harness-skills-plugin

A Claude Code plugin implementing the multi-agent build harness architecture from Anthropic Engineering: Harness Design for Long-Running Apps.

Extends the original pattern with a codebase auditor for existing projects, file-based handoffs that survive context resets, and a polish phase for iterative refinement.

Architecture

flowchart TD
    U([User Prompt]) --> O

    subgraph Harness["Claude Code — Harness Skills"]
        O["🎯 Orchestrator\nharness-orchestrator\nSequencing · Gates · Handoffs"]

        O -->|new project| P
        O -->|existing project| A
        A["🔍 Auditor\nharness-auditor\nCodebase Analysis · Tech Debt · Constraints"]
        A -->|CODEBASE_AUDIT.md| P

        P["📋 Planner\nharness-planner\nProduct Spec · Calibration Anchors · Sprint Plan"]
        P -->|PRODUCT_SPEC.md| B

        B["🔨 Builder\nharness-builder\nSprint Execution · Git Commits · Sprint Contracts"]
        B -->|source code| E

        E["✅ Evaluator\nharness-evaluator\nQA · Scoring · Pivot Suggestions · Polish Mode"]

        E -->|pass| O
        E -->|retry| B
        E -->|polish loop| B
    end

    subgraph State[".harness/ — File-Based State"]
        direction LR
        H[HANDOFF.md]
        S[PRODUCT_SPEC.md]
        Q[QA_REPORT.md]
        C[config.yaml]
    end

    O <-->|read/write| State
    P <-->|read/write| State
    B <-->|read/write| State
    E <-->|read/write| State

    O -->|complete| Done([Shipped Product])

Key insight from the Anthropic post: separated evaluation beats self-evaluation. The evaluator reads builder output cold, from files, with a skeptical prompt — it never shares conversation context with the builder.

Skills

Skill	Role	Key Outputs
harness-orchestrator	Entry point, sequencing, manual gates	config.yaml, HANDOFF.md
harness-auditor	Codebase analysis for existing projects	CODEBASE_AUDIT.md
harness-planner	Prompt → product spec + sprint plan	PRODUCT_SPEC.md, CALIBRATION.md
harness-builder	Spec → working code, sprint by sprint	Source code, git commits
harness-evaluator	QA, grading, feedback, polish loop	QA_REPORT.md
harness-brief	Standalone project brief generator	PROJECT_BRIEF.md

How File-Based Handoffs Work

Each agent reads its inputs from .harness/ files and writes its outputs there — no conversation context is shared between phases. This means:

• Context resets are safe — resume any time with Resume the harness
• Subagent-ready — each phase can run as a separate Claude subagent
• Human-steerable — edit .harness/ files directly to change scope, thresholds, or sprint plans mid-run

Installation

/plugin marketplace add woink/harness-skills-plugin
/plugin install harness-skills@harness-skills-plugin

Usage

New project:

Spin up the harness: Build a real-time dashboard for NYC park conditions.

Existing project:

Run the harness on this project. I want to add real-time notifications.

Resume after context reset:

Resume the harness.

Unattended multi-session run:

bash scripts/run-harness.sh --max-iterations 20

Configuration

.harness/config.yaml controls thresholds and gates:

harness_version: 1
project_mode: greenfield        # greenfield | enhance | refactor | rescue
eval_mode: hybrid               # playwright | automated-tests | manual-gates | hybrid
max_sprint_retries: 2
qa_pass_threshold:
  functionality: 7              # hard floor — does it work?
  design_quality: 6
  code_quality: 6
  product_depth: 6
manual_gates:
  after_audit: true
  after_spec: true
  after_each_sprint: false
  after_final_qa: true

Cost Estimates

Phase	Typical Cost	Duration
Auditor	$1–5	5–15 min
Planner	$0.50–2	3–10 min
Builder (per sprint)	$20–70	30–120 min
Evaluator (per sprint)	$3–5	5–10 min

A 5-sprint build with 1 retry averages $100–200 total.

Based On

Anthropic Engineering: Harness Design for Long-Running Apps