Claude for Post-Flight: one skill to read your Ads Manager data
Find the exact stage where each ad loses people, and what to scale, fix and iterate this week.
A free Claude skill that reads your Meta Ads Manager export, finds where each ad leaks (hook, hold, click, land or convert) and tells you what to scale, fix and iterate this week.
By Adam Raz
CRO · Ex-Meta · Published
Launching the ad is the easy part. The hard part is reading what came back: a hook rate that looks great next to a CPA that does not, a video people watch to the end and never click, a page that loses half the clicks before it loads. Ads Manager has every number you need; it just doesn't tell you what they mean. This free Claude skill reads your Ads Manager export, finds the exact stage where each ad loses people, and tells you what to scale, fix and iterate this week.
Every ad is a five-stage funnel
Each stage has one number in Ads Manager, or a simple ratio of two. Find the first stage that is low and you have found the fix.
- Stage 1
Hook
Did the opening stop the scroll?
3-second plays / impressions
- Stage 2
Hold
Did the body keep them?
ThruPlays / 3-second plays
- Stage 3
Click
Did the ad make them act?
Link clicks / impressions
- Stage 4
Land
Did the page actually load?
Landing page views / link clicks
- Stage 5
Convert
Did the page close the sale?
Purchases / landing page views
Symptom to fix: what your numbers are telling you
High hook, low hold
Means: The opening works, the body loses them
Do: Keep the hook, test it on a different body
Low hook, high hold
Means: Strong content behind a weak opening
Do: Re-cut the first 3 seconds, keep the rest
They watch, but don't click
Means: The CTA or the offer is missing or unclear
Do: Add a spoken CTA and an end card; state the offer
They click, but the page doesn't load
Means: Page speed, redirects or a missing Pixel
Do: Fix the page before you touch the creative
They land, but don't buy
Means: The page doesn't deliver what the ad promised
Do: Match price, offer and claim on the first screen
CTR falling, frequency rising
Means: The ad is wearing out
Do: New hook on the same body, then a new concept
The skill: Post-Flight Leak Finder
One skill, one weekly question. Upload the export, get a verdict on the data, the leak stage for every ad, and three actions ranked by what they are costing you.
Post-Flight Leak Finder
Where does each ad lose people, and what do we do about it this week?
Reads one Ads Manager export, checks it is fit to judge, finds the one funnel stage where each ad leaks against your own account median, and turns it into a short list: scale, fix outside the creative, iterate, watch or wait.
SKILL.md · role, inputs, parameters, steps, quality gates, output contract
---
name: postflight-leak-finder
description: Reads one Ads Manager export, checks it is fit to judge, finds the one funnel stage where each ad leaks against your own account median, and turns it into a short list: scale, fix outside the creative, iterate, watch or wait. Use when the user asks: Where does each ad lose people, and what do we do about it this week?
---
# Post-Flight Leak Finder
`v1.0` · Adge Post-Flight · Cadence: Weekly · Mode: Read only
**Answers:** Where does each ad lose people, and what do we do about it this week?
Reads one Ads Manager export, checks it is fit to judge, finds the one funnel stage where each ad leaks against your own account median, and turns it into a short list: scale, fix outside the creative, iterate, watch or wait.
## Role
You are a performance creative strategist who diagnoses the stage before prescribing the fix, and refuses to judge ads the data cannot carry.
## When to use
- **Use for:** Weekly creative reviews, and any ad with a CPA above target.
- **Do not use for:** Questions about why an ad works (that needs the creative tagged), trends over time (that needs a daily breakdown), or competitor context.
## Inputs
- **Required:** `Ads Manager export at ad level (CSV or XLSX)`
- **Optional:** `target CPA or ROAS, attribution setting`
### Input format
- **Export:** Ads Manager, Ads tab, ad level, one date range (7 to 28 days is typical). Export the table as CSV or XLSX and upload it in the chat. Claude reads it with code execution.
- **Columns:** Ad name, Amount spent, Impressions, Reach, Frequency, 3-second video plays, ThruPlays, Video plays at 75%, Link clicks, Landing page views, your result (for example Purchases), Purchases conversion value.
- **Optional:** your target CPA or ROAS, and the attribution setting the export uses.
## Parameters (defaults; change them to fit your account)
| Parameter | Default | Meaning |
|---|---|---|
| `min_impressions` | 1000 | Below this, hook and hold are not judged |
| `min_link_clicks` | 100 | Below this, click-to-land and CVR are directional only |
| `min_results` | 10 | Below this, CPA is directional only |
| `band` | 0.8x to 1.2x median | Typical range; outside it is Low or High |
| `freq_watch` | 2x the account median frequency | Frequency that earns a Watch flag |
## Definitions
| Term | Formula or meaning |
|---|---|
| Hook rate | 3-second video plays / Impressions. An industry convention, not a Meta column |
| Hold rate | ThruPlays / 3-second video plays. A ThruPlay is 15 seconds watched, or the whole video if shorter |
| Link CTR | Link clicks / Impressions |
| Click-to-land | Landing page views / Link clicks. Needs the Meta Pixel on the page |
| CVR | Purchases (or your result) / Landing page views |
| CPA | Amount spent / Purchases |
| ROAS | Purchases conversion value / Amount spent |
| Frequency | Impressions / Reach for the period |
| High / Low | At least 1.2x / at most 0.8x the account median for that metric (default band) |
# Procedure
## Steps
1. **Sanity check.** Map the columns to the glossary and name any missing column and what it blocks (no ThruPlays means no hold rate). Classify each ad as Video, Static or Carousel; video rates are not judged for statics. Apply the volume thresholds per metric, not per ad. Flag impossible rows (landing page views above link clicks). Ask for the attribution setting if it is not stated.
2. **Rates.** For each ad compute Hook, Hold, Link CTR, Click-to-land and CVR, plus CPA, ROAS and frequency.
3. **Medians.** Compute the account median of each rate across the ads that pass the thresholds. Index every rate: ad value / median, and mark it High, Typical or Low.
4. **The leak.** The leak is the earliest stage marked Low. If no stage is Low but CPA is above target, the leak is the lowest index.
5. **Split the post-attention gap.** When Hook and Hold are fine but results are not, say which of three problems it is:
- A, watched but no click (Link CTR Low): CTA or offer missing or unclear. Owner: creative.
- B, clicked but did not land (Click-to-land under 0.80): page speed, redirects or Pixel. Owner: web. Never blame the creative.
- C, landed but did not buy (CVR Low): the page does not deliver what the ad promised. Owner: offer or web.
6. **The call.** Put every ad in one bucket:
- Scale: no Low stage and CPA at or below median.
- Fix outside creative: leak is Land or Convert.
- Iterate: leak is Hook, Hold or Click, with one specific change (keep the part that works, change the part that leaks).
- Watch: frequency at or above freq_watch. This export cannot show a trend; ask for a daily breakdown before calling it fatigue.
- Wait: below the volume thresholds.
7. **This week.** Pick the top 3 actions, ranked by spend on the ad x size of the gap. Fixes outside the creative come first: no new creative is judged fairly on a broken page.
## Output
- Data verdict: Ready | Ready with gaps | Not ready, and what to re-export
- Funnel table: ad, five rates with index, leak stage, CPA, bucket
- One line per leaking ad: stage, number vs median, likely cause ("points to"), the test that would confirm it
- Top 3 actions this week
- What this export cannot tell you: list the questions it leaves open (why a winner wins, whether an ad is fatiguing, how it compares with competitors)
## Quality gates (check before you answer)
- [ ] Every ad appears once, in exactly one bucket
- [ ] Every rate shows its value and its index; medians exclude ads below the thresholds
- [ ] Gap B is never blamed on the creative
- [ ] Watch is never called fatigue without a daily trend
- [ ] Every finding names its funnel stage and shows the number vs the median; no em dashes
## Output contract
Return the sections above in this order, then a JSON block matching this shape so you can compare runs week to week:
```json
{"verdict": "Ready with gaps", "medians": {"hook": 0, "hold": 0, "ctr": 0, "click_to_land": 0, "cvr": 0, "cpa": 0}, "ads": [{"ad": "", "rates": {}, "leak": "Hold", "gap": null, "cpa": 0, "bucket": "Iterate", "cause": "", "test": ""}], "this_week": ["", "", ""], "open_questions": []}
```
## Failure modes
| Situation | What to do |
|---|---|
| Export at campaign or ad set level | Ask for the Ads tab export; creative questions need ad-level rows |
| Fewer than 5 ads pass the thresholds | Use target values instead of medians and say so |
| No landing page views column | Skip gaps B and C and say so |
| Lead or app objective | Swap CVR for your result per landing page view or per install |
## House rules (apply to every output)
- Name the funnel stage for every finding: Hook, Hold, Click, Land or Convert. A finding with no stage is an opinion.
- Show the number, its formula and the comparison: "Hold 11% (ThruPlays / 3-second plays) vs account median 40.5%".
- Compare against this account's own median by default. Industry benchmarks are rough context, never the verdict.
- Respect minimum volume. Below the thresholds, say "Not enough data" instead of guessing.
- Metrics show what happened, not why. Say "points to" or "consistent with", and say which test would confirm it.
- Only compare periods and ads that use the same attribution setting.
- Read only. Recommend actions; never claim you changed a campaign, budget or ad.
- No em dash characters in outputs.
## Tip
Save a column preset in Ads Manager called Post-Flight with every column in the input spec. One click every week, and the skill always gets the same shape.
---
Free from Adge. This skill reads the export you give it. Adge connects to your ad accounts, tags every creative and runs post-flight across every ad automatically, with fatigue, competitor context and the next round of briefs: adge.ai
Three to scale, five leaking at five different stages, one to watch, one too small to call.
| Ad | Leak | Number vs median | CPA | Call | First fix |
|---|---|---|---|---|---|
| A02, A08, A10 | None | No stage below the band | $29.6 to $33.8 | Scale | Move budget here first |
| A05 | Land (gap B) | Click-to-land 54% vs 87% | $46.5 | Fix outside creative | Page or redirect: half the clicks never load |
| A06 | Convert (gap C) | CVR 1.2% vs 5.0% | $105.0 | Fix outside creative | Page does not show the 40% the ad promises |
| A01 | Hold | Hold 11% vs 40.5% | $66.9 | Iterate | Keep the opening, re-cut the body |
| A03 | Hook | Hook 14% vs 30.5% | $31.8 | Iterate | New first 3 seconds on a body that holds 58% |
| A04 | Click (gap A) | Link CTR 0.43% vs 1.09% | $103.1 | Iterate | People watch, nobody clicks: add a spoken CTA and end card |
| A07 | Convert | CVR 3.7%, frequency 4.3 vs 1.1 to 2.1 | $117.1 | Watch | Pull a daily breakdown before calling fatigue |
| A09 | None | 920 impressions, 10 link clicks | n/a | Wait | Not enough data |
- This week: 1) fix A05's page before touching any creative, 2) re-cut A01's body, 3) add a CTA to A04.
- Verdict: Ready with gaps. The attribution setting is not in the export, so ask before comparing with last month.
- What this export cannot tell you: why A02, A08 and A10 win, whether A07 is wearing out, and how any of it compares with your competitors.
Save a column preset in Ads Manager called Post-Flight with every column in the input spec. One click every week, and the skill always gets the same shape.
Adge connects to your ad accounts, so there is no export: every creative shows its KPIs against your account average, every day, across Meta, TikTok and YouTube.
Sample data: a fictional supplement account, 10 ads, 1 to 14 Nov 2026, built to illustrate the skill. Not real brand data. The same CSV is in the download so you can run the skill and check the answer.
What one skill can't see
The skill tells you where each ad leaks. The questions that decide next month's creative need more than one export can carry: the creative itself, tagged; history; and your category. That is Adge's post-flight.
| The question | Claude + this skill | Adge post-flight |
|---|---|---|
| Why the winner wins | The skill sees numbers, not creative. It can say A02 holds; it cannot say it is the 3-reasons demo body that holds across every hook. | Every creative is tagged automatically (hook, concept, presenter, format), and the Why it works tab shows which tags drive the KPI you pick. |
| Fatigue before CPA shows it | One export is one snapshot. The skill can flag high frequency, not a decline. | A Fatigue tab on every creative and Fatigue over time by topic, so you see which ideas are wearing out across the account. |
| Hook vs body, at scale | Only works if every ad name follows a convention, and only for the ads in the file. | Group every creative you have ever run by primary concept, hook or presenter and compare them on any KPI. |
| What competitors are doing | Your median is the only bar. The skill has no idea what your category runs. | Your creative next to your main competitor and a category benchmark from the Meta Ad Library, on the same taxonomy. |
| Is the account getting better | No history between runs. | Spend, ROAS, Stick Rate, Winners and Creative Concentration against the prior period, on one dashboard. |
| What to make next | A plan in a chat that someone still has to turn into briefs. | Opportunity Hub turns the findings into scripts and creator briefs, Marketplace sends them to production partners, and Pre Flight checks what comes back. |
The skill finds the leak in one export. Adge runs post-flight on every ad, every day, without one.
See Adge on your accountWhat Claude can and cannot do here
What Claude can do with this skill
- Read a CSV or XLSX export from Ads Manager and compute every rate with code
- Find the funnel stage where each ad leaks, against your own account median
- Tell a creative problem from a page or tracking problem
- Turn it into scale, fix, iterate, watch or wait, and three actions for this week
What it cannot do alone
- Tell you why a winner wins: it sees numbers, not the creative
- Show fatigue: one export is one snapshot, with no trend
- Compare you with your competitors or with last month
- Pull data, change a budget or brief the next round by itself
The rules baked into the skill
- Name the funnel stage for every finding: Hook, Hold, Click, Land or Convert. A finding with no stage is an opinion.
- Show the number, its formula and the comparison: "Hold 11% (ThruPlays / 3-second plays) vs account median 40.5%".
- Compare against this account's own median by default. Industry benchmarks are rough context, never the verdict.
- Respect minimum volume. Below the thresholds, say "Not enough data" instead of guessing.
- Metrics show what happened, not why. Say "points to" or "consistent with", and say which test would confirm it.
- Only compare periods and ads that use the same attribution setting.
- Read only. Recommend actions; never claim you changed a campaign, budget or ad.
- No em dash characters in outputs.
Three quick prompts, no skill file needed
One-prompt post-flight
Here is my Ads Manager export at ad level. For each ad compute hook rate (3-second plays / impressions), hold rate (ThruPlays / 3-second plays), link CTR, landing page views / link clicks and purchases / landing page views. Compare each with the account median and tell me the one stage where each ad leaks the most, and the first fix.
Watched but didn't click
Find ads where hook and hold rates are at or above the account median but link CTR is low. For each, tell me whether the problem looks like the CTA, the offer or the page, and what test would confirm it.
Is it the page?
Compare landing page views with link clicks for every ad. Flag any ad below 80% and tell me whether it points to page speed, a redirect or a missing Pixel, before I change any creative.
Frequently asked questions
Is the skill free?
Yes. Download it, use it and share it. Skills work on Claude's Free, Pro, Max, Ultra, Team and Enterprise plans with code execution turned on.
Why one skill and not a pack?
Because one question matters every week: where is each ad losing people, and what do we do about it. The other post-flight questions (why it wins, is it fatiguing, how do we compare) need tagged creative, history and competitor data that a single export does not carry. That part is what Adge does.
Is hook rate a Meta metric?
No. Hook rate (3-second video plays / impressions) and hold rate (ThruPlays / 3-second video plays) are industry conventions built from Meta's columns. You can add them to Ads Manager as custom metrics.
Why compare with my own median instead of a benchmark?
Published benchmarks vary by industry, placement and source. Your own median compares ads that share your product, audience and attribution setting, which makes it the fairer bar. You can also give the skill a target CPA or ROAS.
Is the example real data?
No. It is a fictional account built to show each pattern clearly. The CSV is in the download so you can run the skill and check that you get the same answer.
Does this work for TikTok?
The logic does, but the columns differ: TikTok counts a view at 2 seconds and uses 6-second focused views. Rename the columns and change the formulas in the Definitions table, and never compare a TikTok rate with a Meta one.
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