Performance Analyzer

Instructions

Work the steps below as one dependency-aware pass. Each fill-in template lives in references/analysis-templates.md. Build the Step 2 shell after intake, but run Step 7 before populating or publishing Step 2 Conversions, Revenue, or any rate/cost that depends on them; those fields must cite Step 7's reconciled counted total or remain Unknown/NEEDS_INPUT.

1. Gather performance data — log campaign/period/influencers/platforms and the available sources (native analytics, influencer reports, web analytics, sales, promo codes). Template: step 1.
2. Analyze core metrics — compare reach, impressions, engagements, ER, video views, clicks, promo uses, conversions, and revenue against compatible source-dated targets/benchmarks. Emit field-level comparison states; do not invent an aggregate score or adjective verdict. Template: step 2.
3. Analyze by platform — compare platforms on compatible reach/ER/click/conversion/CPA windows and state observed differences. Put any explanation in a separately labeled hypothesis unless a designed comparison supports it. Template: step 3.
4. Analyze by creator — use opaque creator_ref; rank only comparable rows under the declared rule. Consume ROI/ROAS only from a cited roi-calculator artifact, do not compute it here, and separate observed content anatomy from causal hypotheses. A renew/retest/retire call comes only from the Retro decision gate. Template: step 4.
5. Content performance analysis — compare formats/themes under compatible exposure and attribution bases. Name observed higher/lower associations; describe a hook/message/visual as causal or "winning" only when the supplied design clears the measurement protocol. Template: step 5.
6. Engagement quality analysis — break engagement by type/intent, run evidenced comment sentiment, and surface purchase-intent signals. Use typed observations or Unknown; emit no /10 quality score without a supplied rubric, inputs, and calculation. Template: step 6.
7. Conversion & attribution analysis — draw the observed funnel and use one declared attribution model. Deduplicate events into mutually exclusive counted buckets; preserve promo/UTM/direct overlap as reconciliation evidence, and report Estimated influence outside the counted total. Template: step 7.
8. Generate insights & recommendations — write 3–5 evidence-backed observations, separately labeled hypotheses, and bounded next tests. Add one compact Campaign Retro Card per creator decision requested only when a verified current, non-forked tracker-state artifact proves that exact campaign/creator is measured or closed and the matching measurement-contract and decision-rule refs are supplied; a bare stage string never qualifies. Use only renew | retest | retire | unknown. Template: step 8.

Before naming any creator/format/platform a real winner, clear the comparability, complete-scope, preregistered-rule, and significance bars in measurement-protocol.md — otherwise mark it Keep-testing or NOT_RANKED. When a structured score is needed, apply per-dimension STAR analysis (Suitability/Trust/Appeal/Return dimension reads) from star-benchmark.md, and hand financial inputs to roi-calculator for Return (R) math — this skill contributes inputs but does not compute ROI/ROAS or SQS (the creator-content-auditor gate computes SQS).

For the Retro Card, use renew only when comparable measured evidence clears the preregistered decision rule without a material unresolved limitation; use retest for a plausible but inconclusive or correctable test; use retire only when measured evidence or a documented hard constraint clears the declared stop rule; otherwise use unknown. This operating decision is not a STAR dimension, SQS, or creator-content-auditor verdict—do not simulate or carry forward one.

After an authorized WARM save, offer a handoff to campaign-planner to append the saved analysis/card reference to the relevant tracker row's evidence_refs; the tracker edit needs its own exact authorization, and neither the card nor this skill advances stage. Also offer fit-scorer as an explicit next-cycle handoff with the card's evidence references and hypothesis. Do not invoke it automatically, and do not translate the Retro decision into a STAR/SQS verdict.

Example

User: "Analyze this dated summer-skincare export for 10 creators. It contains opaque creator refs, the metric/target table below, per-creator and per-platform results, one deduplicated attribution model, and a completed significance read. ROI comes from roi-calculator artifact roi-ref-01."

Output (abridged — full version in references/analysis-templates.md):

# Summer Skincare Campaign Performance Analysis — illustrative export-backed read

| Metric | Result | Target | Status |
|--------|--------|--------|--------|
| Total Reach | 2.4M | 2M | ✅ +20% |
| Engagement Rate | 4.2% | 3.5% | ✅ +20% |
| Conversions | 1,847 | 2,000 | ⚠️ -8% |
| Revenue | $142,500 | $150,000 | ⚠️ -5% |
| ROAS (from `roi-ref-01`) | 2.8:1 | 3:1 | ⚠️ -7% |

**Top 3**: the three `creator_ref` rows that clear the declared ranking and significance rule, using only comparable metrics in the export.
**Key learning**: report the export-backed TikTok/Instagram delta only if the comparison windows and attribution bases match; otherwise mark it Keep-testing.
**Recommendation**: renew/drop and reallocation calls remain conditional on the predeclared decision rule rather than invented from the campaign count alone.

Reference Materials

• references/analysis-templates.md — the eight fill-in step templates plus the full worked example.
• skill-contract.md — shared contract and handoff format.
• state-model.md — memory tiers and save-path conventions.
• CONNECTORS.md — verified free/keyless data recipes per connector category.
• measurement-protocol.md — preregistered readback windows, outcome unit, alpha, practical-effect boundary, multiplicity/sequential policy, guardrails, and decision owner. Report statistical and practical flags separately; use experiment.py for deterministic Calculated evidence, and never substitute a universal p-value/lift rule or attribute a business action to the helper.
• The STAR benchmark at references/star-benchmark.md — scoring architecture when a structured score is needed.
• Sibling skills: roi-calculator, report-generator, fit-scorer, campaign-planner.

Next Best Skill

Primary: roi-calculator — convert measured performance into dollar-level ROI, cost-per-result, and payback math.

Alternates (same Report family):

• report-generator — package the analysis into a formal stakeholder report.
• fit-scorer — feed proven performers back into creator scoring for the next round.

Termination note: Maintain a visited-set. If a skill has already been invoked this session, stop and report chain-complete rather than re-running it. Cap the chain at max-depth 3 hops; if results are inconclusive after that, surface the open loops to the user instead of continuing.