gooseworks-ai/goose-skills / skills/marketing/analyze-performance
analyze-performance
VerifiedWeekly performance read across every channel we run: what worked, why, and what to make next, judged on the business's real number (store revenue, qualified pipeline, retained users), not platform claims. Use when: write my Monday report; how are my ads doing — which ad won?; are my ads actually making money (Meta says 4x ROAS but revenue is flat); why did sales drop this week; did the test win; is my tracking working; monthly numbers for the replan; anything on fire today. Not for acting on the recommendations (pausing, scaling, launching) — use run-paid-campaigns; not for replanning the budget — use plan-campaign.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 15.0 KB |
| references/creative-and-channels.md | markdown | 7.0 KB |
| references/diagnosis.md | markdown | 5.5 KB |
| references/local-mode.md | markdown | 2.1 KB |
| references/measurement-contract.md | markdown | 10.3 KB |
| references/monday-report.md | markdown | 8.9 KB |
| references/monthly-and-pulse.md | markdown | 4.4 KB |
| references/revenue-join.md | markdown | 4.6 KB |
| references/run-memory.md | markdown | 14.7 KB |
| references/tests-and-learnings.md | markdown | 5.5 KB |
| skill.meta.json | config | 5.9 KB |
| tests/test-by-hand.md | markdown | 6.9 KB |
references/tests-and-learnings.md
View on GitHub# Calling tests and keeping the learnings log
Load at step 5 and at step 7 (write-back). This playbook is the one owner of the learn leg; the log is what every planner and maker reads (`brand_read` learnings, `knowledge_search` `source_types: ["learning"]`), and it feeds creative generation directly — a noisy "do" steers every future ad.
## Calling a test
Read the test's row in the campaign brief's **Test plan** (hypothesis, variable, cells, readable threshold, end date, primary metric, win/lose/unclear actions). Call it only when it has ended or reached its readable threshold.
| Test | Success criteria (from the brief) | Result | Call | How sure | Action (from the brief's rule) |
|---|---|---|---|---|---|
| Hidden pore-cloggers vs award-proof hooks | CPA ≥ 20% lower, ≥ 40 purchases/arm | −34%; 94 vs 45 purchases; store new orders rose with it | **Win** | likely (one SKU, Goose ads only) | Release reserve to the winning angle → run-paid-campaigns |
- **Win / Loss / Unclear** — against the written criteria only. Under ~30 conversions per arm, Unclear unless the gap exceeds ~50%; reliably detecting a ~20% lift needs ~100+ per arm (rules of thumb).
- Use the primary metric the brief named, on store-joined data where available; a platform-only read caps it at likely.
- A messy result: say whether it is enough to act on or the test should run again (the brief's Unclear rule) — the marketer decides.
- Negative results are learnings too.
- A Win or Loss is a **measured result**: saved as a readout (kind 23 `experiment_readout.<campaign-slug>-t<n>`) through save-learning's readout path in a run with the marketer present (`run-memory.md` § Experiment readouts). Unclear is not.
## Calls carried forward (calibration)
Every report opens its calls section with last week's calls and what happened — run-memory's outcome check (`run-memory.md`): status from a source (`done` / `not done` / `unknown` / `dropped`), then the result. *"Scale the UGC unboxing ad (we said, confirmed) → done 7 Oct; the paid run's read: worked, $31 per extra order vs $40 (Meta's count); store new orders rose 9% over the same days — right." "Raise Sales – Broad 40% → done 14 Oct; too early, readable 28 Oct." "Kill award statics (likely) → not done; still losing."* A money move run-paid-campaigns made is judged with its verdict and baseline recipe, plus this playbook's store-joined check over the same window; never re-measured here on platform numbers alone. Without this nobody can tell whether the calls are any good.
## How sure (on every claim and call)
The three words of the `write-report` skill (`../../write-report/SKILL.md`) §1, in the report and in the run record; never high / medium / low or a score (older records that say high / medium / low mean confirmed / likely / first look).
- **Confirmed** — mature data, n large enough, store-joined, cause confirmed, no conflicting signal.
- **Likely** — one of: platform-only, modest n, single segment, cause partly confirmed.
- **First look** — immature, small n, claims-only mode, or conflicting stories. First-look kill/scale calls are labelled "watch", not acted on.
Each call also carries **"would change if"** — the evidence that would reverse it.
## The learnings log
Written through the shared atom the `save-learning` skill (`../../save-learning/SKILL.md`), which picks the Brand Brain path or the older path and holds their mechanics: the marketer's own words as the statement, one idea per learning, read back to confirm, own superseded learnings retired.
**One fact = one line**, starting `[learning YYYY-MM-DD]` (plus ` A-NN` when it is about an angle-bank angle):
```
[learning 2026-10-13 A-07] Open cold Meta hooks for Unplug strips (US) with hidden pore-cloggers, not awards · likely · review 2027-01-13
```
The full evidence (test, dates, ads, spend, CPAs, scope, what it supersedes) stays in the weekly report's Tests section, which is the long-form record — the fact points back to it by date.
**Kind:** `insight` by default; `do` / `dont` only for a rule the marketer states as a rule in their reply; a first-look finding is never written.
**What qualifies** (anything else stays in the report, not the log):
- an **ended test**, called; or a pattern holding **2+ weeks that clears the same n floors as a test call** (≥ ~30 conversions per arm, and for DTC with UTM-joined store orders or the account-level new-customer number moving with it; or a gap > ~50%) — B2B's one or two SQLs a week and app CPI never qualify; or a confirmed channel surprise; or a confirmed failure with its reason.
- Never weekly snapshots ("CTR was 1.2%" is a metric, not a learning).
- One rule per fact, scoped narrowly (channel, product, audience, market).
- **It must have appeared in a report the marketer saw.** Learnings are *proposed* in the report's decisions block and written only after the marketer's reply — and not those they strike.
**Show what it would retire.** Each proposed learning carries "retires: <old fact>" when it supersedes one, so a yes to it is a yes to the retirement too.
**Hygiene:**
- Before writing, look for the same claim (the atom's read step: `brand_knowledge_search` / `brand_knowledge_browse` kind 23 where listed, else `knowledge_search` `["learning"]`): update the existing entry rather than add a near-duplicate.
- When a new learning contradicts an old one of ours, retire the old one as the atom says.
- When a fact's review date has passed, propose re-confirm / retire in the next decisions block.
- Promote `insight` → `do`/`dont` only when the marketer states it as a rule.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| finishing-a-development-branchobra/superpowersUse when implementation is complete, all tests pass, and you need to decide how to integrate the work | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| systematic-debuggingobra/superpowersUse when encountering any bug, test failure, or unexpected behavior, before proposing fixes | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| test-driven-developmentobra/superpowersUse when implementing any feature or bugfix, before writing implementation code | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| skill-creatoranthropics/skillsCreate new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |