docs(playbooks): add leading words, the no-op test and the context-boundary line to the authoring guidance #8645

Refs: #7582

Agent Brief

Type: Task
Summary: Carry "leading words" into the skill criteria's Body section and give /docs-hygiene:write-for-agents a leading-words section, a sentence-level no-op test and the line that hiding later steps only works across a real context boundary. All additions; no cut.

Current behavior:

• plugins/docs-hygiene/skills/write-for-agents/SKILL.md:129-132 ("Prompt the positive") has one clause: "pretrained leading words are the compact anchors that steer". Nothing tells an author to look for them, and nothing about them reaches skill bodies: the criteria's ## Body section (plugins/playbooks/skills/skill-authoring/reference/skill-criteria.md:67) does not mention them.
• write-for-agents has no test for a sentence that changes nothing, and its "Give every step a completion criterion" section (SKILL.md:95) does not say when hiding later steps works.
• Comparing two wordings is supported: /evals:plugin-eval compares runs with --baseline <before.json> --margin <m>.

Desired behavior:

• skill-criteria.md ## Body: one bullet. Choose leading words on purpose: prefer a term the model already holds (TDD, tautological, tracer bullet, a loop that goes red) over a spelled-out description; keep a chosen term only when a with-and-without comparison (/evals:plugin-eval --baseline) shows it moves behavior. Point at write-for-agents for the doctrine instead of restating it.
• write-for-agents, new section "Leading words": a leading word is a compact concept the model already holds from pretraining, repeated as a token, never re-explained, so it anchors a region of behavior in few tokens. Prefer an existing word to a coined one; a coined word must be defined and recruits no prior. Look for triads spelled out at several sites and pointers spending a sentence on one idea: each is a candidate to collapse into one word.
• write-for-agents, the no-op test, sentence by sentence: an instruction the model already follows by default spends load and says nothing. Whether it changes behavior is settled by running the document, not by debate. A sentence that fails is deleted whole. A word too weak to beat the default is a no-op too; the fix is a stronger word.
• write-for-agents, in the completion-criterion section: hiding later steps from the agent works only across a real context boundary (a hand-off or a subagent dispatch); an inline call leaves the later steps in context.
• One eval case for write-for-agents: given a draft rule with a spelled-out triad and a sentence the model already obeys by default, the output proposes one leading word and deletes the no-op sentence whole.
• No existing sentence is removed from either file.

Key interfaces:

• plugins/playbooks/skills/skill-authoring/reference/skill-criteria.md ## Body.
• plugins/docs-hygiene/skills/write-for-agents/SKILL.md.

Acceptance criteria:

• The diff adds the Body bullet, the "Leading words" section, the no-op test and the context-boundary line, and deletes no line in either file.
• plugins/skill-quality/scripts/check-evals-quality.sh plugins/docs-hygiene/skills/write-for-agents/evals/evals.json exits 0, with the new case asserting the behavior named.
• scripts/check-changelog-fragments.sh --check passes, with one patch fragment each for playbooks and docs-hygiene, naming Matt Pocock's writing-for-agents as the source.
• CI's check-skills passes on the PR (a PR check; this run does not run it).

Out of scope:

• The description standard, check 3 and check 12 (B2's two-attempt slice), the router-under-5k rule, the hand-off phrasing eval, trims and consolidations: they wait for skill-quality: skill authoring, quality and catalog #7582's own sub-plan (B7) and its probes.
• The docs/upstream/mattpocock-skills.md row.
• Editing either skill's description.

Origin:

• Leading words in the criteria with a --baseline comparison; write-for-agents leading-words section, sentence-level no-op test, context-boundary line: Matt's, D4, Q525.

Do not run with: none open.

Slice of #7582, drafted by the poteto-review session from the thread's recorded decisions.