source-before-shape — read the source before committing to a shape

Census id: F29 · Family: F — rules corpus + generated-rule tooling + shipped audit scripts · Provenance: framework pin b069c59328aa8e08671d3a19fd25efa618dfc1c5 (staging FETCH_HEAD, read 2026-09-11; re-anchored from census pin a1337cb301). Every file:line below is at that pin. Raw reference drafted by stage E4 — factual, unpolished by design.

What it bans

Authoring or scoping from internal recall when the authoritative source was one probe away. .claude/rules/source-before-shape.md:8 «# Source-before-shape — read the authoritative source before committing to a shape»; the rule: :26 «Before an authoring or scoping action that commits to an artifact's shape, the authoritative external source MUST be read first, and as an input that drives the shape — not consulted post-hoc as a trailer or a correction:» Origin incidents: :20 «1. BFR-reinvention: PR #858 shipped .claude/skills/night-mode/SKILL.md re-describing the executor + dual-reviewer loop already owned by Superpowers subagent-driven-development … The trailer said the right thing; the body did the wrong thing.»; :21 «2. Scope-from-memory: an autonomous launch prompt was scoped to «implement D1» from recall, when the spec defines the chain as D1 → B». Class header: :13 «> Class: B — compensating mechanism without a CI principle test on the discipline itself: the violation … is a judgment call, not mechanically detectable, so it cannot be gated».

Never (fires)

The anti-pattern family: .claude/rules/source-before-shape.md:57 «- #authoring-from-internal-state (the family) — committing to an artifact's shape or a dispatch's scope from an internal mental model / session recall at an authoring moment that has no action-scoped gate, when the authoritative external source was one cheap probe away and would have changed the artifact.» with sub-cases :58 «#consult-as-trailer-not-input — the prior-art consult happened but only as a post-hoc Prior-art: trailer, too late to shape the body» and :59 «#scope-from-memory-not-source — scoping a dispatch/launch from recall instead of re-reading the spec's scope-chain.»

Always (clean)

The two surfaces: :28 «- Reuse surface …: grep the SSOT … + the existing .claude/skills/ and agents/ for overlap on this capability-area before writing the body. If an own-stack or upstream analog exists, the body is a thin adapter that subordinates to it» and :29 «- Scope surface (scoping a dispatch, launch prompt, or kickoff): re-read the spec's scope / chain section (sed/Read the actual lines) before scoping. A scope stated from recollection is provisional until the source is quoted.»

Enforcement channels

• Rule-index row: .claude/rules/00-rule-index.md:37 «| source-before-shape.md | B | creating a new SKILL.md/agent/module, or authoring a dispatch/kickoff. | paths:(3), edit-time inject |»
• Layer A — edit-time injection, 3 globs: .claude/rules/source-before-shape.md:10 «» (inject text at :11), via inject-matching-rule.sh, once per session (:44).
• Layer B — AI-agnostic auditor: agents/capability-reuse-auditor.md — the semantic overlap pass, run before handoff (:46, :78).
• Honest limitations stated, not hidden (:48-53): the once-per-session token can be «spent» on a routine edit — «Layer A would not have fired on its own origin incident»; packages/core/** deliberately excluded; inline dispatch prompts unreachable.
• Not backend-rendered — no FF diagnostic applies (FF7001/FF7002, packages/core/diagnostics/registry.ts:321/:329); honest status: best-effort first-touch injection + auditor + parents (BFR, §1.11) + principle 11 F1.