Home

project-harness

A Claude skill that sets up a software project before implementation starts, then re-audits it at every phase boundary — so agentic work doesn't drift, and unknowns get surfaced before they become wrong code.

The problem it attacks

An agent that hits a gap in its context does not stop. It fills the gap with a plausible, industry-average assumption and keeps going — producing coherent, confidently wrong work. And even a well-set-up project decays: a rule gets softened during a refactor, a status gets hand-edited under deadline pressure, a doc keeps describing a world that no longer exists. Guardrails fail silently.

project-harness attacks both failure modes:

• a setup mode that closes context gaps before code is written, and
• a phase audit that verifies the guardrails still physically work — by feeding them fake violations and requiring rejection, not by reading documents that claim they work.

Two sources of authority, one spine

The skill distills a know-your-unknowns discovery playbook, a Karpathy-style minimal CLAUDE.md of behavioral rules, and a production harness grown over months on a live algorithmic-trading build. They independently converge on the same spine: surface assumptions before code; verify against executed checks in a loop.

Start here

• Getting Started — install the skill and run your first setup + audit.
• Discovery and the Four Quadrants — how context gaps are closed before code.
• Tiers — how the harness is right-sized (S / M / L) and grown on felt pain.
• The Phase Audit — the seven integrity + drift checks run at each boundary.
• Design Principles — the load-bearing rules behind every mode.

Source of truth

This wiki is the reader's-eye introduction. The repository files are canonical for exact procedure: SKILL.md is the entry point, and the references/ directory holds the authoritative detail for each mode. When this wiki and a repo file disagree, the repo file wins.