memory/ad/fatigue-frequency-manager/

• Reads: the ad set / campaign under review, a daily (or weekly) time-series export with impressions, reach, frequency, clicks/CTR, conversions/CVR and spend; the early-flight baseline window (first stable days after learning-phase exit); target CPA/ROAS; audience size / saturation estimate if the user has it.
• Writes: a user-facing fatigue table plus a reusable summary storable under

  memory/ad/fatigue-frequency-manager/

  .
• Promotes: confirmed Rotate/Widen triggers, the frequency/decay thresholds used, and any measurement-signal risk (CVR drop that may be broken tracking, not real saturation) to

  memory/open-loops.md

  as

  pending-decision

  — this skill does not write

  decisions.md

  directly.
• Done when: decay is read as a slope against a fixed early-flight baseline (not a raw last-day dip); the diagnosis separates creative fatigue (CTR decays, frequency rises, audience not exhausted) from audience saturation (reach plateaus, frequency climbs because the pool is spent); and exactly one trigger is returned per ad set with the threshold that fired named.
• Primary next skill: use the

  Next Best Skill

  below.

Emit the standard shape from skill-contract.md §Handoff Summary Format.

• ~~ad platform

  (own data) — campaign / ad-set time-series report CSV from the native ad manager: date, impressions, reach, frequency, clicks, CTR, spend, CPM, and the platform's reported conversions/CVR.

• ~~web analytics

  (GA4) — Conversions + Traffic-acquisition export to read CVR from the order-ID truth set, so a CVR drop is checked against real orders before it is called saturation.

• ~~ecommerce

  — store export (orders, revenue) to confirm the conversion side when CVR movement is the trigger.

1. Set the early-flight baseline. Take the first stable window after the ad set exited learning phase (frequency still low, metrics settled) as the baseline. If the set is still in learning, stop — decay is not readable yet; the numbers are noise. Note the learning-exit date.
2. Build the trend, not a snapshot. Read frequency, CTR, and CVR as a slope from baseline to now. Snapshot to the ledger so the delta is computed, not eyeballed:

   python3 "${CLAUDE_PLUGIN_ROOT}/scripts/connectors/ledger.py" record <ad-set> --source paid --data '{"frequency": ..., "ctr": ..., "cvr": ..., "reach": ...}'

   , then

   ledger.py trend <ad-set> --source paid --field ctr

   (repeat for

   frequency

   ,

   cvr

   ,

   reach

   ).
3. Diagnose fatigue vs saturation. Separate the two causes — they take different triggers:
   • Creative fatigue → CTR decays and frequency rises while reach is still growing (the pool isn't exhausted, the same people are just seeing a tired ad). Trigger: Rotate-creative.
   • Audience saturation → reach plateaus and frequency climbs because delivery has run out of new people; CTR/CVR fall as it re-serves the same pool. Trigger: Widen-audience.
   • Both can co-occur; name the dominant driver and the secondary one.
4. Check the frequency threshold. Compare current frequency to the working ceiling for the objective (per measurement-protocol.md); a prospecting set tolerates a lower frequency before decay than a warm-retargeting set. State the ceiling you used and whether it was breached — do not assert a universal "frequency 3" rule.
5. Confirm the CVR drop is real, not broken tracking (ROAS-R protection). A falling CVR can be genuine saturation or a measurement-signal fault. Read CVR against the GA4/ecommerce order truth set; if conversion tracking looks broken/unverifiable (ROAS-R1) or double-counted across platforms (ROAS-R2), the decay read is untrustworthy → flag it and hand the signal to the auditor gate rather than firing a trigger on dirty data. See roas-benchmark.md for the Return-dimension vetoes. This skill flags; it does not score or veto.
6. Return one trigger per ad set. For each set output: baseline window · frequency now vs baseline · CTR slope · CVR slope · reach trend · dominant cause · trigger (Rotate-creative / Widen-audience / Hold) · the threshold that fired · caveats. Hold when decay is within noise or the window is too short to call.

• Stop and ask — only when there is no time series at all (a single-day export), or the ad set is still in learning phase. Present the two options: (1) supply the daily time-series export, or (2) supply the learning-exit date, and state that no fatigue read is possible until one is available.
• Continue silently — if audience-size/saturation estimate is missing (infer saturation from the reach-plateau + frequency-climb signature and mark it Estimated); if only some ad sets in the campaign have full data (read those, mark the rest N/A); if CVR is absent but CTR + frequency are present (read a creative-fatigue signal on CTR alone and note CVR was unavailable).

• ROAS Benchmark — the paid-ads scoring framework; this skill works the S (CTR/CVR/frequency decay under spend-efficiency) and R (return-protection) levers; the Return vetoes R1/R2 govern whether a CVR-based read is trustworthy. Only

  ad-account-auditor

  computes the RQS or runs vetoes.
• Measurement & Attribution Protocol — baseline windows, conversion-lag handling, and frequency-ceiling guidance by objective.
• scripts/connectors/README.md —

  ledger.py

  record / trend reference for the decay slope.
• ad-creative-builder — builds the replacement creative when a Rotate-creative trigger fires (this skill diagnoses; it does not produce the ad).

• Rotate-creative fired → ad-creative-builder to produce the fresh ad unit (ad↔LP message-match + claim/policy checks live there).
• Widen-audience fired → audience-segment-builder to expand seed/lookalike segments from the user's own data.
• A measurement-signal risk was flagged (ROAS-R1/R2) → stop and route to ad-account-auditor — the gate scores the RQS and runs the vetoes; do not act on a fatigue read built on untrusted conversion data.
• Hold → terminal; report chain-complete.

max-depth: 3