🛠️ Harness Kit

Five executable skills for building autonomous agent harnesses that don't silently fail

From the author of Inside Claude Code — the reverse-engineering field guide to agent engineering.

The book teaches the patterns. This kit runs them.

---

What it is

A Claude Code plugin that turns the hard-won lessons of building autonomous agents into skills you invoke, not chapters you read. Each one operationalizes a failure you've probably already hit:

Skill	Invoke it when	What it does
failure-mode-audit	before a long run, or after one stalls	Audits a plan or transcript against the 7 known agent failure modes and prescribes the verified fix for each
claude-md-architect	onboarding a repo to Claude Code	Inspects the project + interviews you, writes a lean constraint-first CLAUDE.md
subagent-brief	spawning a sub-agent / swarm step	Writes a brief that survives the messaging asymmetry — so the sub-agent completes instead of silently aborting
context-handoff	nearing the context limit	Captures the minimum state to resume cleanly after a reset or compaction
harness-preflight	right before an unattended run	A 7-point go/no-go check on your harness before you hand it hours of work

Plus /harness-init — scaffolds a sound harness (CLAUDE.md + handoff convention + pre-flight) into any repo in one command.

Why it exists

Most "agent frameworks" hand you more abstraction. The failures that actually waste your time are smaller and more specific: an agent that declares a half-finished task done, a sub-agent that aborts because it was told to "wait for a message" it can never receive, a run that panics near its context limit with capacity to spare. These have names and fixes — Harness Kit encodes them so you stop rediscovering them by hand.

Install

# In Claude Code:
/plugin marketplace add aimasteracc/harness-kit
/plugin install harness-kit

# Then, in any repo:
/harness-init

Skills auto-activate by description, or invoke directly (e.g. "run a failure-mode audit on this plan").

What you get

• 5 skills + 1 command, all plain-text and inspectable — no black boxes, no runtime deps.
• Works alongside any other plugins; nothing to configure.
• Battle-tested framing from a full teardown of how a leading agentic coding tool behaves.

Go deeper

Harness Kit is the toolbox. Inside Claude Code is the manual — the five-layer prompt architecture, the request pipeline, all 12 patterns, and the full treatment of the 7 failure modes, in English / 简体中文 / 日本語.

License

See LICENSE.md. © 2026 aimasteracc. Independent work; not affiliated with Anthropic.