The standing creative feedback loop for every pipeline (format-* skills, cartoon-h3, ugc-omni, vo-broll, dawn-statics, Dawn natives): PLAN the next concepts as combos from what is winning (stakes x emotion x payoff x root cause x mechanism x authority/POV x angle x device), TAG each pack/job at creation with combo_* keys, LAUNCH under that ad-name prefix, READ Meta weekly (read-only), RECODE only new concepts against the fixed codebook, UPDATE the map, recipes and the diff (incl. song vs static frequency/decay).

Lane: strategy (feeds every format-* pack) · not built yet: 0 (see below) · source: /Users/ayden/.openclaw/workspace/skills/workflow-combo-loop

workflow-combo-loop — PLAN → TAG → LAUNCH → READ → RECODE → UPDATE

The 10-05 Dawn combo map (output/dawn-combo-map-2026-10-05/MAP.md) was a one-off. This skill makes it a loop every creative pipeline runs through: concepts are planned as one-variable combos off a paying parent, tagged at birth so the read never has to guess, and the map refreshes itself weekly from Meta (read-only) so the next plan starts from the newest winners, not last week's.

1. Read first

knowledge/ad-formats/_FORMAT-CORE.md (§1 doctrine, §6 launch + read). Binding. knowledge/ad-formats/_MODEL-REGISTRY.md (Sonnet = coding role).
output/combo-loop/dawn/latest/MAP.md: the current map, DIFF vs the last snapshot, frequency/decay per family, recipes.
scripts/combo_loop/codebook_dawn.md: the codebook (dimensions + allowed values). Same codebook for tags and for coding.
references/procedure.md: commands, file layout, the ad-name ↔ tag mapping, how recipes/confidence are computed, an ILLUSTRATIVE run. references/sources.md: where this came from and the receipts.
Stake bank: output/creative-strategy-cooper-2026-10-05/dawn-stakes.md.

2. When to use / when not

Use: before any new Dawn concept in any pipeline (Step 0 of every format-/workflow- pack spec); when Fish asks what to test next, what's winning by dimension, or whether a format's frequency/decay holds; weekly refresh; tagging a pack/job before launch.
Not: judging a single ad set (that is media-buyer: 7-day rule, never pause); writing copy (format beat sheet → copy routing in core §1); the hub atlas avatar × angle labels (atlas_label.py, still used by dawn-native-brief; this loop codes 12 dimensions, the atlas 2).
Curvelle: config exists in scripts/combo_loop/brands.json but refresh refuses until it has its own codebook + breakeven ROAS (media-buyer owns the number).

3. Inputs

output/combo-loop/<brand>/latest/recipes.json from the last refresh (plan step).
The format skill name (--format format-song, dawn-statics, dawn-native, …) and the brand.
For refresh: META_SYSTEM_TOKEN in the workspace .env (GET only), the claude CLI for Sonnet coding.
For tags: the pack (output/concept-packs/<slug>.md), the ugc-omni/yapper job.json, the ui-mockup spec.json, or a hub ads.json entry with meta_ad_name.

4. Outputs

output/combo-loop/<brand>/<date>/: MAP.md (DIFF, frequency + decay, recipes, coded-this-run, uncoded, invalid tags), recipes.json (per family: winning value per dimension with spend/ROAS/n/W/L/confidence, runner-up, avoid list, top/weak pairs, paying parents), units.json, agg.json, family.json, ads.json, weekly.json. latest → newest date.
output/combo-loop/<brand>/codes.json: cumulative codes (baseline 10-05 + Sonnet + tags never overwritten).
output/combo-loop/<brand>/plans/<date>-<format>.md: the plan block a pack pastes.
output/combo-loop/<brand>/tagged/<prefix>.md: registered tags for work with no pack file.

5. Process

PLAN (free, no network). python3 scripts/combo_loop/plan.py --format <skill> --brand dawn [--n 4] → 3-5 combos, each = a named paying parent (spend, ROAS, CPA) + ONE changed dimension: exploit (move toward the recipe winner), stake swap (a stake-bank row that lands on her), explore (a winner from another family never tried here). Thin families (video today) plan from the song recipe (story formats) or the static recipe (callouts) and say so. Pick with Fish; never stack two variables.
TAG (free). Paste the plan's yaml block into the pack front matter / job or spec JSON ("combo": {...}); set combo_ad_prefix to the exact string the launched Meta ad names will start with. No file (natives, statics) → python3 scripts/combo_loop/tags.py register --prefix N1006-A combo_avatar=... .... Gate: python3 scripts/combo_loop/tags.py check <file> PASS. tags.py keys lists allowed values.
LAUNCH. Normal launch path (core §6): own ad set per concept, preflight, dof guard, Fish's go. Ad names MUST start with combo_ad_prefix (copies/SCALE dupes keep the prefix; "LP Split · ", "LW · ", "- Copy", "- SCALE" are stripped before matching).
READ + RECODE + UPDATE (free Meta GETs; Sonnet via subscription). python3 scripts/combo_loop/refresh.py --brand dawn (weekly via launchd, or on demand): pulls lifetime + weekly per-ad insights (≤1 req/s), assigns every ad to a unit (tags > existing code by ad base name > existing code by body hash), codes only NEW concepts ≥$100 with the codebook verbatim (cap 40/run, sure/unsure per field), aggregates per family (song / video / static), writes the dated snapshot + DIFF vs the previous one. --no-code skips Sonnet; --offline re-aggregates saved pulls.
Review (Fish). Read the DIFF: recipe flips, cells that moved ≥0.05 ROAS, new winners/losers, frequency/decay per family. Check unsure fields on new codes before trusting a recipe that rests on them; correct a code by editing codes.json (set "c": "sure" with your value).
Next plan starts from the new latest/ (step 1). The loop closes.

6. Tools & models

Built 10-06: scripts/combo_loop/refresh.py, scripts/combo_loop/plan.py, scripts/combo_loop/tags.py, scripts/combo_loop/common.py, scripts/combo_loop/brands.json, scripts/combo_loop/codebook_dawn.md.
Coding role = Sonnet via scripts/claude_cli.py run_sonnet (OAuth subscription; ~1-2 min per 8 concepts). No paid API.
Meta: GET only (insights, creatives). The system user cannot read video length (#10 permission): video concepts are coded from the script (hub meta_ad_name scripts, 10-05 song lyrics) or primary text, and say so.
Schedule: output/combo-loop/com.fish.combo-loop.plist (weekly Monday 06:15, read-only, NOT installed; Fish installs).
Built 10-06 late: hub page for the map — scripts/build_pipeline_hub.py combo_map_md() → f/combo-map.html (nav link + foundations card on the pipeline hub), reads output/combo-loop/<brand>/latest/MAP.md per brand (dawn, curvelle). Discord DIFF poster — scripts/combo_loop/discord_post.py (default prints only; --post required to actually send, via DISCORD_BOT_TOKEN/DISCORD_PIPELINE_QA_CHANNEL). Curvelle codebook — scripts/combo_loop/codebook_curvelle.md (mirrors codebook_dawn.md's 14 fields from knowledge/brands/curvelle/*); brands.json curvelle.codebook points at it. breakeven_roas for Curvelle stays null (media-buyer's number) — refresh.py still refuses Curvelle until that lands.

7. Gates & QA checklist

[ ] Plan: each concept names ONE variable and a paying parent with CPA; stake lines come from the bank with their source tag (VOC grep rule, corrections 10-06).
[ ] Tags: tags.py check PASS; combo_ad_prefix equals the launch script's ad-name prefix (read the launch script before launch).
[ ] Refresh: MAP.md written, latest moved, Meta GET count printed, no token in logs; "Uncoded" list empty or explained by the cap.
[ ] Before quoting a recipe: confidence ≥ med, or say "signal, not proof" (songs: few spenders).
[ ] Never act on the map with a Meta write from the loop; launches and budget changes go through media-buyer + Fish.

8. Failure modes

Prefix drift: ads launched with a name that doesn't start with combo_ad_prefix fall back to Sonnet coding (lower confidence). Fix the tag's prefix, re-run refresh.
Two variables changed: the read is unattributable; the plan block lists "keep fixed" for that reason.
Proxy recipes: thin families (video) borrow the song/static recipe; a result there tests the format AND the variable. Read it as a format test first.
Body-hash drift: a static whose primary text changes becomes a new concept; that's correct (it is a new message), but its history stays under the old key.
Coding from captions: video concepts without a script are coded from the primary text; Sonnet marks those fields unsure. Add a hub script + meta_ad_name or tag the pack.
Small-n songs: 3-5 spenders carry the song recipe; the DIFF shows when it changes. Don't extend a song finding to statics or vice versa (MAP §2 "what changes between formats").

9. Sources

10-05 build: output/dawn-combo-map-2026-10-05/MAP.md + its scripts (promoted, not moved), plan ~/.claude/plans/dawn-combo-map-2026-10-05.md.
Fish 10-06: "map stakes / emotions / payoffs / root cause / mechanism and authority/pov and angles with each other … plan like that in the pipelines … and also have feedback loops".
references/sources.md for receipts; hub atlas precedent project_angle_atlas.md (labels at planning time, ad-name prefix).