Harness Kit (as a skill)

This skill packages the Claude Harness Kit — a portable, stack-agnostic setup ritual
plus building philosophy. The kit's documents live in this same directory.

Pick the flow first:

• Brand-new project → drive from claude-project-kickoff.md (steps below).
• Existing, working project → drive from claude-project-adoption.md instead: it
  re-sequences the same kit for a live codebase, and its first step is three gated beats —
  evaluate the repo thoroughly (read-only), interview the user for what the code
  can't answer, then propose a written, adversarially-reviewed plan the user approves
  before any edit. Only then: floor, baseline, contract, wiki-from-history. It points back
  into the kickoff guide's sections as it goes — keep both at hand.

How to run a kickoff (new project)

1. Read claude-project-kickoff.md — the driver. It contains the one-time machine
   hardening (Part 0), the per-project setup ritual with its nine-question intake
   (Part 1), the building principles (Part 2), and the autonomous/multi-agent playbook
   (Part 3).
2. Execute Part 1 for this project, adapting every example to the user's actual stack —
   the guide is explicit that every concrete tool it names is an example, not an
   assumption.
3. Pull companions when their step comes up, not before:
   • llm-wiki-kickoff.md → §1.5b, before scaffolding the wiki (Standard tier and up)
   • claude-audit-base.sh → §1.6, copy to scripts/audit.sh, wire the TOOLING section
   • claude-eval-base.sh + evals-template/ → §1.6b, copy to scripts/eval.sh + evals/ (non-throwaway — the judgment verifier)
   • scripts/harness-metrics.sh + HARNESS_LOG.md → §1.6a, copy the script into the project + seed HARNESS_LOG.md at the repo root (non-throwaway — the ROI gauge)
   • readme-template.md → §1.5c, the human-facing README
   • prd-template.md → §1.7, hand to the user to fill in if no spec exists
   • templates/ + CHEATSHEET.md → Part 0 / §1.3, the settings layers
   • securing-claude-sessions.md → only if the user wants the security model explained
4. Confirm setup, remind the user to enter auto mode and restart so the sandbox
   initializes (§1.7), then ask for the spec.

Boundaries (the kit's own)

• The kit is scaffolding, used once. Never copy its files into the project repo, and
  never @-import them from the project's CLAUDE.md. Only the ritual's outputs
  persist: CLAUDE.md, .claude/settings.json, scripts/audit.sh, scripts/eval.sh,
  scripts/harness-metrics.sh, evals/, HARNESS_LOG.md, wiki/, README.md, the filled-in PRD.
• This skill cannot install the machine-wide managed floor (it is root-owned by design) —
  walk the user through Part 0 by hand instead.
• If the project already exists and is set up, this skill does not apply: work from the
  project's own CLAUDE.md and wiki, not the kit.