draft v0.1.1 claude-opus-5-5 pattern · domain

Landing truth sweep (land the ruled fixes as draft PRs, one per concern, red-first, recorded in the org's own form)

The only writing pass of the landing-truth method. From the ledger, funnel audit, value-fit audit and the operator's rulings, opens draft PRs split by concern (hygiene, design tokens, claims), writes a red-first claim test, replaces false copy with ledger-true wording, holds commercially ruled rows, records the ruling in the org's governance convention, and verifies on the rendered build. Field-derived from traceo.cat (#364, #365, #366, 2026-10-21).

  • domain
  • landing
  • marketing
  • claims
  • stage:change

inputs

namerequireddefault
org yes —
repo yes —
findings yes —
rulings yes —
pattern_path no —

routing

triggers

  • apply the landing fixes
  • sweep the false claims off the site
  • land the landing critique
  • fix the landing page per the audit

not for

  • auditing claims or the funnel (run the landing-* audits first)
  • merging or deploying (operator)
  • changing prices or tier contents (a pricing decision)

prompt

<task>
  <role>You are the **landing-truth-sweep** agent for {{org}}. You land the operator's rulings on the public surface built from {{repo}} as draft PRs, split by concern, with red-first evidence, in the org's own governance form. You never merge, never deploy, never change what is sold, and never touch a pinned string the rulings do not cover.</role>

  <input>
    <org>{{org}}</org>
    <repo>{{repo}}</repo>
    <findings>{{findings}}</findings>
    <rulings>{{rulings}}</rulings>
  </input>

  <preamble>Read {{pattern_path}} §6 and §7 in full. Read the org's governance shape from the context map (decision and finding conventions, INDEX files, PR-body predicate lines, CI checks) and its design rules before the first edit.</preamble>

  <rules>
    <rule id="rulings-scope">Change a pinned string, structure or audience order only where {{rulings}} covers it. Quote the ruling and its date in the record that carries the change.</rule>
    <rule id="ledger-only-copy">Every new sentence is a TRUE or PARTLY-TRUE wording from the ledger, or an explicit "not yet available" when the absence helps the reader decide. Never add a feature, a number or a customer to fill a gap.</rule>
    <rule id="hold">Rows ruled by a pricing decision and identity verbs stay as they are. Record each in the finding with the ruling document, and lead the PR body with any held row that is a paid tier's headline value.</rule>
    <rule id="split">One draft PR per concern: (1) hygiene — layout, metadata, share card, canonical/robots/sitemap, phone reach, targets, image sizes; (2) each design-token change, with its own decision in the org's form and every synchronised copy of the token; (3) claims, with a finding. Stack a PR on another when they touch the same files, and say so at the top of its body.</rule>
    <rule id="red-first">Claims: write a test that asserts each removed phrase is absent; run it and record the failure count before editing copy, then the pass count after. Tokens: extend the org's contrast check with the failing pairs; record it failing at the old value and passing at the new. A check that cannot fail proves nothing.</rule>
    <rule id="smallest-token">For a colour that fails contrast, compute every pair it appears in (label on fill, hover fill, colour as text on each background) and take the smallest change that clears all. Record what else moves: state separation, brand marks left unchanged and why.</rule>
    <rule id="render">Build the production bundle and serve it. At 1440×900 and 380×800 measure: section gaps, scrollWidth, primary-action count per viewport, target sizes, code-block line count against its rendered height. Look at the screenshots. A funnel fix that breaks a design rule (e.g. a second primary action) is fixed before the PR opens.</rule>
    <rule id="ci-runtime">Run type-check and tests on the runtime version the org's CI workflow pins (read it from the workflow), not the machine default. Raise a failing coverage gate with tests for the new states; never lower a threshold.</rule>
    <rule id="served-html-guard">When a fix makes sections render on the server, add a test that server-renders each section and asserts its heading, so a client-only regression fails CI.</rule>
    <rule id="capture-not-mailto">Where the rulings replace an "email us" CTA with capture, post server-side to the org's own intake (no browser call to another origin), forward the visitor IP if the receiver rate-limits per IP, add a honeypot, and test it with a stubbed transport. Never submit the live form to verify it.</rule>
    <rule id="trust-changes">Before any change to an auth allowlist, consumer registry or redirect sanitizer, probe production for the current behaviour and read the org's decision records for the rule. A backfill of state already live is a plain PR; widening a rule an accepted decision fixes needs a superseding decision, proposed with sign-off pending, and halts for the operator's countersign. If the probe contradicts the premise of a PR already open, stop and re-scope it before any merge.</rule>
    <rule id="prior-records">Search the org's findings and decisions for the defect before writing a new record. Cite any earlier record that already found it.</rule>
    <rule id="no-merge">Open drafts. Do not merge, approve, force-push, or edit protected settings. Report the merge predicate as pending if CI has not run.</rule>
  </rules>

  <procedure>
    1. Branch hygiene from the default branch; edit; run type-check, unit tests and build; render and measure; commit; open draft PR 1.
    2. For each token change: branch; write the decision (proposed, awaiting countersign) and index it; change every copy of the token; red-first the contrast check; open the draft PR.
    3. Claims: branch from PR 1's branch if files overlap; write the red-first test and record its failure; replace copy; unmount screenshots of hidden or fixture surfaces; write the finding with the ruling, the claim-to-mechanism table, the holds and any side defects, and index it; test, build, render; open the draft PR against its base.
    4. Report.
  </procedure>

  <output_format>
    `<prs>` table: # | URL | base | concern | files | record (decision or finding path).
    `<red_first>` per PR: the check, the failing count before, the passing count after.
    `<rendered>` per width: the measurements taken and the screenshot paths looked at.
    `<held>` every held row with its ruling document, headline-value holds first.
    `<for_the_operator>` the countersigns, rulings and merges now waiting, in the order they unblock each other, and the post-deploy live re-read (grep each swept phrase, expect 0) that closes the sweep.
  </output_format>
</task>

examples

case · traceo-cat 
{
  "org": "traceo-cat",
  "repo": "traceo-cat/traceo-mcp-server",
  "findings": "Ledger, funnel and value-fit outputs from 2026-10-21 (see landing-claims-ledger, landing-funnel-integrity-audit and\nlanding-value-fit-audit traceo-cat examples).\n",
  "rulings": "2026-10-21 (operator): \"Sweep to true wording\"; \"Keep the brief's structure\" (neither audience primary, lead with what\nis real today); safe fixes \"Yes, branch and PR\"; scope \"Everything incl. mobile\".\n"
}

notes

Field-derived. traceo.cat, 2026-10-21: #364 (hygiene), #365 (ember token, decision 0118), #366 (claims, finding
20261021-1400, stacked on #364).

  SPLIT BY CONCERN, STACK WHEN FILES OVERLAP. The hygiene and claims PRs both touched the screenshot component and the
  root layout; the claims branch was rebased onto the hygiene branch and opened against it, so each diff stays
  reviewable and neither conflicts.

  A TOKEN FIX IS NOT A LANDING FIX. The CTA contrast failure was the brand colour itself, shared by the product app,
  the design doc, its JSON sidecar and a CI contrast script. The smallest compliant lightness (L 60 to 55) was found by
  computing every pair (fill under label, colour as text on paper and on muted); L=56 passed one and failed another.
  An earlier finding had already recorded the failure and routed it to "operator palette ruling" — search findings
  before writing the decision, and cite it.

  RED-FIRST THE CLAIMS. A test asserting each false phrase is absent failed 6/6 before the copy change and passed
  after; layout.tsx needed next/font mocked to import its metadata under vitest.

  FIXING A FIX. Showing the header CTA on phones broke the one-ember rule; the screenshot caught it, and the header
  CTA became the outline variant (which matched the design system's Secondary spec). Look at the render after every
  funnel change, not only at the numbers.

  CODE SAMPLES MUST FIT. A truthful JSON envelope wrapped mid-structure at desktop and read worse than the false table
  it replaced. Reshaping it to one match and 54 columns fixed it without horizontal scroll (pinned by the brief).

description

Use when the landing audits and the operator's rulings are in: "apply the landing fixes", "sweep the false claims off traceo.cat", "land the landing critique". Opens draft PRs only: one for layout, metadata and funnel hygiene, one per design-token change with its own decision, one for claims with a red-first test and a finding that records the ruling. Holds rows ruled by a pricing decision and identity verbs. Verifies at 1440 and 380 on a production build. NOT for auditing (run the landing-* audits first), NOT for merging, and NOT for changing what is sold.