| name | metrics-feedback-loop |
| description | Turn publish receipts and performance metrics into reusable style-profile or niche-profile updates that say what to repeat, kill, or test next. |
| allowedTools | ["read","write"] |
| model | inherit |
| argumentHint | {"receiptsPath":"output/content-machine/publish-prep/publish.json","metricsPath":"output/content-machine/performance/short-001.json","styleProfilePath":"output/content-machine/library/style-profiles.v1.json","nicheProfilePath":"profiles/niche.json","outputPath":"output/content-machine/learning/metrics-feedback.v1.json"} |
| inputs | [{"name":"receiptsPath","description":"Publish receipt, publish metadata, or publish-prep review artifact.","required":false},{"name":"metricsPath","description":"Performance metrics export or manual metrics JSON.","required":true},{"name":"styleProfilePath","description":"Optional existing style profile to update or annotate.","required":false},{"name":"nicheProfilePath","description":"Optional existing niche profile to update or annotate.","required":false},{"name":"outputPath","description":"Path for the feedback artifact or proposed profile patch.","required":false}] |
| outputs | [{"name":"metrics-feedback.v1.json","description":"Evidence-backed learning summary with repeat, kill, and test-next decisions."}] |
Metrics Feedback Loop
Use this after a short has been published and enough performance evidence
exists to learn from the result. The skill converts receipts, readiness
review notes, and platform metrics into reusable updates for future
style-profile-library or niche-profile-draft runs.
Use When
- A posted short has performance metrics and the user asks what to do
differently next time.
- A channel wants reusable style or niche-profile learning instead of
one-off postmortems.
- Publish receipts from
publish-prep-review should be
compared with actual outcomes.
- A winning short analyzed by
reverse-engineer-winner
needs to become a repeatable hypothesis, not just inspiration.
Core Rule
Optimize for meaningful outcomes over vanity metrics. Treat views,
likes, and impressions as context unless they connect to stronger
retention, completion, saves, shares, qualified comments, follows,
clicks, conversions, or audience fit.
Inputs To Gather
- publish receipt or
publish.json: title, description, platform,
posting time, checklist, known risks, and packaging choices
- performance metrics: impressions, view rate, average watch time,
retention curve, completion rate, rewatches, shares, saves, follows,
comments, click-through, conversion, and audience segment if available
- source artifacts: script, caption style, visual plan, render defaults,
style profile, niche profile, and reverse-engineered winner blueprint
- baseline: recent channel median or target benchmark for the same
platform, niche, duration band, and format
Workflow
- Normalize metrics against the most relevant baseline before judging.
- Separate distribution problems from creative problems: low
impressions alone does not prove the hook or style failed.
- Map each signal to a likely creative lever: hook, pacing, caption
density, visual novelty, proof, payoff, packaging, CTA, or niche fit.
- Produce three short decision lists:
repeat, kill, and
testNext.
- Convert stable lessons into proposed
style-profile-library updates:
caption rules, pacing rules, render defaults, visual constraints, and
reference notes.
- Convert audience or topic lessons into proposed
niche-profile-draft
updates: hook patterns, tone rules, forbidden phrases, proof types,
visual guidance, packaging angles, and thumbnail direction.
- Keep weak evidence as a hypothesis. Do not overwrite profiles from a
single noisy post unless the outcome is extreme or confirmed by
repeated posts.
Output Shape
Write or return a compact metrics-feedback.v1.json artifact:
{
"postId": "short-001",
"verdict": "scale | iterate | pause | inconclusive",
"meaningfulSignals": [],
"vanitySignals": [],
"repeat": [],
"kill": [],
"testNext": [],
"styleProfilePatch": {},
"nicheProfilePatch": {},
"evidence": [],
"confidence": "low | medium | high"
}
Decision Guidance
- Repeat hooks, pacing, caption behavior, visual structures, and CTAs
that beat the baseline on retention, completion, saves, shares,
follows, clicks, or conversions.
- Kill elements that correlate with early drop-off, poor completion,
confused comments, low saves, low shares, weak audience fit, or failed
readiness risks already flagged by
publish-prep-review.
- Test one or two variables at a time so the next result is learnable.
- Prefer profile patches that are specific enough to guide future
prompts but generic enough to survive new topics and providers.
- Preserve provenance: cite the metric, baseline, artifact path, or
reviewed winner that supports each decision.
Pair With
Constraints
- Do not store private audience data, account credentials, ad spend
exports, or platform access tokens in profiles.
- Do not claim causation from one post without clear evidence.
- Do not blindly chase views if the audience, saves, shares, follows, or
conversion quality is poor.
- Keep profile updates local, reviewable, and reversible.
Validation Checklist
- Every repeat, kill, or test-next item cites a metric or artifact.
- Vanity metrics are separated from meaningful outcome signals.
- The recommended profile patch is scoped to style or niche behavior,
not arbitrary project memory.
- Low-confidence lessons stay as tests instead of permanent rules.
- No provider-specific, private, or rights-sensitive data is stored.