Editorial Review

Announce at start: "I'm using the editorial-review skill to run quality gates on this artifact."

Overview

The final gate before an artifact reaches the repository. This skill coordinates two layers of quality checking: deterministic tools for things agents are bad at (counting tokens, computing readability scores, validating section completeness) and agent-based review for things tools can't catch (sarcasm, assumed context, tone drift, the conference-talk test).

This is the tone firewall in practice — the enforcement mechanism behind ADR-0002.

When to Use

• Before committing any documentation artifact (ADR, handoff, design doc, end-user doc, changelog, release announcement)
• When another skill's final step says "run through editorial review"
• As an on-demand check during drafting, if you want early feedback before the commit path
• When manually-written docs are staged for commit — they need the same quality gate as agent-written docs

When NOT to Use

• For any private context (journal entries, PRIVATE_MEMORY.md, private memory files) — private context bypasses the firewall entirely. That's the point.
• As a writing tool — this skill reviews, it doesn't draft. Use the appropriate writing skill first.

Quick Reference

Layer	Tool	What it checks	When it fails
Deterministic	bito tokens	Token count vs. budget	Handoffs over 2,000 tokens
Deterministic	bito readability	Flesch-Kincaid grade level	Above persona target (≤8 for Doc Writer, ≤12 for Technical Writer)
Deterministic	bito completeness	Required sections present and substantive	Missing or placeholder sections
Agent-based	${CLAUDE_PLUGIN_ROOT}/agents/editorial-reviewer.md	Conference-talk test, negative refs, assumed context, tone consistency	Any passage that would be uncomfortable on stage

Process

Step 1: Identify the artifact

Determine:

• Artifact type: adr, handoff, design-doc, end-user-doc, changelog, or release-announcement
• Target persona: Which persona voice should the artifact match?
• File path: Where the artifact lives

Step 2: Run deterministic checks

Dialect: Check for BITO_DIALECT environment variable or the project's bito config for a dialect preference (en-us, en-gb, en-ca, en-au). If set, pass --dialect <value> to bito analyze to enforce dialect-consistent spelling. The consistency checker will flag wrong-dialect spellings in addition to mixed usage.

Run the applicable bito checks. Not all checks apply to all artifact types.

Artifact type	Token check	Readability check	Completeness check
Handoff	Yes (budget: 2000)	No (density metric, not grade)	Yes (template: handoff)
ADR	No	Yes (max-grade: 12)	Yes (template: adr)
Design doc	No	Yes (max-grade: 12)	Yes (template: design-doc)
End-user doc	No	Yes (max-grade: 8)	No (structure varies)
Changelog entry	No	No (too short for meaningful score)	No
Release announcement	No	Yes (max-grade: 8)	No

Commands:

# Token check (handoffs only)
bito tokens <file> --budget 2000

# Readability check
bito readability <file> --max-grade <target>

# Completeness check
bito completeness <file> --template <type>

If any deterministic check fails, report the failure. The artifact should be revised before proceeding to agent review — no point checking tone on a document that's missing required sections.

Step 3: Run agent-based review

The context: fork frontmatter on this skill delegates to the editorial-reviewer agent automatically. The agent receives this skill's instructions as its task, along with the artifact content, type, and target persona from the earlier steps.

The agent works through six checks:

1. Conference-talk test. Would every sentence be comfortable to say on stage at a technical deep-dive?
2. Negative references. Any people, projects, or products mentioned disparagingly? (Factual comparisons are fine.)
3. Assumed context. Does the document stand alone? Any "as discussed" without a link?
4. Tone consistency. Does the voice match the target persona?
5. Readability. Short paragraphs? Scannable headers? Jargon defined?
6. Completeness (qualitative). Beyond section presence — is the content substantive?

Step 4: Report results

Produce a structured review report:

If all checks pass:

## Editorial Review: [artifact type]

**Overall:** PASS

**Deterministic checks:**
- Tokens: [N/A or PASS (count/budget)]
- Readability: [N/A or PASS (grade/target)]
- Completeness: [N/A or PASS]

**Agent review:** No issues found. The artifact meets professional standards and is ready to commit.

If any checks fail:

## Editorial Review: [artifact type]

**Overall:** FAIL — N issues found

**Deterministic checks:**
- Tokens: [PASS/FAIL with details]
- Readability: [PASS/FAIL with details]
- Completeness: [PASS/FAIL with details]

**Agent review issues:**

### Issue 1: [Category]
**Passage:** "[quoted text]"
**Problem:** [what's wrong]
**Suggested revision:** "[revised text]"

### Notes
[Optional observations that aren't failures but worth considering]

Step 5: Iterate if needed

If the review fails:

1. Report the specific issues to the writing agent or user
2. Let them revise the artifact
3. Re-run only the failed checks (deterministic and/or agent) on the revised version
4. Repeat until PASS

Do not auto-fix agent review issues. The reviewer identifies problems and suggests revisions; the writer decides how to address them. The writer persona may have good reasons to push back on a suggestion.

Integration

• Called by: Every other writing skill as its final step (curating-context, capturing-decisions, writing-design-docs, writing-end-user-docs, writing-changelogs)
• Also triggered by: Pre-commit hook (${CLAUDE_PLUGIN_ROOT}/hooks/pre-commit-docs) for manually-written or externally-modified docs
• Depends on: bito CLI (cargo install bito) for deterministic checks, ${CLAUDE_PLUGIN_ROOT}/agents/editorial-reviewer.md for agent-based review
• Enforces: ADR-0002 (tone firewall on commit path), ADR-0003 (real tools for measurement, agents for judgment), ADR-0005 (token budget for handoffs)

Common Mistakes

Mistake	Fix
Skipping deterministic checks and going straight to agent review	Run tools first. Agent review on a document missing required sections wastes an API call and produces noisy feedback.
Auto-fixing agent review issues without writer input	The reviewer suggests; the writer decides. Flag the issue and let the writing skill or user choose how to address it.
Applying the wrong readability target	Match the target to the persona, not the artifact type. Doc Writer docs target grade ≤ 8; Technical Writer docs target grade ≤ 12.
Running editorial review on PRIVATE_MEMORY.md	Never. Private memory is explicitly outside the firewall. If someone asks to review it, decline.
Treating "Notes" as failures	Notes are observations, not blockers. A note might say "consider shortening this paragraph" — that's advice, not a gate.
Re-running all checks after a minor revision	If only the agent review failed, re-run only the agent review. Deterministic checks are idempotent but unnecessary to repeat if the relevant content didn't change.