A drift corrector for cooperative-communication discipline. The method —
re-anchor, audit the work in flight, correct forward, report, and the tone
that firing this is not an accusation — lives in
${CLAUDE_PLUGIN_ROOT}/context/re-anchor-audit-correct.md.
Read it; this file adds only what is specific to communication discipline.
The discipline this re-anchors
Cooperative communication: give the reader what they need — no less, no
more — relevantly, clearly, and with your uncertainty and limits disclosed.
This discipline rests on maxims their primary sources own; this skill
points at them and does not restate them (it dogfoods
/discipline:point-dont-copy — a paraphrase of the maxims here would be a
second copy that drifts):
- Grice's Cooperative Principle and its four maxims — Quantity, Quality, Relation, Manner — in Grice, "Logic and Conversation" and Cooperative principle.
- The AI-augmented maxims for human-AI dialogue — Transparency and Benevolence — in Miehling et al., "Language Models in Dialogue: Conversational Maxims for Human-AI Interactions" (arXiv:2403.15115).
Resolve the source of truth per the method doc's ladder: if the consuming
project declares a communication or writing-quality convention in its own
CLAUDE.md / .claude/rules/, re-anchor THAT as well and audit against it.
The rubric this skill audits
Its own delta — the four axes this corrector owns, mapped onto the pointed-to maxims:
- Quantity, both directions. Under-informing is a finding: an asked-for detail, a load-bearing caveat, or a necessary step omitted. Over-informing is a finding too: padding, restatement, and detail the reader did not need. Both fail the reader.
- Relation. Answer the question actually asked. An adjacent answer to a question the reader did not ask, a tangent, or a buried lede (the answer present but hidden below setup) fails relevance even when technically correct.
- Manner. Unambiguous references (no pronoun or "it" the reader must resolve), ordered structure, and plain clarity over cleverness or needless jargon.
- Transparency. Disclose the boundaries: flag genuine uncertainty rather than asserting through it, and state the edges of your knowledge or capability rather than letting the reader assume completeness.
Delegations — axes this skill routes elsewhere
- Quality (truthfulness, evidence) →
/discipline:do-your-research. The Gricean Quality maxim — say what you have grounds for — is that skill's discipline, not this one's. A claim's truth routes there; this skill owns how the message is shaped, not whether it is factually backed. - Pure verbosity remediation →
/discipline:tighten-your-output. The Quantity audit here NAMES padding as a finding, but batch prose-trimming is that skill's job; route large verbosity cleanups to it rather than hand-trimming at scale here.
Both are sibling skills in this plugin, so these are direct routes.
Benevolence — deliberate out-of-scope
The paper's other augmented maxim, Benevolence, is deliberately NOT audited here: it is platform / safety-layer territory (the model's alignment posture), not an auditable communication-drift axis a mid-session corrector can meaningfully score. Recorded as an explicit exclusion, citing the paper above — not an oversight.
Audit targets and the dual moment
- Two target kinds. Audit recent agent RESPONSES and agent-authored ARTIFACTS alike — docs, PR bodies, prompts, commit messages, skill text. A buried lede in a PR body fails the same rubric as one in a chat reply.
- Dual moment. Valid as posture-setting at any time (the method doc's conversation-start case — acknowledge the discipline as active and stop, nothing to audit yet), AND as an audit once output exists. Fire it early to set the posture or later to check the work.
What this skill does NOT do
- Does not audit truthfulness. Whether a claim is factually backed is
/discipline:do-your-research; this skill audits the shape of the message, not its evidence. - Does not restate the maxims. The definitions live in the cited sources; this skill audits against them, it does not reproduce them.
- Does not fabricate a finding. A response that gave the reader exactly what they needed, relevantly and clearly, audits clean; say so.
Gotchas
- Correct and unhelpful is still a finding. An answer can be true, and fail Relation or Quantity — right, but not to the question asked, or buried under detail the reader did not need. Judge by what the reader needed, not by whether the words are accurate.
- The buried lede is the common miss. The answer is present but below the setup; Relation and Manner both fail when the reader has to dig for the thing they asked for. Lead with it.