Mirrored verbatim from utopia-models/knowledge _rules/survey-the-territory-before-generating.md by the allowlisted publish pipeline. Edit it there, not here.

name: survey-the-territory-before-generating
cases: [2026-08-16-observe-the-territory-before-generating-candidates, 2026-08-27-bullets-are-built-for-the-reader-not-mined-from-the-candidate, 2026-08-27-swapping-a-noun-is-not-reverse-engineering, 2026-08-27-familiar-beats-differentiated-when-the-goal-is-being-hired, 2026-08-27-defensibility-is-not-a-resume-property, 2026-10-17-stock-footage-recurred-because-the-rip-rule-lived-in-an-unbound-subprocess, 2026-10-17-open-tylers-named-reference-sites-before-building, 2026-10-17-rank-the-competitor-headline-set-before-picking-copy, 2026-10-17-every-customer-facing-claim-carries-a-recorded-source, 2026-10-17-a-new-physical-output-step-starts-from-the-printer-ui-and-a-tool-survey]
status: active
mechanism: prose — the trigger needs judgement a hook cannot observe
last-recurrence: 2026-10-17
audience: agent

When the output must land with an outside audience, census that audience FIRST. The census is the deliverable, not the scaffolding.

Fires when

Producing anything judged by people outside the fleet (name, resume bullet, landing page, cold email, pitch, product description) when material about US is to hand. 🔴 The tell is that the generating step feels ready to start — I am working supply-side. Also: "propose N options", a brief whose first verb is invent/draft/write/generate, editing an artifact produced the wrong way (a revision inherits its shape). 🔴 A site build (2026-10-17): open every reference site Tyler names FIRST and match its structure. Media and copy are RIPPED from successful competitor sites — never stock, never authored; take the strongest ranked headline; every claim keeps its source. Procedure: 8f444ccd79:cleaning/subprocesses/interactive/site-build.md (loads only with a bound plan).

Trigger detail (full text — kept out of the always-loaded bundle)

producing anything judged by people outside the fleet — a name, a resume bullet, a landing page, a cold email, a pitch, a product description — and material about US is already to hand. 🔴 The tell is that the generating step feels ready to start. If I can begin writing from what I already know, I am about to work supply-side. Also fires on "propose N options", on any brief whose first verb is invent/draft/write/generate, and on editing an existing artifact that was produced the wrong way — a revision inherits the original's shape, so a measured word-level fix is still supply-side work. 🔴 A SITE BUILD (2026-10-17, utopiacleans.com). Open every reference site Tyler names FIRST and match its structure. Media and copy are RIPPED from successful competitor sites — never stock, never authored; rank the competitor headline set and take the strongest, never the first; every page claim keeps its source. Full procedure: 8f444ccd79:cleaning/subprocesses/interactive/site-build.md (loads only with a bound plan).

The rule

1. Survey before generating, and make the survey the WHOLE job. A census run as a secondary task becomes scaffolding for the invention; the report's weight lands on the invented thing. Dispatch the survey alone, read it, and only then generate.
2. The generator must be BLIND to our material. A subagent told what we have will find a way to use it. If it asks what we have, the method is working.
3. The artifact is a set of TEMPLATES with placeholders, committed, before anything is filled. Not a mental step — a file. And the check is per-slot: "this fills template X, which answers this quoted audience language." If both cannot be named, the work is an edit.
4. Where nothing in our material fits a slot, the slot stays empty. Never substitute the closest thing we have; that is the supply side re-entering through the back door.
5. 🔴 Familiar beats distinctive when the goal is a DECISION by a person. Match the register the audience already reads. Distinctiveness belongs in the content, never in the shape.

Cost when it fired

PLAN-179 (2026-08-16): returned 12-18 invented names with rationales; the top three were rejected outright. Tyler's correction named the method, not the output: "run research on other AI Ads Agencies to figure out what names they have." The case was written the same day and sat unabsorbed for 11 days — there was no room in the corpus, which was at 19,997 of a 20,000 B ceiling that had never been measured.

PLAN-225 (2026-08-25 → 27): the same failure, five repetitions, three days. "You are way too focused on me, instead of the person reading this resume… BULLET POINTS SHOULD BE 90% DONE BEFORE EVEN LOOKING AT MY KRATOS WORK OR FLEET INFRASTRUCTURE!!!!" Then, after a measured fix that swapped one noun: "Feel like you are STILL NOT reverse engineering." Tyler's standard: "You should be able to run research and write bullet points and content in ONE PASS, not require 3 days of iterations" — and "these are not things that I should need to catch."

When it was finally done correctly, the census produced findings unreachable from the supply side: only 2 of 12 target postings mention out-of-work evidence at all · they count calls, developers, users, documents/day and zero of twelve count commits or lines of code · 58.5% of hired-candidate bullets claim no business outcome, so demanding one from every bullet reads as unlike the resumes that worked.