gooseworks-ai/goose-skills / drafts/marketing/learn-my-brand
learn-my-brand
VerifiedLearn my brand once and stop asking me every time. Use when someone says learn my brand, set up my brand, here's our brand book, remember how we talk, what can we claim, can we say this, approve our claims, sign off our voice, our price or product changed, the client approved the claims, pick up where we left off, or check our claims are still current. Reads what GooseWorks already found about the brand, asks only the gaps as drafts to confirm, audits every claim (including ones research marked approved) and gets the ones that matter signed off with proof, scope and review date, turns the voice into rules the owner confirms, checks logo, colours and fonts, and saves it all as it goes so every other job loads it instead of asking. Ends with an unknowns list and a read-back. Not for who our customers are and what they say (use research-audience) or writing positioning from scratch (not covered yet).
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 14.5 KB |
| atoms/brain-write-back/ATOM.md | markdown | 6.8 KB |
| atoms/claims-audit/ATOM.md | markdown | 3.9 KB |
| atoms/voice-calibrate/ATOM.md | markdown | 3.1 KB |
| references/brain-map.md | markdown | 10.2 KB |
| references/brand-record-template.md | markdown | 6.2 KB |
| references/category-rules.md | markdown | 9.5 KB |
| references/claims-craft.md | markdown | 13.4 KB |
| references/intake-craft.md | markdown | 7.8 KB |
| references/voice-craft.md | markdown | 6.4 KB |
| skill.meta.json | config | 4.5 KB |
| tests/test-by-hand.md | markdown | 12.0 KB |
tests/test-by-hand.md
View on GitHub# Test by hand: learn-my-brand
## Setup
1. Connect GooseWorks **staging** to your agent.
2. Link the playbook: `ln -s "$PWD/drafts/marketing/learn-my-brand" ~/.claude/skills/learn-my-brand`
3. Pick a brand the team knows deeply, with the documents and links the founder would really hand over. Run it twice more on: one regulated brand (skincare or supplements; staging's Clearday works) and one B2B brand. Not only a well-known DTC brand.
## Part 1: the intake run
Say "Learn my brand" and answer as the founder would.
It must:
- Read the brain first and ask for documents before any question; import what you hand over.
- Ask only gaps, as drafts to confirm; about 15–20 decisions and 20–30 minutes in total (basics 5–8 decisions); under time pressure, the shortcuts in SKILL.md's time budget. Never ask a fact the site, the brain or a document already holds.
- Surface at least one contradiction if one exists (site vs ads, old vs new) with both sides and a source; never resolve it silently.
- Treat research's "approved" claims as candidates: on Clearday, flag at least "Rated 4.9 out of 5 stars 234 Reviews" (needs a capture date and review-by), the press quotes (need article and date), and the value props "100% free of acne-triggering ingredients" and "Dermatologist-developed & 6x award-winning" (need proof or rewording).
- Load the skincare rule pack and say once that this is not legal advice.
- Stop twice for sign-off: claims (by risk tier, high-risk one at a time with proof or "no proof on file") and voice (triad test, keep-or-kill, then the whole guide).
- Record unknowns with the playbook that will ask, and never guess.
- End with a read-back screen and three probes answered from a fresh read of the brain.
Fails if: it asks a known fact, invents a claim, proof, customer or number, accepts one "ok" for a long claims list, writes an unconfirmed fact as the owner's correction, or never stops for sign-off.
### Variant: the approver is not in the chat
Say "I'm the agency; the founder approves claims." Claims must end as "pending client approval", with a forwardable sign-off sheet, and the marker must carry `pending-client` (nothing beyond plain product facts approved).
### Variant: time pressure and resume
Halfway through the claims step say "I've got five minutes left." It must switch to the product-facts-only default, save what is confirmed, mark the record "in progress, resume at step N", and stop cleanly. In a new session, say "let's finish the brand setup": it must resume at step N without re-asking anything saved.
### Variant: client sign-off comes back
After the agency variant, paste: "From Dana (CEO), email today: approve D1 and D3, reject D2, voice guide is fine." It must map each answer to its row, record "Dana via <agency user>, email, <date>" with the words verbatim, write the approved claims and the marker, and leave D2 as not approved.
### Variant: pre-launch with no website
"Learn my brand — we launch our habit app in three weeks; there's only a TestFlight link and my Twitter." It must use the placeholder brand (not create a second one), not ask for a website setup it can't do, ask for 5–10 of the founder's posts as the voice sample, and default the launch date to embargoed.
### Variant: update mode
After a full run, say "Our free-shipping threshold changed to $35." It must ask whether the site already shows $35, touch only that fact (the one `Claims you may use` fact holding it, the brand record with the old line under Retired, and the marker re-written naming the new version — not 20 product rows), not re-run the intake, and list which saved outputs of other playbooks now rest on the old threshold.
### Variant: one-claim mode
Say "make-static-ads cut the line 'clinically proven to reduce breakouts in 7 days' — can we approve it?" It must run the claims gate for that line only: classify (clinical + efficacy, high risk, drug-claim territory for a cosmetic in the US), ask for proof, and recommend reject or a safe rewording.
### Variant: scoped sign-off from a maker
In a fresh session with no sign-off on file, ask for three static ads. The maker runs `../../../../skills/marketing/claims-gate/SKILL.md` §4: it asks the owner to approve the 3–5 lines the ads need (each with type, proof on file, recommendation), records the approved ones as one `Claims you may use —` fact each plus a `signed-off` marker naming a new Brand record version, and uses only those lines. Then run this playbook's read-back: the scoped lines show as approved; everything else stays a candidate.
### Variant: client approver on a brand that is already signed off
After the client sign-off variant (some claims approved, `approver: client Dana`), have the agency user ask a maker for ads needing a new benefit line. claims-gate must read `approver: client Dana`, not ask the agency user, put the line on the sign-off sheet in a new record version, and keep the marker `signed-off` (the approved claims stay usable). It fails if the marker turns `pending-client` or the agency user's yes is recorded as approval.
## Part 2: the fresh-session check (the spec's Done-when)
Open a new session that reads only the brand brain (no pack file, no earlier chat). Ask ten questions. Pass: ≥ 9 correct with no question bounced to the owner, and zero hard fails. Hard fails: any unapproved, expired, out-of-scope or strengthened claim; any invented fact. "Unknown, assuming X" from the ledger counts as correct; a guess is a fail.
1. What is the hero product, the standing offer, its price and where it is sold?
2. Who is the customer and what is their top pain? (Top objection: "unknown — asked when we research your buyers" is a correct answer if that research hasn't been done.)
3. Our one-line positioning (or: is it unknown, and who will write it)?
4. Give three claims I can use in a US Meta ad for <product>, verbatim, with qualifiers.
5. Trap: can I say "clinically proven to clear acne" / "dermatologist recommended"? (Expected: no, why, and the approved alternative.)
6. Staleness: can I use the star rating and the press quote today? (Expected: checks the review-by date.)
7. Scope: can the US free-shipping line run in a UK email?
8. Voice: rewrite "Elevate your skincare journey with our game-changing formula" as an email subject, a TikTok caption and a support reply. (Owner blind-picks against a generic rewrite; banned-words check must pass.)
9. What must we never say, and who approves a new claim?
10. What logo file, colours and fonts do we use, what is missing, and what unknown would block a launch email next week?
Adversarial probes in the same session: ask for a line that uses a rejected claim; ask for a line using a banned word. Both must be refused or flagged with the approved alternative.
## Staging run log
See the bottom of this file (filled in by the builder session).
---
## Staging run log — 2026-10-16 (builder session, GooseWorks staging, brand: Clearday)
Ran steps 1–2 for real (all free reads). Stopped before the first write (step 7), because writes would change a staging brand that other sessions are testing against. The writes are listed below as "would have run".
**Step 1 — reads (all returned, all free):**
- `brand_read` sections summary, kit, products, competitors, learnings, onboarding, accounts, sources, `detail: full` → 143 KB answer (too big for one read; the agent must read it in parts: kit full alone is ~70 KB; lean summary/kit/products/onboarding/learnings/accounts is ~10 KB, but lean competitors and sources add ~50 KB, and an unfiltered `media_list` is ~90 KB; the playbook now reads kit full once, competitors via `competitor_read`, sources only when needed, and media by kind). Kit: voice is a prose paragraph that already holds research's own Do/Don't lines and three site lines; tone and aesthetic adjectives; 5 value props; 6 proof points; 7 approved claims; palette, Inter typography (mapped, not the real font: kit fonts say "Exposure"); 20 products; onboarding profile empty (role, ad spend, goals, platforms missing); 1 learning (a `dont` from the user about video scripts); accounts none; 17 saved sources, all `confirmed: false`, incl. the brand's own TikTok.
- `knowledge_search` `["claim"]` → 7 "approved" claims, all created by research, `applicability: unknown`: "Rated 4.9 out of 5 stars 234 Reviews", "Free Returns", "Free Shipping on Orders $45+", four press quotes (Allure ×2, Byrdie, NewBeauty).
- `knowledge_search` `["document"]` "Brand record approved claims brand book" → empty (no brand record yet: a first run).
- `knowledge_search` `["evidence"]` → the same four press quotes as evidence rows, `approved_ad_claim: false` (so the brain holds them twice: as evidence not approved, and as claims approved).
- `competitor_read view: suggestions` → 5 unconfirmed (Skinfix, Naturium, Sofie Pavitt Face, La Roche-Posay unsure, Topicals unsure); tracked roster has 5+ from research.
- `media_list kind: logo` → 2 brand-scope logos plus 3 copies inside video projects (→ fixed: the playbook now passes `scope: "brand"`).
- `account_whoami` → `user_context` all null. (Round 2: no longer read; the founder's situation now comes from `brand/situation.md`.)
**What step 4 would flag on this brand (from the real data):**
- Value prop "100% free of acne-triggering ingredients — every single product": absolute + implied efficacy → reword ("formulated without <named list> of common pore-clogging ingredients") or proof.
- "Dermatologist-developed & 6x award-winning formulas": needs the named dermatologist and the six awards with years.
- "Clinically safety-tested, non-comedogenic": needs the test reports; "non-comedogenic" needs a comedogenicity test.
- Site line in the kit voice "Acne-safe, deep hydration that also prevents breakouts": "prevents breakouts" is US drug-claim territory for a cosmetic → never-say with alternative.
- Description "created by actress Lili Reinhart": a celebrity endorsement/association → needs the owner's confirmation of how it may be used in paid media.
- Rating "4.9 out of 5 stars 234 Reviews": needs platform, capture date, review-by (90 days).
- Press quotes: need article URLs and dates; paid-use permission for publisher names.
- "Free Shipping on Orders $45+": offer; needs current threshold and a review date.
**Would have run (writes, not executed):**
- `brand_update` `knowledge_intent: user_correction` + owner's words: `patch.kit` {voice, audience, description, value_props (rewritten), brand_type}; `patch.knowledge` {positioning, tagline}; `patch.products` offers; `patch.onboarding` {role, platforms}.
- `brand_update` `patch.facts`, one call through `../../../../skills/marketing/save-learning/SKILL.md`: retire superseded facts; add `Not public —` (dont), `Banned words —` (dont, owner's words first), `Policy:` (must, operating rules only), one `Claims you may use — C<n> …` (must) per approved claim (free shipping and free returns included), one `Never say —` (dont) per never-say line, and the marker (`Claims sign-off (learn-my-brand, <date>, signed-off, approver: owner, voice: provisional, record "…")`). One idea per fact, read back and compared.
- Brand record through `../../../../skills/marketing/save-to-brain/SKILL.md`: `file_write` (`brand/records/2026-10-16-v1.md`, brand coworker scope) + `media_upload` (scope brand, kind document, source workspace) + `knowledge_import` (`media_asset_id`, title "Brand record — Clearday — 2026-10-16 v1") + `job_get` (`kind: knowledge_ingest`). (Route changed from inline bytes to the workspace after round 3, for hosts with no shell.)
**Not verified on staging.** Creating a throwaway staging brand for a write test was blocked by this session's permission check (shared resource). Settled since by the cross-set contracts review (server code): facts need `knowledge_intent: user_correction` (refused without it), are stored as the user's and never overwritten by the distiller; deleted media drops out of search. (Today's text-length cut is handled inside `../../../../skills/marketing/save-learning/SKILL.md` until the Brand Brain rebuild removes it.) Still unverified: whether retired facts are hidden from every read, and which voice field `kit.voice` shows.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first. | Stars: 196k | Updated | Claude CodeCodex | Verified | Install |