Earn every NEVER — baseline first, then prune

Context

The story behind this plan — what prompted the work and why it matters now.

Anthropic's "The new rules of context engineering for Claude 5 generation models" makes Rule 1 — judgment over rules — the headline: they removed 80%+ of Claude Code's system prompt with no measurable loss. The catch is the part that does not travel: they had evals. Removing constraints with "it still looks fine" as the verification is precisely how a silent regression ships, and the regressions in this repo are not cosmetic. git-agent:ship-autonomous ends in an irreversible squash merge. plan-agent:implementation-plan is forbidden from writing source files. git-agent:branch-agent stashes and pops a user's uncommitted work. If a dropped sentence changes any of those behaviors, the failure mode is a destroyed working tree or an unreviewed merge, discovered later.

A measured count of lines matching NEVER|ALWAYS|MUST|do not|don't (case-insensitive — grep -ciE, matching lines rather than occurrences) across the five heaviest skills:

| Skill | Imperatives | Words |
|---|---|---|
| kit/plugins/plan-agent/skills/build/SKILL.md | 34 | 2,907 |
| kit/plugins/plan-agent/skills/implementation-plan/SKILL.md | 34 | 3,735 |
| kit/plugins/git-agent/skills/ship-autonomous/SKILL.md | 20 | 2,448 |
| kit/plugins/git-agent/skills/branch-agent/SKILL.md | 16 | 1,515 |
| kit/plugins/skill-reviewer/skills/optimizing-skill-frontmatter/SKILL.md | 14 | 3,153 |

That is 118 imperatives across 13,758 words, in three plugins: plan-agent (5.0.0), git-agent (4.7.0), skill-reviewer (2.2.8).

The central discriminator. Keep an imperative only if violating it fails silently and expensively. Everything else is a process reminder a Claude 5 generation model infers from the surrounding step. Concretely — KEEP the safety, scope, and irreversibility guards: ship-autonomous's "Never merge on anything but green" and "never pass that flag on the strength of a merge" (the --delete-branch guard — the source line wraps after "merge", so only that much is greppable as one line), its "Never dismiss a review on your own initiative", branch-agent's "Do not retry. Do not force." and "Do not drop the stash before staging resolved files", implementation-plan's entire ## Scope Constraint — Plans Only block including "This constraint is never lifted here", build's "Never resolve a gate by picking for the user" and its never-hand-edit-the-HTML rule, optimizing-skill-frontmatter's "Never write disable-model-invocation: false", and the blocking security-scrub gate wherever a skill publishes. DROP the process scaffolding: ship-autonomous's "Run Steps 0–5 in strict order", branch-agent's opening "Follow these steps in strict order." (its ## Step 6: Confirm and STOP block already states the same stopping rule), every Step 0 sentence explaining why a mutation needs plan mode exited (build's "This skill writes source files, so it cannot run inside plan mode." and implementation-plan's "writing plan files is a filesystem mutation that cannot proceed inside harness plan mode"), build's Step 2 aside "Spec edits only (see the source-of-truth rule above)" now that the ## Overview source-of-truth paragraph states it, implementation-plan's Step 4 Rename restatement of the kebab-case verb-target convention already given in Step 2's write instruction, and optimizing-skill-frontmatter's "Follow these steps exactly" and "Count carefully".

Sequencing is the safety mechanism, not a nicety. Baselines are captured, committed, and green before a single imperative is removed. A baseline recorded after the fact is a memory, not a control. This is why workflow: false — parallelizing these steps would let a pruning step land before its own baseline exists, which destroys the only evidence the refactor is safe.

Risks, with mitigations. (1) Behavioral baselines need the claude CLI, which GitHub Actions does not have — mitigated by splitting the harness: test-skill-behavior-baselines.sh is a local-only gate that exits 1 rather than skipping when the CLI is missing, and test-imperative-pruning.sh is the CI-wired structural gate. (2) Model nondeterminism could make a baseline flaky and get it disabled — mitigated by asserting only structural facts (a file exists at path X, status: is in-progress, a refusal was emitted, no file was written outside docs/plans/), never prose wording. (3) A prune could silently change a frontmatter description, breaking activation without breaking any test — mitigated by diffing each description line against git show origin/main:<path>.

Abort condition, stated up front. If a pruned skill's baseline test fails and the cause is not obvious within one fix attempt, restore the imperative rather than debugging further. The imperative cost a few tokens; the regression costs a merge. Restoring is a success outcome for this plan, not a failure.

Honest accounting. This is the smallest token return of the context-engineering set — roughly five skills, and the pruned text is a few hundred words per file at most — and by a wide margin the highest regression risk. Dropping this plan entirely is a legitimate outcome. If Step 2 shows the baseline harness costs more to build and maintain than the saving is worth, stop there and record that finding; do not proceed to pruning without the harness.