Overview

The harness is the authored layer on top of Claude Code: everything under .claude/ — skills, rules, agents, hooks — plus the documents that orient an agent in this repo (CLAUDE.md, the glossary, the learnings). Claude Code is the runtime; the harness is what makes an agent reliably useful here rather than generically capable.

This dashboard is a view-only render of that harness, generated straight from the source so it cannot drift.

The leverage is the harness, not the model: the model is maybe a tenth of what makes an agent reliable here, the rest is this authored layer. So when an agent misbehaves, the first suspect is the harness — a missing tool, a loose rule, a forgotten guardrail — not the model. Debug the harness first.

The closed loop

Work flows through the harness as a loop, not a line. An idea is shaped into a spec (brainstorm), red-teamed before it costs anything (check-reasoning), sliced into independently-shippable issues (to-issues), classified for readiness (triage), placed and ordered on the board (organise), then executed one gated commit at a time (build-sprint), looped across a whole epic by build-epic. Every change is verified against the gate (qa-tests, qa-code), and whatever the work taught is captured back into memory (find-learnings) — which the next plan reads first. The output of one turn becomes the input to the next.

The five pillars

The harness is audited along five pillars — context, tools, memory, guardrails, specialists — by qa-harness. Each has an owning skill: optimise-context for context, qa-skills for tools, find-learnings for memory. See the five pillars.

The spine

Underneath sits a fixed ordering of engineering principles — the spine. Read top to bottom; each clause depends on the one before it:

• permission before capability
• rollback before autonomy
• verification before delivery
• context budgets before long dialogue
• lifecycle before multi-agent
• institutions before proficiency

You don't grant a capability you can't permission, don't automate what you can't roll back, don't ship what you haven't verified.

Altitude

Every skill sits at one of three altitudes — how wide it operates, i.e. its blast radius. They nest: each higher altitude contains the ones below.

The wider the box, the wider the blast radius: L0 (qa-code, debug) touches one artifact, L1 (qa-skills, optimise-context, qa-philosophy) owns a whole pillar's set, and L2 (qa-harness) reasons over the entire harness.