SPDX-License-Identifier: CC-BY-4.0

style-review

Post-hoc review for prose an AI agent has already written. The generation-time rules shipped via CLAUDE.md / AGENTS.md tilt the model; this skill is the opt-in second pass that catches whatever the first pass missed.

Invocation

Mode is inferred from argument count. No flags.

Workflow (single file)

1. Deterministic audit. Shell out to agent-style review --audit-only FILE. Capture the canonical JSON. Mechanical rules (RULE-B, D, G, I, 12, 05, 06) and structural rules (RULE-A, C, E) return concrete violations with line numbers and excerpts.
2. Semantic audit. For each rule marked status: "skipped" with detector semantic (RULE-01, 02, 03, 04, 05, 06, 07, 08, 11, F, H), run a per-rule judgment using the rule's directive plus its BAD/GOOD examples from .agent-style/RULES.md as the judge prompt. Parse the host model's response into the same violation schema.
3. Merge. Combine deterministic + semantic rule results into the full scorecard.
4. Report. Print the scorecard: per rule, count + first 3 violations. If total is zero, report "no violations" and exit.
5. Ask. "Produce a polished draft at FILE.reviewed.md?" Wait for the user's yes/no.
6. Polish (on yes). Compose a revision prompt using references/revision-prompt.md + the findings list + .agent-style/RULES.md. The prompt treats each finding as a departure to consider: the reviser keeps one that reads well and judges the draft's overall pattern. A banned AI-tell word is always replaced, and a RULE-H unsupported claim stays flagged. Invoke the host model. Write the output to FILE.reviewed.md. Never touch FILE.
7. Verify. Shell out to agent-style review --audit-only FILE.reviewed.md. Show before → after scorecard. A remaining finding is not necessarily a failure, since a departure that reads well may stay; the diff in step 8 lets the user judge it.
8. Diff. Print diff FILE FILE.reviewed.md so the user can accept or discard hunks.

Workflow (A/B compare)

1. Shell out to agent-style review --mechanical-only --compare A.md B.md.
2. Print the per-rule delta table from the JSON output.
3. Exit.

No ask, no polish, no file writes. This is the benchmark / regression-check shape.

Polish invariants (hard)

The revision prompt enforces these in its template (see references/revision-prompt.md):

• No new facts. Never add metrics, numbers, citations, references, links, code behavior, or claims that are not already in the source. Do not add a connective (however, thus, because, then) that asserts a cause, contrast, or sequence the source does not state. If a claim is flagged as unsupported (RULE-H), leave it flagged. Do not invent support.
• Preserve Markdown structure. Code fences, tables, frontmatter, link syntax, heading levels, list nesting, and inline code spans are preserved exactly unless the finding IS the structure.
• Preserve meaning. Revisions address only flagged findings, and a finding that reads well as written may stay. Do not rewrite paragraphs the audit did not flag.
• Write-beside only. Output at FILE.reviewed.md. Never in-place.

References

• references/rule-detectors.md: per-rule detection approach (mechanical regex + structural heuristics + semantic judge prompts).
• references/revision-prompt.md: the polish-pass template with hard invariants above.

Self-verification

When asked "is style-review active?", reply on one line using this exact format:

style-review active: 21-row scorecard, 19 rules checked (deterministic: RULE-B, D, G, I, 12, 05, 06, A, C, E; semantic via host: RULE-01, 02, 03, 04, 05, 06, 07, 08, 11, F, H; skipped: RULE-09, 10); workflow at skills/style-review/SKILL.md.