write-report

Every playbook that hands over research, a plan, a strategy or a report writes it through this atom.

1. Write for a human

The 30-second test. A marketer who reads only the top knows what you looked at, what you found and how sure you are, and what to do first — and can judge the move from the findings right above it. If not, rewrite the top.

Shape, in this order:

1. Title: what it is, for which brand, the date. Example: "GooseWorks — market research, 18 Oct 2026".
2. In short, at most 150 words, in this order:
   • What we looked at: one line — scope, sources, date, and the main limit on the evidence (one review site, one country's search results, no access to the ad account).
   • 2–3 numbered key findings, each with its "so what" for this brand and how sure, in the body's own word: confirmed (two independent sources, or the brand's own data), likely (one good source or a clear pattern), first look (one sample, place or read; needs a re-check).
   • So, do this first (because of 1 and 2): the one move, who does it, by when, naming the findings it rests on. When they are only a first look, or the move needs something not yet confirmed, the first move closes that gap (confirm X, run the test, get the data).
   • Only when they apply, one line each: what changed since the last report; where a sibling report disagrees (below).
3. Decisions for you: a short dated box — each decision with your recommendation, the section it rests on ("see How they compare"), and the date it matters by, if any.
4. The story: context → what we found → what it means for you → what to do. Each section opens with its point in one sentence, then the evidence. Where a finding means nothing for this brand, say "nothing to do here" rather than invent a move.
5. What to do: at most 6 numbered moves, each with who does it, what it needs, and how we will know it worked.
6. What we're not doing, and why: the tempting moves you decided against, one line each — where herd-following and over-reaction get stopped.
7. What we couldn't check: one line each, with why.
8. Sources: links, dated, at the end, plus one line on source skew ("Most buyer quotes come from Trustpilot, which over-represents complaints").

Length: the body (everything after "In short") stays under 3,000 words. Cut repetition and detail no decision needs. An appendix may hold specialist tables; the report must work without it.

Agree with the sibling reports, or say why not. Before writing, read every same-brand report from the last 14 days whose job overlaps (market, competitors, audience, a campaign or social plan, outreach) — saved documents (brand_knowledge_search where listed, else knowledge_search; local mode: gooseworks-local/<brand-slug>/ instead) and the local gooseworks-reports/<brand-slug>/ folder. Compare lead buyer, first move, positioning and prices. A different lead buyer or first move is named in "In short" (anything else, in the story) and settled: merge the two first moves, or say why this one wins.

A worked example of the top: example-top.md, next to this file (a fictional brand — never reuse its names, numbers or findings). Read it before writing your first In short.

Language: plain words, short sentences, active voice; one idea per paragraph; bullets, tables for comparisons, bold on what matters. Every fact carries a date and a link, inline: "Loxo starts at $149/user/month (pricing, 17 Oct)". A fact from a file the marketer gave (an export, a screenshot) links to that file in their folder under a readable name ("your Shopify export, 6–12 Oct"); the path sits only inside the link. In a hosted chat with an uploaded file, name the file and its date. Say how sure with the three words above — confirmed, likely, first look (anything weaker, "provisional" or "a guess", is a first look, with why) — the same word in "In short" as in the body; never scores or made-up percentages.

Never in a report: the banned words in §3; field dumps, JSON or raw lists of everything collected; hedging filler, empty headings. One exception: two or three outside tools the marketer must run themselves (an email finder, an email platform), named as examples, never an endorsement — the guide's rule (the marketing-guide skill §3, "Talking to the marketer").

2. Deliver it so it opens

Do all that apply. Never leave a report only on the GooseWorks server.

1. Host with files (Claude Code, Codex, Cursor): write it as Markdown in the marketer's current folder at gooseworks-reports/<brand-slug>/<YYYY-MM-DD>-<topic>.md, give a clickable link, [<title>](gooseworks-reports/…), and offer to open it. If the host can render a page or artifact, use that too. Never overwrite a report: suffix the local file, the brand copy's path and its title by save-to-brain's never-overwrite rule (the save-to-brain skill (../save-to-brain/SKILL.md) step 1).
2. Hosted chat (Claude.ai, ChatGPT): an artifact or document if the host can; otherwise the full report in the chat, with its headings, not a summary.
3. With GooseWorks connected, always save it to the brand through the save-to-brain skill (../save-to-brain/SKILL.md), then get a link with gooseworks:file_url: "Saved to your brand: open". Local mode (the marketing-guide §5): save-to-brain's local branch keeps the copy the next run reads; give the step 1 link and say "saved in this folder". Never call gooseworks:file_url and never say it is saved to the brand. Hosted chat without files: step 2, and say it isn't saved anywhere, so they should keep a copy.

In the chat, after delivering, at most 3 lines: the first move with the findings it rests on (a few words), the link, the decisions you need. Never re-paste the report as a summary.

3. Check before you hand it over

A failed check blocks delivery. Fix, re-check, then hand over. Never ship over budget or with a known miss, or with a note that it is over.

• Banned words: id-like codes (A-07, R14, G1, T1, S1, D2…), local or workspace paths shown as text (research/market/…, *.md, *.json, gooseworks-reports/; link targets are fine: web links, and the marketer's own files under a readable name), tool or skill names (gooseworks:, brand_read, knowledge_search, research-market, playbook, atom, MCP), model or vendor names (Claude, Sonnet, Opus, GPT, ScrapeCreators, fal, ElevenLabs — outside tools allowed by §1's one exception), process words ("I called", "per the checklist", "Step 4"), made-up precision ("confidence 0.7", "82% likely"). Scan the text and fix every hit.
• Budget: "In short" ≤ 150 words; body ≤ 3,000; ≤ 6 moves. Count with a tool or by hand, never estimate. Over → cut repetition, then detail no decision needs, then move tables to the appendix; count again.
• Sections present: In short (what we looked at, numbered findings, "So, do this first"), Decisions for you, What we're not doing, What we couldn't check, Sources with a source-skew line.
• In short is never stronger than the body. Each sentence carries the certainty word of the body passage it comes from: a body "likely" or "first look" is never an In short fact; "It is growing" needs a body that says confirmed. A first-look read (one browser, country, sample or day) says so ("first look, from one browser in India").
• "Do this first" assumes nothing unconfirmed. A move that needs an asset the brand hasn't confirmed — customers who will be named, a budget, data or access, a live offer, a person's time — becomes "confirm X" or moves to Decisions for you.
• Every number in In short counts exactly what is claimed beside it. "230 live ads" next to "paying to say it works inside our tool" is wrong when only 4 say that: write "4 of their 230 live ads say…". Re-check each against its source.
• Siblings: an overlapping same-brand report from the last 14 days was read, and a different lead buyer or first move is named and settled.

Then re-read it as the marketer, for the first time: does it pass the 30-second test, could a non-marketer follow it, does every recommendation trace to a dated source, does the link open?