Post-2.0: rule-carrying prose does not fire, and we keep adding more of it — audit where instructions live and how they collide

[BaseInfinity]

Deferred: file now, work it after the v2.0.0 major release. Not a blocker for 2.0.

The evidence

PR fix/563-636-clean-round-1-and-base-ancestry (#563 + #636) ran eight review rounds on ~2 files. Rounds 6–8 were self-inflicted: the driver volunteered a guard mid-review on a field disclosed as binding to nothing, and two full rounds were spent reviewing it. The guard was deleted at the end (#643).

skills/sdlc/SKILL.md already contained four separate rules that each independently end that at round 6 or 7:

• :136 — reviewers may return WRONG SHAPE (stop, redesign)
• :152 — "No test costs more rounds than the change it guards"
• :163 — "Blame the line: a blocker in code nobody asked for — cut it, don't repair it"
• :273 — "If you are repairing a guard for the second time this cycle, ask whether it should exist"

None of them fired. Not concealed, not unknown — read, and not applied under momentum. Fable's diagnosis: they are ambient driver-memory prose, and this repo has a proven failure record for that mechanism and a proven success record for the other one (:273's own note: "Reviewers only pick outcomes the prompt offers. Rounds 1–3 kept it off the menu; round 4 put it on and it was taken immediately").

The immediate remedy is scoped and shipping separately (reframed #617: shape + target as required reviewer fields). This issue is the general problem behind it.

The general problem

Our instructions live scattered across .md and config surfaces, each one authoritative for something, none of them reconciled against the others:

• skills/sdlc/SKILL.md (ships), skills/setup/, skills/update/, skills/feedback/
• CLAUDE_CODE_SDLC_WIZARD.md, AI_SETUP_LANES.md, ARCHITECTURE.md, ROADMAP.md
• CLAUDE.md — repo-local, and a different CLAUDE.md in every consumer repo
• cowork/ — byte-parity copies of skills/, a second place the same prose lives
• .claude/settings.json, hooks/hooks.json, .claude-plugin/ manifests
• global ~/.claude/settings.json, whose env block silently overrides explicit tool params

Questions this issue is to answer, with evidence, not opinion:

1. Do rules collide? Find pairs of instructions across these surfaces that contradict, or that a reader would apply inconsistently. SKILL.md:163 vs. the recheck contract at :161 is one candidate — both are about what a finding authorizes, in different words, in the same file.
2. What is the firing rate of a prose rule? We have the data: rules in prompt templates (copied verbatim into a review every round) vs. rules in ambient skill prose (relied on from memory). Measure, don't guess — this is the whole basis for preferring one mechanism.
3. Is there a budget? SKILL.md keeps growing, and skill listings are budget-clamped (a skill over ctx x 4 x 0.01 chars goes name-only). More prose is not free and past some size it is actively harmful — a rule nobody reaches is worse than no rule, because we cite it as coverage.
4. Which surface should own which kind of rule? Rough hypothesis to test: things that must fire → prompt templates and hooks. Things that inform a judgement → skill prose. Things that are facts about the repo → CLAUDE.md. Today they are mixed arbitrarily.
5. Config collision specifically. Global settings silently overriding per-call params has already bitten us once (subagent model/effort pins). Enumerate every override path where a local instruction can be silently defeated by a broader config.

Why after 2.0

This is an audit-and-restructure of the shipped instruction surface. It will move prose that consumers depend on, so it wants a major boundary behind it, not in front of it. #617's reframe covers the specific defect that cost us the eight rounds; this covers the class.

Related: #617 (design judgement is never the momentum-holder's), #516 (two harnesses colliding at the root docs), #485 (replace hand-rolled doc guards with declarative consistency checking), #643 (the deleted guard that produced the evidence).

[BaseInfinity]

Measured: three documented rules skipped in one session, 2026-08-17/18

This issue argues that rule-carrying prose does not fire. Here is a measurement
rather than a hunch, from a single session driving PRs #669, #670 and #674.

Three rules. All written down. All skipped by the driver. None enforced by
anything mechanical.

Rule	Location	Cost of skipping it
Falsify your own claims before the reviewer sees them	skills/sdlc/SKILL.md:159	~4 avoidable review passes. The claims that burned rounds were exactly the ones that skipped this
WRONG_SHAPE → escalate to the design authority, do not patch	skills/sdlc/SKILL.md:170	2 passes. Worse: the driver then wrote a weaker version of this rule into ROADMAP.md — a "second WRONG_SHAPE" threshold — which review caught as a standing-rule regression
Reviewers reconcile with each other; you relay, never merge their words	skills/sdlc/SKILL.md:184, plus a full section at CLAUDE_CODE_SDLC_WIZARD.md:4334	Not run at all until the maintainer asked "did the reviewers talk to each other?"

The reconcile case is the sharpest. It is not a passing mention — the wizard doc
carries a titled section, an explicit "reconcile once, not per round", and an
explicit "do not resolve conflicts yourself; cross-feed their positions
verbatim." The driver instead posted both positions side by side and left the
disagreement open, which is presentation, not reconciliation.

When it was finally run, it changed the answer. Sol conceded the
disagreement, accepted the other model's causal decomposition, chose the other
model's proposal over its own, and found a real bug in that proposal (the guard
fails open on a stale origin/main and ignores non-main base branches). The
step produced something better than either reviewer had alone — which is the
argument for it, and it still did not fire on its own.

Why this supports the no-new-rules ruling rather than contradicting it

The maintainer has ruled twice against baking new mandatory steps into the
skill. Tonight is evidence for that ruling: all three rules already existed
and adding a fourth would not have helped. The failure is not missing prose.
It is that prose is the enforcement mechanism.

grep -rln reconcil hooks/ scripts/ returns nothing. Nothing checks any of the
three.

Not proposing a fix here

Sequencing is the maintainer's. Recording the measurement so the next
discussion starts from data. Related: #616 now carries two mechanical proposals
from the cost review, and the reconciled recommendation is base-freshness
enforcement in scripts/run-review-leg.sh — notably a script, not a rule.

[BaseInfinity]

Correction to the comment above — it flattered the ruling

Review caught the framing. The previous comment concluded that three skipped
rules are "evidence FOR the no-new-rules ruling rather than contradicting it."
That is one reading, and it is the reading most convenient to whoever wrote it.

The honest version keeps both halves:

• The evidence does reject adding more duplicate prose. All three rules
  already existed; a fourth would not have helped. That much stands.
• The evidence equally supports mechanically enforcing the obligations
  already written. That is a different proposition from the ruling the
  maintainer actually made — which was specifically against new mandatory steps
  in skills/sdlc/SKILL.md and against a commit-hook lane.

"Three written rules did not fire" is an argument about enforcement, not
about how many rules exist. Collapsing it into support for the standing
ruling reads as deference while quietly widening a targeted decision into a
general position. The maintainer ruled on a specific proposal; they have not
ruled that existing obligations may never be mechanised — and #616's reconciled
recommendation is precisely a mechanical enforcement of one.

The measurement in the comment above is unchanged and still stands. Only the
conclusion drawn from it is corrected.