Agent Integration

Wiring writetighter into an agent workflow requires a small skill file to handle the shell calls and output rules. Use one of these two patterns:

• Local model — the agent runs revise directly against a configured endpoint.
• Delegated model — the agent runs prompt to get revision instructions, then hands them to a subagent.

Both examples below are templates. Adapt the frontmatter and command structure to your specific agent harness. Both follow the core writetighter rules: it never edits files, and clarification questions are asked, never guessed.

Local model

Use this when your agent already has a configured model endpoint and can call revise directly.

---
name: tighten
description: Run drafted prose through writetighter before shipping. Reports findings; you apply them.
argument-hint: '[text or file to tighten; optional --kind]'
---

$ARGUMENTS: the prose to tighten, plus an optional `--kind`

1. Resolve `$KIND` (default `description`) and `$DRAFT`.

2. Lint (deterministic):

   `printf '%s' "$DRAFT" | writetighter lint --stdin --kind "$KIND" --format json`

3. Revise (contextual):

   `printf '%s' "$DRAFT" | writetighter revise --stdin --kind "$KIND"`

4. Process revisions:
   - For each `rewrite`: Adopt `replacement` only if it preserves all commands, paths, numbers, and identifiers in `source_text`.
   - For each `clarification`: Surface the `question`. Never guess.

5. Update `$DRAFT` with accepted changes, re-lint, and present the result.

Delegated model

Use this when the agent has no configured endpoint, or when you want to delegate the model call to a cheaper, specialized subagent.

---
name: tighten-delegate
description: Tighten prose by delegating revision instructions to a subagent.
argument-hint: '[text or file to tighten; optional --kind]'
---

$ARGUMENTS: the prose to tighten, plus an optional `--kind`

1. Resolve `$KIND` and `$DRAFT` as in the local-model version.

2. Lint locally:

   `printf '%s' "$DRAFT" | writetighter lint --stdin --kind "$KIND" --format json`

3. Get revision instructions:

   `writetighter prompt --kind "$KIND" --format json`

4. Spawn a subagent (e.g. Haiku) with the prompt text and `$DRAFT`. Instruct it to return the same `revisions[]` JSON shape as `revise`.

5. Apply the subagent's output following the same verification rules as the local-model version.

Which one to use?

• Local model: Use it if you got it!
• Cloud model: Use with an api key if you like. No need for a pricey one!
• Delegated model: Good if you've just got a "coding plan" and don't want to configure any API stuff.

Both rely on the commands documented in the CLI reference.

Packaged plugin

For Claude Code or Codex, you can skip the manual skill file and install the bundled tighten plugin instead — see the Claude Code plugin and Codex plugin guides. The plugin formalizes the workflow below, including subagent delegation when revise is unconfigured.