instantiate-harness

instantiate-harness

This is the only skill a user runs once per project (or per task type within a project). It scaffolds .harness/ from templates and the user's answers.

When to invoke

• User says they want to set up the harness for a project.
• User has a new kind of task in an existing project (different domain → different .harness/).
• A .harness/ already exists but is broken and needs to be rebuilt from scratch.

Procedure

1. Confirm the working directory is the project root. If a .harness/ already exists, ask before overwriting.
2. Run the wizard (see wizard.md for the full script). Eight questions, all multi-choice with sensible defaults:
   • Domain (code · research · writing · outreach · data · design · other)
   • The user prompt that triggers a run (free text)
   • Verification action (domain-specific suggestions)
   • Top 3–5 weighted criteria
   • Model per role (planner · generator · evaluator)
   • Session strategy (reset · continuous)
   • Budget cap (iterations · hours · dollars)
   • Pivot policy (consecutive-failure threshold)
3. Write files into .harness/:
   • config.json from templates/config.template.json with the answers filled in
   • planner.md, generator.md, evaluator.md from templates with placeholders substituted
   • rubric.json from the criteria the user picked
   • calibration/README.md instructing the user to drop 3–5 worked examples
   • empty state/ and traces/ directories
4. Print next steps for the user:
   • "Drop 3–5 calibration examples into .harness/calibration/ before the first run."
   • "When ready, run /run-harness."

Rules

• Never write into the plugin repo from this skill. Only writes go to the project's .harness/.
• Confirm before overwriting an existing .harness/.
• Use AskUserQuestion, not free-text prompts. The wizard is meant to be quick.
• Domain-specific defaults: when the user picks "research", default the verification action to "walk every citation and fact-check". When they pick "outreach", default to "dry-run each message against the rubric and grep for template-smell". The wizard branches on the domain answer.

What good looks like

A successful instantiation produces a .harness/ that:

• Validates against the JSON schemas under runtime/schemas/.
• Contains four prompt files that read coherently to a human.
• Has at least one verification action wired to at least one criterion.
• Is runnable: /run-harness --dry exits without schema errors.

Anti-patterns

• Asking the user to write the entire planner prompt themselves. The skill writes it from a template; the user only fills the slots.
• Skipping calibration. Even three rough examples beat zero. Tell the user explicitly.
• Letting the user pick more than 5 criteria. Push back. Reference principle P7.