install

harness-kit

You describe a feature. Claude Code writes the spec, the plan, the code, and the tests, and opens the PR. Nothing moves forward until it passes a check.

Start

1 min read

harness-kit is a set of Claude Code agents that carry one idea through three phases: Spec, Build, and Ship. Every step writes a markdown file, a script checks its structure, a judge scores it, and only a passing step unlocks the next. You decide twice: the direction, and the PR.

Use it for features worth doing right. Skip it for one-line fixes and throwaway spikes. To see it move first, pull an idea through the pipeline in your browser.

/plugin marketplace add Pierry/harness-kit
/plugin install harness-kit@harness-kit

Restart Claude Code, open your repo, run /harness-kit:install, and restart again. It asks which eval judge to use and whether to turn on graph engineering; the defaults are fine. You need python3, git, and the gh CLI. To update: /plugin update harness-kit, restart, /harness-kit:update.

Phase 1: Spec

1 min read

You type a four-line brief, or build one in the brief builder.

/golden-path

Squad: checkout
Problem: Returning guests abandon checkout when a card is declined once.
Hypothesis: If we add one-tap retry, completion rises 5 points.
Success metric: checkout completion, from 71% to 76% within 30 days

It writes the PRD (problem, customers, metrics, rollout, risks) and you approve the direction: a wrong problem caught here costs one rewrite, not a feature. Then it writes the PRP, the engineering spec, with the files to change and acceptance criteria.

With Jev, Jev scores both documents instead of Claude. With graph engineering, each criterion gets a stable id like REQ-001 and trace/{feature}.yml is created.

Phase 2: Build

1 min read

It writes the plan (files, order, risks, tests), implements it in small commits that follow your conventions and linters, and runs your tests. Each step loops until it passes:

flowchart LR
    write[agent writes] --> check{structure ok and score 8.0+?}
    check -->|no, up to 3x| write
    check -->|yes| next[next step]

The score comes from small yes/no checks, such as "every metric has a baseline", so a failure names exactly what to fix. With graph engineering, the plan records each file it expects to touch, and dev fails if it touches any other file without a written reason.

Phase 3: Ship

1 min read

It writes the PR (summary, test plan, links) and you approve it; it opens as a draft. A monitor clears the pipeline when it merges.

With graph engineering, the PR carries a table linking each requirement to its code and the test that proves it, and the next feature marks those links VALIDATED or STALE. Every approved step logs its score to the quality dashboard, so you can see quality move over time. The cockpit shows it all live in a terminal: npm i -g @pieerry/harness-kit, then hk-tui.

Options

1 min read

Switch any time: /hk:eval jev or local, and /hk:graph manifest, full, or off. full adds a local graph of your decisions and call graph (FalkorDB with Graphiti on free NVIDIA models, and Joern), with no Docker. Install with HK_STATUSLINE=1 if you want the pipeline in your status line. When installed, semble, repowise, and context7 sharpen code search; otherwise it uses grep.

Learn more

1 min read

The wiki, also in Portuguese and Spanish, has Getting Started with every step in detail, how scoring works in Evals and Jev and System One, traceability in Graph Engineering and Graph Theory, and the sources in References. The method is harness engineering from Birgitta Böckeler. The three agents (product-manager, staff-software-engineer, system-architect) are mapped in AGENTS.md.

One honest limit: neither judge is checked against human ratings yet, so treat 8.0 as a signal, not a measurement. Issues and PRs welcome. MIT license.