create-harness

Create Harness

You are creating a reusable harness — a set of instructions that defines how to process tasks from any workflow. Harnesses are stored in .claude/harnesses/ so they can be applied to different workflows.

Step 1: Gather Configuration

If a harness name was provided as $ARGUMENTS, use it. Otherwise, ask.

Prompt the user for:

1. Harness name — short, slug-friendly (e.g., tdd, review-and-fix, research-only)
2. Description — one sentence describing what this harness does
3. Steps — walk the user through defining the harness behavior:
   • What should happen for each task? (e.g., "write tests first, then implement, then verify")
   • Are there constraints? (e.g., "never modify files outside src/", "always run tests after changes")
   • Should it pause between tasks or continue automatically?
   • Any special handling for blocked tasks?
   • Any verification or quality gates before marking a task done?

Step 2: Write the Harness

Read the harness template from harness.yaml.tmpl.

Create .claude/harnesses/<harness-name>.yaml by populating the template:

• {{HARNESS_NAME}} — the slug name
• {{DESCRIPTION}} — the description sentence
• {{CREATED_DATE}} — today's date in ISO format (e.g., 2026-03-29)

Populate steps, constraints, done_criteria, and pause_between_tasks based on the user's answers from Step 1.

Translate the user's description into concrete, actionable steps. Each step's instruction should be clear enough that Claude can follow it without ambiguity.

Step 3: Confirm

Tell the user:

• The harness was created at .claude/harnesses/<name>.yaml
• Summarize the steps and constraints
• How to use it: /checkpoint:run-harness <harness-name> <workflow-name>
• Remind them harnesses are reusable across any workflow

Rules

• Create .claude/harnesses/ if it doesn't exist
• If a harness with the same name exists, ask whether to overwrite or pick a new name
• Steps should be specific and actionable — avoid vague instructions like "do the work"
• Constraints should be enforceable — things Claude can actually check
• Done criteria should be verifiable — things Claude can confirm before marking a task complete