---
title: Ad Creative Fatigue Skill for Claude Code
canonical: https://skillselion.com/skills/mbfinotti/advertising-skills/ad-creative-fatigue
content_type: text/markdown
description: ad-creative-fatigue by mbfinotti - This skill decides whether a running ad creative is genuinely wearing out. It checks confounders such as budget changes.
---

# Ad Creative Fatigue Skill

Title: Ad Creative Fatigue Skill for Claude Code

> ad-creative-fatigue by mbfinotti - This skill decides whether a running ad creative is genuinely wearing out. It checks confounders such as budget changes.

Canonical: https://skillselion.com/skills/mbfinotti/advertising-skills/ad-creative-fatigue

This skill decides whether a running ad creative is genuinely wearing out. It checks confounders such as budget changes, learning-phase resets, audience saturation, CPM inflation, seasonality and tracking breakage. It returns a verdict with confidence and the best remedy for the effort. A solo founder reaches for it when an ad's results fall and they need to know whether to replace the creative.

## Key facts

- Installs: 26,822
- GitHub stars: 2
- Author: mbfinotti
- Repository: mbfinotti/advertising-skills
- Works in Grok Bot: yes
- Install command: `npx skills add https://github.com/mbfinotti/advertising-skills --skill ad-creative-fatigue`
- Last updated: October 26, 2026

## What Ad Creative Fatigue does

Checks whether a running ad creative is wearing out or whether a confounder such as a budget change or tracking break explains the drop.

Declining ads get blamed on creative when budget or tracking is the cause.

A verdict on fatigue with a ranked fix.

Best for: Founders running always-on paid campaigns.

Not for: Brand-new ads with too little data to judge.

- Separates true creative fatigue from confounders
- Verdict with a confidence level
- Highest-return remedy per unit of effort

## SKILL.md

## Creative Fatigue

Read a creative's performance-over-time data and decide whether it is genuinely wearing out, how confident that call is, and what to do about it. The core discipline is refusal: many things can cause a declining metric, and creative wear-out is only one of them. Most refresh decisions are made too early, off one metric, against no baseline.

This skill runs a differential diagnosis in a fixed order, confounder screen before the fatigue call, never after:

1. Baseline
2. Confounder screen
3. Decay measurement
4. Confidence gate
5. Verdict

Two further tensions stay live throughout:

- Fatigue is often hook fatigue rather than whole-ad fatigue (Ben Heath).
- Some ads improve with repeated exposure, the "wear-in" effect Les Binet describes in brand advertising, though Meta's 2023 research found no wear-in for direct-response objectives - wear-in is a real verdict only where the objective is brand, not DR.

There are no universal thresholds here: every trigger is derived from the creative's own trailing baseline and the account's own history. Named practitioner numbers appear only as attributed starting reference points.

This skill ends at the verdict and the recommended action.

- Producing the replacement asset belongs to `mbfinotti/advertising-skills@ad-copy-variants`, `mbfinotti/advertising-skills@ugc-ad-scripts`, and `mbfinotti/advertising-skills@ad-creative-brief`.
- Sample-sizing a pre-launch test belongs to `mbfinotti/advertising-skills@ad-creative-test-plan`.
- Scoring a video opening before launch belongs to `mbfinotti/advertising-skills@ad-hook-analyzer` (here, hook rate is only read as a decay signal on ads already running).
- When the confounder screen shows the problem is not the creative at all (structure, targeting, tracking), hand off to `mbfinotti/advertising-skills@ad-account-diagnostic`.
- Retargeting-sequence and frequency-cap architecture design is `mbfinotti/advertising-skills@retargeting-funnel`.
- Scaling and pacing decisions after a win are `mbfinotti/advertising-skills@paid-media-scaling` and `mbfinotti/advertising-skills@ad-budget-pacing`.

### Interview

Ask before diagnosing anything. One question per message; offer multiple-choice answers where possible; skip anything already answered or visible in supplied data.

- Which platform(s) is the creative running on? (Meta / Google Ads - search, Performance Max, or YouTube / LinkedIn / TikTok / other)
- B2B or B2C/ecommerce?
- Funnel stage: cold prospecting or retargeting/warm? (Tolerated exposure differs sharply between the two.)
- What can you actually export, and at what granularity: per-creative per-day, per-creative weekly aggregate, or only ad-set/campaign level? Per-creative per-day is the working assumption; anything coarser weakens every step downstream.
- Daily spend and conversion volume on the creative in question? (Decides whether the confidence gate can clear at all.)
- Rough audience size for the ad set, and is the audience list-based/ABM or broad?
- How many creatives are live in the same ad set? (Sibling mix changes what "losing spend share" means.)
- When was the ad, or its parent ad set/campaign, last edited - creative, budget, bid, audience, optimization event? Exact date matters.
- Did budget change during the window under suspicion? By how much?
- By what date does the result have to land? (A hard deadline promotes the Action Ladder's same-day rungs - rotation, budget shift, frequency cap - and demotes iteration and new concepts.)
- Do you want a one-off win on this creative, or a compounding asset? (A compounding mandate promotes iteration and new concepts even where a configuration fix would hold this week.)
- What is your effort ceiling: creative production capacity and lead time, in-house editing hours, whose sign-off list work needs, and how reversible the change has to be? (A refresh recommendation nobody can execute is worthless - this answer deletes rungs from the Action Ladder rather than reordering them.)
- What is the target CPA or ROAS the creative is judged against?

### Workflow

1. Run the Interview; collect every answer before touching the data.
2. Fix the unit of analysis. Analyse at creative and concept level, never at campaign averages: forty rows for forty variants of five concepts hides the real pattern, and one fatigued ad dragging an ad-set average looks like a campaign-wide problem. Where the platform rotates variants automatically (Advantage+-style delivery, dynamic creative), read the asset-level breakdown - automated rotation masks individual asset decay. Conversely, a creative silently losing budget share with no manual change is itself a decay signal.
3. Build the creative's own baseline. Never compare single days.

   Sourced methods to offer as starting defaults, picked per the user's data granularity and overridden by their own history - the two are not ranked against each other, since both cost the same single export and only granularity decides:

   - 3-day moving average against the creative's own trailing 14-day baseline (AdSights).
   - Rolling 7-day window against a 30-day baseline (Segwise).

   Exclude known outages and match the comparison window's day-of-week composition to the baseline's. If your harness can read the export or run the calculation, compute it; otherwise emit the exact export steps and spreadsheet formulas (columns, moving-average ranges, delta formula `(current - baseline) / baseline`) for the user to run and report back.

4. Run the confounder screen, before any fatigue talk. Work through [references/confounders.md](references/confounders.md) and mark every confounder pass (ruled out) or fail (present):

   - budget/bid change
   - learning-phase reset
   - audience overlap/saturation
   - auction CPM inflation
   - seasonality and window composition
   - tracking breakage
   - attribution-window skew
   - placement/device mix shift
   - statistical noise
   - landing-page/offer change
   - sibling-mix shift

   Any fail that explains the decline ends the fatigue inquiry: the verdict is `confounded` (or `saturating` for the audience case) and the fix targets the actual cause. Structure/targeting/tracking root causes hand off to `mbfinotti/advertising-skills@ad-account-diagnostic`.

5. Measure decay against the baseline using [references/signal-reference.md](references/signal-reference.md). Apply the field's consensus decision rule: no single metric confirms fatigue - require at least two signals moving together across two or more consecutive periods (Segwise; AdSights). Weight leading signals (link CTR decay, hook-rate decay on video, falling first-time-impression ratio, silent budget-share loss) over lagging ones. Frequency is a lagging, confirming signal only - Meta's own analytics team notes reported frequency is measured at ad/ad-set level while fatigue happens at creative level, and is a period average, not the marginal effect of the next impression. Do not build the call on frequency.
6. Apply the Confidence Gate (below). If it fails, the verdict is `insufficient data` - state what extra spend or days would clear it and stop; refuse to recommend a refresh on noise.
7. Separate fatigue from saturation before finalising. The discriminating logic:

   - Fatigue: costs rise while conversion rate holds.
   - Saturation: both degrade, alongside a falling first-time-impression ratio and flattening reach.

   Where the user can run it, the definitive test is Meta's own:

   - A new creative restoring performance on the same audience proves fatigue.
   - Performance recovering only on a fresh audience proves saturation.

   Their fixes are opposite, new creative versus audience expansion, so the split matters.

8. Issue the verdict from the Verdict Ladder, then pick the remedy from the Action Ladder - top of its efficiency ordering first, re-ranked against the Interview's deadline, one-off-versus-compounding and effort-ceiling answers - and fill one Fatigue Verdict block per creative (shape below; worked versions in [references/examples.md](references/examples.md)).
9. Set the re-check date (one full comparison window after any action) and log the decision for the Measuring section's scorecard.
10. If your harness has persistent memory, memorize the account's baselines, per-creative verdicts, actions taken, and re-check dates, so the next run starts from history instead of re-deriving it.

### The Fatigue Verdict

Deliver one block per creative analysed:

```
FATIGUE VERDICT - <creative id/name>, <date>
platform      : <platform> | funnel stage: <cold prospecting | retargeting>
window        : <comparison window> vs baseline <baseline definition>
volume        : <spend> spent, <impressions> impressions, <conversions> conversions in window

signals
  <signal>    : <current> vs <baseline> (<+/-x%>)   [leading|lagging]
  <signal>    : <current> vs <baseline> (<+/-x%>)   [leading|lagging]
  ...

confounder screen
  budget/bid change        : pass | FAIL - <note>
  learning-phase reset     : pass | FAIL - <note>
  audience saturation      : pass | FAIL - <note>
  auction CPM inflation    : pass | FAIL - <note>
  seasonality/window mix   : pass | FAIL - <note>
  tracking breakage        : pass | FAIL - <note>
  attribution-window skew  : pass | FAIL - <note>
  placement/device mix     : pass | FAIL - <note>
  statistical noise        : pass | FAIL - <note>
  landing page/offer change: pass | FAIL - <note>
  sibling-mix shift        : pass | FAIL - <note>

confidence    : high | medium | low - <one-line basis: gate math, signal count, window length>
verdict       : fatigued | saturating | confounded | insufficient data | healthy | wear-in
action        : <chosen action-ladder rung + who produces the asset, if any>
ruled out     : <every rung the account's stated constraints delete, each with the constraint that deleted it - or "none">
expected      : <what should recover, by roughly how much, based on which evidence>
re-check      : <date - one full comparison window after action>
```

### Confidence Gate

No verdict may be issued below the data floor, and the floor is the user's own math, not a copied constant:

- **Noise check.** For a rate signal (CTR, hook rate, CVR) the observed decline must exceed sampling noise. Approximate the baseline rate's noise band as `p ± 2 × sqrt(p × (1-p) / n)`, with `n` the impressions (or clicks, for CVR) in the comparison window. A "decline" still inside that band is noise, whatever it looks like on a chart.
- **Two-signal rule.** At least two signals, at least one of them leading, moving the same direction across two or more consecutive periods (Segwise; AdSights). One metric, one period, never clears the gate.
- **Conversion floor.** For CPA/ROAS-based claims, enough conversions in both baseline and comparison windows that the delta survives the same noise check on CVR. Low-conversion accounts (most B2B) should gate on leading engagement signals instead and say so in the verdict's confidence line.

When the gate is not met: report `insufficient data`, compute the additional days or spend needed (days ≈ what it takes for `n` to make the noise band narrower than the observed delta at current traffic), and refuse to recommend a refresh. Killing a creative on an underpowered read is the most expensive false positive this skill exists to prevent.

### Verdict Ladder

Each verdict maps to its own action - they never all collapse to "make new creative". These verdicts are deliberately not ranked against each other: the evidence picks exactly one, so an efficiency ordering over them would be false precision. The ranking happens one level down, inside the Action Ladder.

| Verdict             | Meaning                                                                                                                                                                  | Action                                                                                                                                                                                 |
| ------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `fatigued`          | Two-plus signals decayed vs own baseline, confounders ruled out, gate cleared                                                                                            | Action Ladder, highest return per unit of effort first - rung 1 unless the evidence rules it out                                                                                       |
| `saturating`        | The audience pool is depleting - falling first-time-impression ratio, flattening reach, CVR degrading with costs                                                         | Audience expansion, fresh seed, or exclusions - not new creative; Action Ladder rung 6, promoted to first under this verdict; see `mbfinotti/advertising-skills@ad-audience-targeting` |
| `confounded`        | A non-creative cause explains the decline                                                                                                                                | Fix the actual cause; hand structural/tracking causes to `mbfinotti/advertising-skills@ad-account-diagnostic`                                                                          |
| `insufficient data` | Confidence Gate not met                                                                                                                                                  | Report the gap, state days/spend needed, re-check then; no refresh                                                                                                                     |
| `healthy`           | Signals inside the noise band of the creative's own baseline                                                                                                             | Leave it alone; note next scheduled review                                                                                                                                             |
| `wear-in`           | Performance improving with exposure - plausible for brand objectives (Les Binet); Meta found no wear-in for direct response, so treat a DR "wear-in" read with suspicion | Leave it alone; do not rotate it on a frequency number                                                                                                                                 |

### Action Ladder

Ranked by return per unit of effort - what each rung buys, against what it costs you to ship it. Not by price: the cheapest rung and the one worth doing first are rarely the same rung, and "buys a week" is not the same purchase as "restores the ad".

- efficiency (the default order, and the numbering below): hook swap > rotate & rest > budget shift > iterate the winner > frequency cap > audience expansion > new concept > pause
- effort, lightest first: rotate == budget shift == pause > frequency cap > hook swap > audience expansion > iterate > new concept - the three-way tie is a real equality, not indecision: each is one console change, produces no asset, needs no sign-off, and is reversible the same day
- durability of what it buys: new concept > iterate > audience expansion > hook swap > frequency cap > rotate > budget shift > pause - a concept is the parent of every iteration run off it, so it outlives any single one of them
- compliance cost, heaviest first: audience expansion (customer-list and ABM uploads need a lawful-basis attestation and a data-processing review, and an uploaded list is not easily un-shared) > every other rung (near-zero - all are platform settings or new assets)

Excerpt. The rest of the file is on the canonical page and in the source repository.

## Ad Creative Fatigue FAQ

### Is Ad Creative Fatigue safe to install?

skills.sh reports 3 of 3 security scanners passed. Review the Security Audits panel on this page before installing in production.

## Machine access

- [Sitemap](https://skillselion.com/sitemap.xml)
- [llms.txt](https://skillselion.com/llms.txt)
- [llms-full.txt](https://skillselion.com/llms-full.txt)
- [Developer resources](https://skillselion.com/developers)
- [OpenAPI spec](https://skillselion.com/openapi.json)
- [MCP handshake](https://skillselion.com/.well-known/mcp)
- [API catalog](https://skillselion.com/.well-known/api-catalog)
- [AI catalog](https://skillselion.com/.well-known/ai-catalog.json)
- [Pricing](https://skillselion.com/pricing.md)
- [Auth](https://skillselion.com/auth.md)
- [AGENTS.md](https://skillselion.com/AGENTS.md)
- [Contact](https://skillselion.com/contact)
- [Home](https://skillselion.com/)
