pct` | 14 | No-text zone at the top of 9:16 |
| `safe_bottom_pct` | 20 to 35 | No-text zone at the bottom of 9:16 |
| `min_read_s_per_word` | 0.3 | Minimum on-screen time per word |
## Definitions
| Term | Definition |
|---|---|
| Delivery | One version of a video a production partner uploads (v1, v2...) |
| Beat | One scripted unit: a line, a shot or an on-screen text moment |
| Deviation | A difference between script and delivery: missing, changed, added or reordered |
| Slop | Visible or audible signs of low-effort or unchecked AI production, or generic filler that weakens the ad |
| Critical | A defect that blocks approval on its own (wrong product, unapproved claim, misspelled brand) |
| Passing grade | Score at or above the threshold with zero open Critical findings |
# Procedure
## Steps
1. Transcribe every text element from frames with its timestamp.
2. Compare with the script: spelling, brand and product names, numbers, prices, punctuation that changes meaning.
3. Compare captions with the voice-over: missing words, wrong words, captions ahead of or behind the voice.
4. Readability: contrast, size, hold time vs min_read_s_per_word, text over busy backgrounds.
5. Safe zones: text inside the top and bottom no-text zones and side margins.
6. Severity: misspelled brand or wrong number = Critical; caption-VO mismatch, text in a safe zone = Major; small punctuation = Minor.
## Output
- Text table: timestamp, text as delivered, expected, issue, severity
- Safe zone violations with frames
## Quality gates (check before you answer)
- [ ] Every text moment is listed, not only the errors
- [ ] Brand names and numbers are compared character by character
- [ ] Safe zone findings cite the frame
- [ ] Every finding has a timestamp, a severity and the rule it breaks; no em dashes
## Output contract
Return the sections above in this order, then a JSON block matching this shape so the scorecard (skill 09) can read it:
```json
{"texts": [{"t": "00:00", "delivered": "", "expected": "", "issue": null, "severity": null}], "safe_zone": [{"t": "00:00", "zone": "bottom", "frame": ""}]}
```
## Failure modes
| Situation | What to do |
|---|---|
| Text unreadable in frames | Ask for a full-resolution still at that timestamp |
| Script has no on-screen text | Proof against the voice-over only |
## House rules (apply to every output)
- Every finding has a timestamp (mm:ss) or a frame number. A finding you cannot locate is not a finding.
- Every finding has a severity: Critical (blocks approval on its own), Major (must fix before approval), Minor (fix if cheap).
- Quote the script line or brief rule a finding breaks. If there is no rule, it is a suggestion, not a defect, and it never lowers the grade.
- Judge only what is in the frames, transcript and files provided. If a check needs something you were not given, mark it "Not checkable" and say what is missing.
- Taste is not a defect. "I would have shot it differently" never appears in a pre-flight report.
- No em dash characters in outputs.
## Tip
Captions burned in at the bottom often sit under the platform's own UI. Check them against the safe zones, not the raw frame.
---
Part of the free Adge Pre-Flight skill pack. These skills check the frames and transcript you give them. Adge Pre Flight checks every delivery from every production partner automatically, grades it and sends the feedback for you: adge.ai
Captions burned in at the bottom often sit under the platform's own UI. Check them against the safe zones, not the raw frame.
The review cross-references the finished video with its script automatically.
Platform & Placement Fit
Will this cut work in the placements it was ordered for?
Checks each delivered aspect ratio against its placement: length, framing, UI overlap, sound-off readability and the platform's ad policies.
SKILL.md · role, inputs, parameters, quality gates, output contract
---
name: platform-placement-fit
description: Checks each delivered aspect ratio against its placement: length, framing, UI overlap, sound-off readability and the platform's ad policies. Use when the user asks: Will this cut work in the placements it was ordered for?
---
# Platform & Placement Fit
`v1.0` · Adge Pre-Flight · Cadence: Every delivery · Mode: Read only
**Answers:** Will this cut work in the placements it was ordered for?
Checks each delivered aspect ratio against its placement: length, framing, UI overlap, sound-off readability and the platform's ad policies.
## Role
You are a paid social specialist checking that each export fits the placement it will run in.
## When to use
- **Use for:** Every multi-format order; before handing files to the media buyer.
- **Do not use for:** Re-judging creative choices already approved in the script.
## Inputs
- **Required:** `every delivered aspect ratio, keyframes per ratio, the placement plan`
- **Optional:** `platform policy notes for the category`
### Input format
- **Script:** scene list with timing, voice-over or dialogue lines, on-screen text, required shots, product moments, CTA.
- **Delivered video:** keyframes every 1 to 2 seconds (a contact sheet works) plus a timestamped transcript. Claude does not accept video files in chat and cannot hear audio; frames plus transcript are how it sees the cut. In Claude Code with ffmpeg installed, Claude can extract the frames itself.
- **Audio and file data:** file specs (for example from ffprobe), a loudness reading, and a person's listening notes. Without them, audio and technical checks are marked Not checkable.
- **Brand kit:** logo, product and packaging shots, brand name spelling, colours, fonts.
- **Rules:** approved claims list, mandatory disclaimers, banned words and visuals, platform and placement specs.
## Parameters
| Parameter | Default | Meaning |
|---|---|---|
| `placements` | Reels 9:16, Feed 4:5, Stories 9:16 | Ordered placements |
| `max_length_s` | 30 | Upper length for the placement plan |
## Definitions
| Term | Definition |
|---|---|
| Delivery | One version of a video a production partner uploads (v1, v2...) |
| Beat | One scripted unit: a line, a shot or an on-screen text moment |
| Deviation | A difference between script and delivery: missing, changed, added or reordered |
| Slop | Visible or audible signs of low-effort or unchecked AI production, or generic filler that weakens the ad |
| Critical | A defect that blocks approval on its own (wrong product, unapproved claim, misspelled brand) |
| Passing grade | Score at or above the threshold with zero open Critical findings |
# Procedure
## Steps
1. For each ratio: is the subject and product inside frame after the crop, or cut off?
2. UI overlap: captions, CTA and product under the platform's buttons, captions or profile bar?
3. Length vs placement; first and last frames work as a loop where the placement loops.
4. Sound-off: the message survives muted in every placement.
5. Policy: category-specific risks (health claims, personal attributes, before/after).
## Output
- Placement table: placement, ratio, fit (Pass / Fix / Fail), issue, timestamp
## Quality gates (check before you answer)
- [ ] Every ordered placement has a row
- [ ] Crop issues name what is cut off
- [ ] Policy flags cite the policy area
- [ ] Every finding has a timestamp, a severity and the rule it breaks; no em dashes
## Output contract
Return the sections above in this order, then a JSON block matching this shape so the scorecard (skill 09) can read it:
```json
{"placements": [{"placement": "Reels", "ratio": "9:16", "fit": "Pass", "issue": null, "t": null}]}
```
## Failure modes
| Situation | What to do |
|---|---|
| Only a master delivered | Simulate crops from the master and say so |
| Placement plan unknown | Check against Reels, Feed and Stories defaults |
## House rules (apply to every output)
- Every finding has a timestamp (mm:ss) or a frame number. A finding you cannot locate is not a finding.
- Every finding has a severity: Critical (blocks approval on its own), Major (must fix before approval), Minor (fix if cheap).
- Quote the script line or brief rule a finding breaks. If there is no rule, it is a suggestion, not a defect, and it never lowers the grade.
- Judge only what is in the frames, transcript and files provided. If a check needs something you were not given, mark it "Not checkable" and say what is missing.
- Taste is not a defect. "I would have shot it differently" never appears in a pre-flight report.
- No em dash characters in outputs.
## Tip
A 4:5 crop of a 9:16 master usually cuts the bottom caption and the product. Ask partners to deliver both, never to crop at upload.
---
Part of the free Adge Pre-Flight skill pack. These skills check the frames and transcript you give them. Adge Pre Flight checks every delivery from every production partner automatically, grades it and sends the feedback for you: adge.ai
A 4:5 crop of a 9:16 master usually cuts the bottom caption and the product. Ask partners to deliver both, never to crop at upload.
Pre Flight returns a Fit Score by market, platform and placement for each creative.
Decide: grade it and tell the vendor
Roll every finding into a weighted score with a passing grade, then turn the fails into feedback a partner can act on in one round.
Pre-Flight Scorecard
Does this delivery pass, and if not, what exactly is blocking it?
Rolls the findings of skills 01 to 08 into one weighted score with a passing grade: any open Critical fails the delivery, and the score must clear the threshold.
SKILL.md · role, inputs, parameters, quality gates, output contract
---
name: preflight-scorecard
description: Rolls the findings of skills 01 to 08 into one weighted score with a passing grade: any open Critical fails the delivery, and the score must clear the threshold. Use when the user asks: Does this delivery pass, and if not, what exactly is blocking it?
---
# Pre-Flight Scorecard
`v1.0` · Adge Pre-Flight · Cadence: Every delivery · Mode: Read only
**Answers:** Does this delivery pass, and if not, what exactly is blocking it?
Rolls the findings of skills 01 to 08 into one weighted score with a passing grade: any open Critical fails the delivery, and the score must clear the threshold.
## Role
You are the pre-flight gatekeeper who decides whether a delivery is approved, sent back, or escalated.
## When to use
- **Use for:** After running the checks; the single verdict that goes into the order tracker.
- **Do not use for:** Re-scoring findings by taste; use the severities as given.
## Inputs
- **Required:** `findings from skills 01 to 08`
- **Optional:** `the previous version's scorecard, to show progress`
### Input format
- **Script:** scene list with timing, voice-over or dialogue lines, on-screen text, required shots, product moments, CTA.
- **Delivered video:** keyframes every 1 to 2 seconds (a contact sheet works) plus a timestamped transcript. Claude does not accept video files in chat and cannot hear audio; frames plus transcript are how it sees the cut. In Claude Code with ffmpeg installed, Claude can extract the frames itself.
- **Audio and file data:** file specs (for example from ffprobe), a loudness reading, and a person's listening notes. Without them, audio and technical checks are marked Not checkable.
- **Brand kit:** logo, product and packaging shots, brand name spelling, colours, fonts.
- **Rules:** approved claims list, mandatory disclaimers, banned words and visuals, platform and placement specs.
## Parameters
| Parameter | Default | Meaning |
|---|---|---|
| `pass_threshold` | 80 | Minimum score to pass (this pack's default; set your own) |
| `weights` | Script 25, Hook 15, Brand 15, Slop 15, Technical 10, Text 10, Platform 10 | Points per area |
| `penalties` | Major -5, Minor -1 | Deducted within an area, floor 0 |
| `critical_rule` | any open Critical = fail | Overrides the score |
## Definitions
| Term | Definition |
|---|---|
| Delivery | One version of a video a production partner uploads (v1, v2...) |
| Beat | One scripted unit: a line, a shot or an on-screen text moment |
| Deviation | A difference between script and delivery: missing, changed, added or reordered |
| Slop | Visible or audible signs of low-effort or unchecked AI production, or generic filler that weakens the ad |
| Critical | A defect that blocks approval on its own (wrong product, unapproved claim, misspelled brand) |
| Passing grade | Score at or above the threshold with zero open Critical findings |
# Procedure
## Steps
1. Start each area at its full weight. Deduct per finding: Major -5, Minor -1, floor 0 per area.
2. Claims (skill 03) is a gate, not points: any Not approved claim is Critical.
3. Score = sum of areas. Verdict:
- **Pass:** score >= pass_threshold and zero open Critical
- **Revise:** score below threshold or any Critical that can be fixed in edit or re-render
- **Escalate:** Critical that needs a re-shoot, or a third failed version
4. Compare with the previous version: fixed, still open, newly introduced.
## Output
- Verdict, score and threshold
- Area table: area, weight, deductions, score
- Blocking findings (Criticals and the Majors that cost the most points)
- Version progress (if a previous scorecard exists)
## Quality gates (check before you answer)
- [ ] Arithmetic is shown per area and sums to the total
- [ ] Any open Critical forces Revise or Escalate whatever the score
- [ ] Version progress lists fixed, open and new findings
- [ ] Every finding has a timestamp, a severity and the rule it breaks; no em dashes
## Output contract
Return the sections above in this order, then a JSON block matching this shape so the scorecard (skill 09) can read it:
```json
{"verdict": "Revise", "score": 0, "threshold": 80, "areas": [{"area": "Script", "weight": 25, "deductions": 0, "score": 25}], "blocking": [], "progress": {"fixed": [], "open": [], "new": []}}
```
## Failure modes
| Situation | What to do |
|---|---|
| A check was not run | Score that area 'Not checked', exclude it from the total and lower the threshold proportionally, and say so |
| Findings lack severity | Send them back to the check that produced them |
## House rules (apply to every output)
- Every finding has a timestamp (mm:ss) or a frame number. A finding you cannot locate is not a finding.
- Every finding has a severity: Critical (blocks approval on its own), Major (must fix before approval), Minor (fix if cheap).
- Quote the script line or brief rule a finding breaks. If there is no rule, it is a suggestion, not a defect, and it never lowers the grade.
- Judge only what is in the frames, transcript and files provided. If a check needs something you were not given, mark it "Not checkable" and say what is missing.
- Taste is not a defect. "I would have shot it differently" never appears in a pre-flight report.
- No em dash characters in outputs.
## Tip
Publish the weights to your partners. Vendors hit a bar they can see; they argue with one they can't.
---
Part of the free Adge Pre-Flight skill pack. These skills check the frames and transcript you give them. Adge Pre Flight checks every delivery from every production partner automatically, grades it and sends the feedback for you: adge.ai
Publish the weights to your partners. Vendors hit a bar they can see; they argue with one they can't.
A delivery only moves to final approval once it earns a passing grade. Below the bar, it goes back to the partner.
Vendor Feedback Writer
Turn the findings into feedback a production partner can fix in one round.
Writes the revision note: timestamped, ranked by severity, each item with what is wrong, why it matters and the exact fix, plus clear acceptance criteria for the next version.
SKILL.md · role, inputs, parameters, quality gates, output contract
---
name: vendor-feedback-writer
description: Writes the revision note: timestamped, ranked by severity, each item with what is wrong, why it matters and the exact fix, plus clear acceptance criteria for the next version. Use when the user asks: Turn the findings into feedback a production partner can fix in one round.
---
# Vendor Feedback Writer
`v1.0` · Adge Pre-Flight · Cadence: Every revision · Mode: Generates
**Answers:** Turn the findings into feedback a production partner can fix in one round.
Writes the revision note: timestamped, ranked by severity, each item with what is wrong, why it matters and the exact fix, plus clear acceptance criteria for the next version.
## Role
You are a production manager writing revision notes to an external production partner.
## When to use
- **Use for:** Every Revise or Escalate verdict.
- **Do not use for:** Vague notes ('make it pop'), taste notes, or new asks that were not in the script.
## Inputs
- **Required:** `scorecard and findings from skill 09, script, partner name, revision deadline`
- **Optional:** `previous feedback sent on this order`
### Input format
- **Script:** scene list with timing, voice-over or dialogue lines, on-screen text, required shots, product moments, CTA.
- **Delivered video:** keyframes every 1 to 2 seconds (a contact sheet works) plus a timestamped transcript. Claude does not accept video files in chat and cannot hear audio; frames plus transcript are how it sees the cut. In Claude Code with ffmpeg installed, Claude can extract the frames itself.
- **Audio and file data:** file specs (for example from ffprobe), a loudness reading, and a person's listening notes. Without them, audio and technical checks are marked Not checkable.
- **Brand kit:** logo, product and packaging shots, brand name spelling, colours, fonts.
- **Rules:** approved claims list, mandatory disclaimers, banned words and visuals, platform and placement specs.
## Parameters
| Parameter | Default | Meaning |
|---|---|---|
| `max_items` | 8 | Most items per note; batch the rest as Minor |
| `tone` | direct, specific, respectful | Voice of the note |
| `deadline_from` | order SLA | Where the due date comes from |
## Definitions
| Term | Definition |
|---|---|
| Delivery | One version of a video a production partner uploads (v1, v2...) |
| Beat | One scripted unit: a line, a shot or an on-screen text moment |
| Deviation | A difference between script and delivery: missing, changed, added or reordered |
| Slop | Visible or audible signs of low-effort or unchecked AI production, or generic filler that weakens the ad |
| Critical | A defect that blocks approval on its own (wrong product, unapproved claim, misspelled brand) |
| Passing grade | Score at or above the threshold with zero open Critical findings |
# Procedure
## Steps
1. Order items: Critical, then Major, then Minor. Merge findings with the same fix.
2. For each item write: timestamp | what is wrong (observable) | why it matters (the rule or script line) | the exact fix | re-shoot or edit.
3. Add what was good in one line, so the partner keeps it in the next version.
4. Acceptance criteria: the checks that must pass in vN+1 and the score needed.
5. Never add a new creative ask. If something outside the script would help, put it in a separate "optional" line that does not affect approval.
## Output
- Subject line: order, script, version, verdict
- Numbered revision items
- Keep as is
- Acceptance criteria and due date
## Quality gates (check before you answer)
- [ ] Every item has a timestamp and an exact fix
- [ ] No item asks for something the script did not
- [ ] Acceptance criteria name the score threshold
- [ ] Every finding has a timestamp, a severity and the rule it breaks; no em dashes
## Output contract
Return the sections above in this order, then a JSON block matching this shape so the scorecard (skill 09) can read it:
```json
{"subject": "", "items": [{"n": 1, "t": "00:00", "severity": "Critical", "issue": "", "why": "", "fix": "", "type": "edit"}], "keep": "", "acceptance": {"must_pass": [], "threshold": 80}, "due": ""}
```
## Failure modes
| Situation | What to do |
|---|---|
| More than max_items findings | Send the top items and batch Minors into one line |
| Third revision on the same issue | Escalate to the account owner instead of sending another note |
## House rules (apply to every output)
- Every finding has a timestamp (mm:ss) or a frame number. A finding you cannot locate is not a finding.
- Every finding has a severity: Critical (blocks approval on its own), Major (must fix before approval), Minor (fix if cheap).
- Quote the script line or brief rule a finding breaks. If there is no rule, it is a suggestion, not a defect, and it never lowers the grade.
- Judge only what is in the frames, transcript and files provided. If a check needs something you were not given, mark it "Not checkable" and say what is missing.
- Taste is not a defect. "I would have shot it differently" never appears in a pre-flight report.
- No em dash characters in outputs.
## Tip
Feedback that quotes the script line gets fixed. Feedback that describes a feeling gets argued with.
---
Part of the free Adge Pre-Flight skill pack. These skills check the frames and transcript you give them. Adge Pre Flight checks every delivery from every production partner automatically, grades it and sends the feedback for you: adge.ai
Feedback that quotes the script line gets fixed. Feedback that describes a feeling gets argued with.
Feedback goes to the partner automatically. Revision deadlines run on each partner's working days, and due-soon and overdue reminders go out for you.
The rules baked into every skill
- Every finding has a timestamp (mm:ss) or a frame number. A finding you cannot locate is not a finding.
- Every finding has a severity: Critical (blocks approval on its own), Major (must fix before approval), Minor (fix if cheap).
- Quote the script line or brief rule a finding breaks. If there is no rule, it is a suggestion, not a defect, and it never lowers the grade.
- Judge only what is in the frames, transcript and files provided. If a check needs something you were not given, mark it "Not checkable" and say what is missing.
- Taste is not a defect. "I would have shot it differently" never appears in a pre-flight report.
- No em dash characters in outputs.
Six quick prompts, no skill file needed
One-prompt pre-flight
Here is the script and a contact sheet plus transcript of the delivered video. Check it against the script beat by beat, list every deviation with a timestamp and severity (Critical, Major, Minor), then give a pass or revise verdict.
Slop sweep
Look at these frames for signs of unchecked AI generation: hands, warped text, product or logo shape changing between frames, faces, lip sync. List each with the frame number and severity.
Claims sweep
Here is the transcript, the on-screen text and our approved claims list. List every claim made, spoken or shown, and mark each Approved or Not approved.
Caption proof
Proofread every on-screen text in these frames against the script and the voice-over. Flag spelling, brand names, numbers, caption sync and text inside the 9:16 safe zones.
Write the revision note
Turn these findings into a revision note for our production partner: numbered, timestamped, severity first, each with the exact fix. No new creative asks.
Did v2 fix v1?
Here are last version's findings and the new version's frames and transcript. For each finding say Fixed, Still open or Made worse, and list anything new.
Manual pre-flight vs Adge pre-flight
The skills will make your reviews faster and fairer. They also show where manual pre-flight stops scaling: someone has to extract the frames, paste the script, send the note and chase the next version. Adge does all of that by itself, for every partner.
| Claude + these skills | Adge Pre Flight | |
|---|---|---|
| Input | Frames and a transcript you extract per video | The delivered file itself, straight from the partner's upload |
| Script | Pasted in by hand | Attached to the order it was written for |
| Coverage | The videos you have time to check | Every version, every hook, every aspect ratio |
| Consistency | Depends on who ran it and which prompt | One rubric and one passing grade for every partner |
| Feedback | Copied into an email or Slack | Sent to the partner automatically, timestamped, with a deadline |
| Enforcement | Nothing stops a failing video from shipping | Only a passing grade moves a video to final approval |
| Follow-up | You remember to chase | Due-soon and overdue reminders on each partner's working days |
Where to run the skills
Claude Projects
Put the SKILL.md files, your brand kit and your claims list in one Project for the whole team.
Claude skills
Zip each skill folder and upload it in Customize > Skills (code execution must be on).
Claude Code
The most complete setup: with ffmpeg on your machine, Claude can pull frames and specs from the video files itself and run the pack as a batch.
Frequently asked questions
Are the skills free?
Yes. Download the pack, use it and share it. Skills work on Claude's Free, Pro, Max, Ultra, Team and Enterprise plans with code execution turned on.
Can Claude watch the video?
No. Claude reads images and text; it does not take video files in chat and cannot hear audio. Export a contact sheet (a frame every 1 to 2 seconds, plus close-ups of hands, faces, text and product) and a timestamped transcript. In Claude Code, Claude can extract the frames itself if ffmpeg is installed.
So what still needs a person?
Listening: voice quality, music, levels and clipping. Watching: lip sync and anything between your frames. The skills mark these Not checkable rather than guess, unless you add a reviewer's notes.
Will this catch every AI artifact?
No. It catches what is visible in the frames you give it. One-frame glitches between your samples will be missed, which is why the skill asks for close-ups of the risky moments.
Is a passing grade of 80 right for us?
It is the pack's starting point, not a standard. Raise it for hero assets, lower it for fast iterations. What matters is that partners know the bar before they deliver.
Can I use this with my agency or UGC creators?
Yes. The vendor feedback skill is written for external partners: timestamped, tied to the script, no taste notes.
What does Adge add on top?
Pre-flight that runs by itself on every delivery from every partner, against the script attached to the order, with feedback sent automatically and a passing grade enforced before anything reaches final approval.
Want this applied to your account?
20 minutes, no slides: your market, the first ads we'd make for you, and your pricing.
Get your ad ideas
