Creative Fatigue Rotation Planner

Creative Fatigue & Rotation Planner (read-only)

A read-only skill. It reads ad-level performance over two time windows, measures how far each creative has decayed from its own baseline, and prints a ranked rotation plan. It never swaps, uploads, pauses, or appeals anything — the actual rotation is handed to manage-creative.

Tool chain (verified against the live MCP):

Stage 0: resolve advertiser (paste advertiser_id/bc_id, or auth_advertiser_get / bc_get) → advertiser_info_get
Stage 1: confirm windows (recent 3d vs baseline 30d) + ask if a library scan is wanted
Stage 2 (read-only):
  ① report_integrated_get (ad level, recent 3 days: ctr / conversion_rate / spend / impressions)   → sync
  ② report_task_create → report_task_check → report_task_download (ad level, last 30 days)          → async
  ③ smart_plus_ad_get / ad_get                                                                       → ad ↔ creative names
  ④ (optional) file_video_ad_search / file_image_ad_search                                          → fresh candidates
Stage 3: compute decay per creative → classify Retire / Scale / Watch → spend at risk
Stage 4: render the rotation plan (+ optional swap-in shortlist) → route the swap out

🔒 Read-only scope (one-strike rule)

This skill must NEVER call any write tool — no ad_update, smart_plus_ad_update, *_status_update, *_material_status_update, file_*_upload, adgroup_appeal, *_create (except report_task_create, a read-only report task), or *_delete. Only the read tools in the dispatcher table below plus tool_execute / tool_get. If asked to plan AND swap, produce the plan, then refuse the swap and route to manage-creative. A silent write = catastrophic failure.

What this skill is NOT for (route elsewhere)

It sounds like what you really want is [to execute a swap / a full report / a delivery diagnosis] — there's a dedicated skill for that. Want me to switch to <skill>?

MCP backend (dispatcher)

Runs through the TikTok Ads MCP. layer (tt-ads-mcp-layer: ~40 tools + tool_execute/tool_get dispatcher) or flat (tt-ads-mcp-flat: all ~400 callable directly). Mapping below names each real tool; under flat, call it directly.

Complete tool mapping (single source of truth). All READ-ONLY:

Metric values come back as strings — parse to numbers. The conversion metric is conversion (singular); conversion rate is conversion_rate. There is a dedicated creative_fatigue_get tool, but it is allowlist-only — if it errors (not allowlisted), fall back to the report-metric method below, which is the primary path. The optional library scan (④) is off the critical path: if it errors or is empty, say "No unused library assets found" and continue — never block. ⚠️ Paywalled (n/a) metrics → show n/a, never guess.

Error codes

Honest layering: E101 / E103 are skill states; E102 / E104 / E105 / E106 wrap real platform errors (40001, FAILED, 40002, 5xx) and MUST surface the raw code + message.

═══════════════════════════════════════════

STAGE 0 — Resolve the advertiser

═══════════════════════════════════════════

Fast path (preferred): if the user gives an advertiser_id (or bc_id), use it directly — some tokens can't enumerate accounts:

Paste the advertiser ID whose creatives I should check (a long numeric string).

Else try to list: auth_advertiser_get; if empty, bc_get; if both empty and no id → E101_NO_ACCOUNT, ask for an id. Confirm: advertiser_info_get([advertiser_id]) → keep currency, timezone; 40001 → E102_NO_PERMISSION.

═══════════════════════════════════════════

STAGE 1 — Confirm windows + options

═══════════════════════════════════════════

Creative fatigue settings:
  Recent window:   last 3 days (default)
  Baseline window: the 30 days BEFORE the recent window (default; NON-overlapping)
  Library scan:    suggest fresh videos/images to swap in?  (yes / no, default yes)
Default: 3-day recent vs prior-30-day baseline, library scan on. Tell me if you want to change it.

Store recent_days (3), baseline_days (30), library_scan. Date sanity: baseline end_date ≤ today; reject future dates before any call.

═══════════════════════════════════════════

STAGE 2 — Pull data (read-only)

═══════════════════════════════════════════

Checking your creatives for fatigue, this usually takes 20–40 seconds...

① Recent report (sync) — report_integrated_get (L0): advertiser_id, report_type = BASIC, data_level = AUCTION_AD, dimensions = ["ad_id"], metrics = ["spend","impressions","clicks","ctr","conversion","conversion_rate","cpm","video_play_actions","ad_name","campaign_name"], recent date range. Parse string metrics to numbers. cpm + video_play_actions are required — many accounts are video-views/reach ads whose only valid engagement metric is cost-per-view or CPM (verified live: most ads on 7415418166391832593, 7595534274158903297, 7631551715865657362 are video-views).

② Baseline report (async) — ad level over the prior non-overlapping baseline window (default the 30 days ending the day before the recent window), same metrics as ① → report_task_create → poll report_task_check → on success report_task_download. On FAILED → E104.

③ Ad ↔ creative names — smart_plus_ad_get (L0) and/or ad_get (tool_execute) to map ad_id → ad/creative name + material id.

④ (optional) Library scan — only if library_scan = yes: file_video_ad_search / file_image_ad_search. If it errors or is empty → "No unused library assets found," continue.

If no ad clears the data floor (Stage 3) → E103_NO_ADS, stop honestly.

═══════════════════════════════════════════

STAGE 3 — Measure decay + classify (two complementary signals)

═══════════════════════════════════════════

📎 Exact thresholds, formulas, exclusions, and the creative-churn fallback live in fatigue-criteria.md. Summary contract below.

0. Data floor + engagement metric by objective: an ad needs volume to judge (Signal A: baseline impressions ≥ 1000; Signal B: recent impressions ≥ 1000). Below both → "thin data," skipped. Pick each ad's engagement metric by objective: conversion ad (conversion objective or records conversions) → CVR / CPA; TRAFFIC / BRAND_CONSIDERATION / ENGAGEMENT → CTR; VIDEO_VIEWS → cost-per-view (spend / video_play_actions) or view-rate; REACH → CPM. Never judge a VIDEO_VIEWS / REACH ad on CTR (they have ~0 clicks by design — verified live: NUXE / vivo / CP video ads), and never a CVR an awareness ad lacks. One advertiser often mixes objectives (traffic + video-views + live-conversion ads together — verified live: 7119328366494220289, 7591834723639656465), so pick the metric per ad, not per account — but anchor on the campaign (group by campaign_name, already in the pull): all ads in one campaign share an objective. If you didn't fetch the campaign objective, infer it from the metric signature: conversion > 0 → conversion ad (CVR/CPA); video_play_actions ≫ clicks and ctr ≈ 0 → video-views/reach (cost-per-view / CPM, never CTR); video_play_actions = 0 with real clicks → click objective (CTR / CPC). ⚠️ A near-zero-CTR ad inside a click/consideration campaign (where other ads earn real CTR) is a bottom-tier-CTR Retire candidate, NOT a "cheap view" — only treat an ad as video-views/reach when its whole campaign optimizes for views/reach. Verified live: Keyshu consideration campaign spans CTR 0.05%→4.10% across same-objective ads (the 0.05% ad is a Retire, not a cheap view).

1. Signal A — time-decay (preferred, precise): for an ad present in BOTH windows — engagement-metric fatigue = the ad's engagement metric (step 0) worsened ≥ 40% vs its baseline (CTR or view-rate ↓ ≥ 40%, or cost-per-view / CPM ↑ ≥ 40%); CVR fatigue (conversion ads) = recent_cvr < baseline_cvr × 0.7.

2. Signal B — vs current peers (always available; handles creative churn): many advertisers rotate creatives fast (e.g. flash-sale bursts), so few ads persist across windows and Signal A finds little. For each CURRENT ad (recent impressions ≥ 1000), rank it against the account's current same-objective ads on its engagement metric (CTR / view-rate / CPM per step 0) plus recent_cvr / recent CPA for conversion ads.

3. Coverage + mode (report honestly): count ads assessable by Signal A (cross-window history) vs only by Signal B (fresh). If most ads are fresh → say "this account rotates creatives quickly — showing current winners/losers rather than time-decay."

4. Classify each live ad into ONE bucket (tag the WHY):

• Retire 🔴 — time-decay fatigue (A) OR a conversion ad with meaningful spend + clicks but conversion = 0 (clicks that don't convert) OR bottom-third vs current same-objective peers (B: its engagement metric — and CVR/CPA for conversion ads — well below the median); AND recent_spend ≥ floor.
• Scale 🟢 — top-third vs current peers (strong engagement metric; conversion ads: high CVR / low CPA), not decaying.
• Watch 🟡 — mild decline, mid-pack, or thin data.

5. Spend at risk = Σ recent_spend over Retire ads. Zero Retire → "No creatives to rotate right now." Tag each Retire with its reason: decayed / underperforms current peers / clicks-don't-convert.

═══════════════════════════════════════════

STAGE 4 — Render the rotation plan

═══════════════════════════════════════════

🎬 Creative Fatigue & Rotation Plan — {advertiser_name} ({advertiser_id})  ·  {currency}
Recent {recent_days}d vs baseline {baseline_days}d  ·  Ads analyzed: {N}  ·  Skipped (thin data <1000 impr): {M}
Mode: {time-decay ({A} ads have cross-window history)  /  current winners-losers (this account rotates creatives fast)}

★ Spend at risk (recent {recent_days}d on Retire creatives): {spend_at_risk}

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 🔴 Retire — fatiguing, rotate out soon
  #  Ad / creative   Why (decayed / underperforms peers / no-conversions)   Recent CTR/CVR   vs baseline-or-peer   {recent}d spend  → manage-creative
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 🟢 Scale — still winning, put more behind these → optimize-budget
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 🟡 Watch — early decline / keep an eye on
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ♻️ Fresh candidates to rotate in (optional, from your library)
  {if off/empty → "No unused library assets found."}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Assumptions & limits
  · Fatigue = recent CTR < 60% of the 30-day baseline; for conversion ads also recent CVR < 70% of baseline.
    Awareness ads (no conversions) are judged on CTR only.
  · Excluded: ads with < 1000 baseline impressions ({M} skipped). "Spend at risk" is recent-window
    spend on Retire ads, not a guaranteed loss. Candidate list is suggestion-only and may be empty.
    Paywalled metrics show n/a.

🔍 Next steps (this skill changes nothing itself):
  · Swap / retire / upload creatives → manage-creative   · Budget behind winners → optimize-budget
  · Full creative numbers → get-performance-report        · Delivery issues → diagnose-campaign-health

Sort Retire by recent_spend desc (★ on the biggest). If every bucket is empty, say so and still show analyzed/skipped counts + limits.

Worked examples

• Happy path (clear fatigue, real spend-at-risk): examples/01-happy-path.md
• "Plan and swap" redline (must route the swap out): examples/02-scope-redline-swap-attempt.md
• Thin data (no fabrication): examples/03-insufficient-data.md

Additional resources

• Fatigue thresholds, formulas, exclusions: fatigue-criteria.md
• Report parameters / real metric names / sync-vs-async: params-quickref.md
• Test suite (eval contract + cases): evals/README.md, evals/cases.jsonl