| name | docs-humanizer |
| description | Review and improve tracked documents in target-repos/<repo>.git/.wt/main with a Japanese anti-AI-writing checklist, a mechanical audit pass, and a codex worktree workflow. Use when editing README, AGENTS, CLAUDE, docs/, issues/, specs, or other repository documents managed from moonrepo. |
Docs Humanizer
When To Use
Use this skill when you want to inspect or rewrite documents under target-repos/<repo>.git/.wt/main so they read like deliberate human-written technical documentation instead of generic LLM output.
Typical targets:
README.md, README.mbt.md
AGENTS.md, CLAUDE.md
docs/**/*.md
issues/**/*.md
- Any other tracked
.md, .mdx, .txt, .rst, .adoc
Preconditions
- Run from the
moonrepo workspace.
- Target repo exists at
target-repos/<repo_name>.git/.wt/main or in a codex-created worktree under target-repos/<repo>.git/.wt/codex/.
- Commands available:
bash, git, jq, rg.
Primary Commands
Mechanical audit:
bash .agents/skills/docs-humanizer/scripts/audit-docs.sh <repo_name>
Or audit an explicit worktree path:
bash .agents/skills/docs-humanizer/scripts/audit-docs.sh --path /abs/path/to/worktree
Start the full codex workflow:
bash .agents/skills/docs-humanizer/scripts/docs-workflow.sh <repo_name> <task-slug>
Equivalent just entrypoints:
just docs-audit <repo_name>
just docs-audit-all
just docs-review <repo_name> <task-slug>
Workflow
- Run the mechanical audit first. Treat the audit as a triage pass, not as the final judge. It only covers patterns that are easy to detect with regex.
- Prioritize high-signal documents first:
README*, AGENTS.md, CLAUDE.md, docs/, then issue/spec files.
- Remove machine-like patterns first:
- em dash / fullwidth dash
- inline-heading bullets
- chatbot residue
- templated introductions and conclusions
- excessive hedging and vague sourcing
- Then do a content pass using
references/anti-ai-checklist-ja.md.
- Re-run the audit until remaining matches are either gone or clearly intentional.
- If you change commands, examples, or process docs, validate the commands against the repository where possible.
Editing Rules
- Prefer direct statements over inflated significance.
- Prefer simple predicates over roundabout copula-avoidance phrases.
- Keep bullets plain unless the distinction is meaningful enough to deserve a true subsection.
- Delete filler transitions such as repeated
さらに / 加えて when the sentence already flows.
- Replace vague claims with concrete facts, file paths, commands, dates, or observed behavior.
- Introduce a small amount of authorial judgment where it clarifies tradeoffs, but do not add fluff.
Reference
- Read
references/anti-ai-checklist-ja.md before rewriting.
- The checklist is derived from the Zenn article below and the humanizer projects it references:
Validate After Apply
bash .agents/skills/docs-humanizer/scripts/audit-docs.sh <repo_name>
git -C target-repos/<repo_name>.git/.wt/main diff --stat
Expected: obvious mechanical findings are reduced, tracked document diffs stay intentional, and rewritten docs read as specific technical writing rather than generic AI prose.