# Prose Check

> Self-evaluate outward-bound prose against the house rubric — the three tics, end-on-the-fact, TL;DR footer. Use when drafting a PR body, issue body, doc, rule, or commit body before it ships.

- Skill: `laurigates/prose-check` (Agent Skill, multi-file: 5 files)
- Install (CLI): `npx skillmds@latest add laurigates/prose-check`
- Raw SKILL.md: https://api.skillmd.com/api/skills/laurigates/prose-check/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: laurigates (https://skillmd.com/u/laurigates)
- Updated: 2026-09-09
- Page: https://skillmd.com/skills/laurigates/prose-check

---


# /prose:check

Run a draft through the house prose rubric before it ships.

## When to Use This Skill

| Use this skill when... | Use something else when... |
|---|---|
| A PR body, issue body, doc, rule, or commit body is drafted and about to ship | The text is already written and just needs shortening — use `prose-distill` |
| Checking whether a draft carries the three tics `communication.md` says to cut | Turning notes into a plan — use `prose-synthesize` |
| A long answer needs a TL;DR (ELI5) footer decision | The artifact must survive a zero-context reader — use `agent-patterns-plugin:cold-read-gate` |
| Wording a ticket body in the neutral register | Assembling the GitHub issue itself — use `git-plugin:github-issue-writing` |

Not for chat responses that never hit a file — there is nothing to lint. Draft
to a file first if you want the check.

## The Rubric This Encodes

`~/.claude/rules/communication.md` is the canonical source. It names three forms
to cut on sight:

| Form | Shape |
|---|---|
| **Chiasmus / mirrored clauses** | A-not-B, B-is-A — content mirrored across a pivot |
| **Significance-assertion** | telling the reader the fact matters instead of stating it |
| **Aphorism / general maxim** | a sentence that would work as a standalone epigram |

The tell is **position plus shape**: the sentence lands at the end of a
paragraph or section, *and* it generalizes past the specific claim.

Everything under `styles/House/` and in `scripts/` is a **derived encoding** of
that rule, not a second source of truth. Each rule file carries a `link:` back
to the rule that owns its criterion, and `scripts/check-prose-house-style.sh` in
the repo root pins the derived copy against drift.

## Execution

```bash
prose-plugin/skills/prose-check/scripts/prose-check.sh --kind doc path/to/draft.md
```

| Flag | Effect |
|---|---|
| `--kind doc` | default; the three tics plus token and grammar layers |
| `--kind answer` | adds the TL;DR (ELI5) footer check for a complex answer |
| `--kind ticket` | same checks; ticket rules (`TicketHype`, `TicketPlaceholder`) carry their own weight |
| `--long-words N` | paragraph-final sentence word threshold (default 40) |
| `--strict` | exit 1 when any candidate is found |

Read the rollup first — `=== PROSE CHECK ===` carries `STATUS=` and
`ISSUE_COUNT=`. Only open the per-layer sections when the count is non-zero.

## Reading the Output — Candidates, Not Verdicts

**Every line the script emits is a sentence to judge, not a defect to fix.**
Whether a mirrored clause is a chiasmus, or a general statement is an aphorism,
is irreducibly a judgment call. The deterministic layers exist to narrow the
candidate set so that judgment lands on a handful of flagged sentences instead
of a whole document — the split in
`.claude/rules/offload-to-deterministic-substrate.md`, applied to writing.

Two consequences worth internalising:

- **A hedge carrying real uncertainty is a true negative that still shows up
  here.** `communication.md` is explicit: hedges that carry real uncertainty,
  stated caveats, and explicit noise floors all stay — they are information.
  Do not strip a `Hedge` hit reflexively.
- **A document that discusses filler words will be flagged for containing
  them.** `prose-distill/SKILL.md` scores weak-word hits because listing filler
  words is its subject matter. That is not a defect in the document or in the
  check.

Verdict types the script layer emits:

| `TYPE=` | Means |
|---|---|
| `significance_assertion` | paragraph-final sentence carries a portentous noun from the rule's own list |
| `chiasmus` | paragraph-final sentence has a negation plus content mirrored across a pivot |
| `aphorism` | paragraph-final sentence restates its own subject in the predicate |
| `long_final_sentence` | paragraph closes on a sentence at or above the word threshold |
| `long_sentence` | any sentence at or above the threshold — position-independent |
| `missing_tldr_footer` | `--kind answer` only; complex enough to warrant a footer, none present |

## Fixing What You Confirm

`communication.md` gives the fix in one line: **end on the fact.** State the
mechanism or the number and stop. If the significance genuinely is not obvious
from the fact, add one plain clause — never a mirrored or generalizing one.

The script never rewrites. It locates; you decide.

## Hook

`hooks-plugin/hooks/prose-house-style-nudge.sh` runs the same check
automatically after a `Write`/`Edit` of a `.md` file and before
`gh pr create` / `gh issue create`. It **never blocks** — style is a nudge, per
`.claude/rules/hook-block-vs-nudge.md` — and is opt-in behind
`CLAUDE_HOOKS_ENABLE_PROSE_CHECK=1`.

## Agentic Optimizations

| Context | Approach |
|---|---|
| One draft file | `prose-check.sh --kind <kind> <file>`; read the rollup, open sections only if `ISSUE_COUNT` > 0 |
| Many files | pass them all in one invocation — the script batches and reports per file |
| Tool missing | check the `*_AVAILABLE=false` line before concluding a document is clean |
| Pre-commit / CI | add `--strict` to turn candidates into a non-zero exit |

For the three-layer architecture, the optional-tool/binary-discovery behaviour,
and the vale/segmenter mechanics that bite when adding a House rule, see
[REFERENCE.md](REFERENCE.md).