Claude Code harness

Curated,
not dumped

A small Claude Code toolkit for verifiable AI-assisted work: 10 skills, 2 agents, 5 rules. Most collections optimize for count; this one keeps only what survived real use.

View on GitHub Release v0.2.0 How it evolves

What's inside

10 skills

Only what earned its place

vet, proof, workflow-upgrade, goal, konsilium, ship-secure, pre-push, system-health, teach, longread. Each one is a procedure that survived real use, not a prompt someone thought sounded good.

2 agents

Two by design, not sixty

An adversarial verifier that tries to refute a finding instead of confirming it, and a read-only scout that reads an incoming tool's repo and reports before anything is installed.

5 rules

Always-on behavior

Safety guards (dependency intake included), code principles, root-cause discipline, detective mindset, and a think-before-code checklist for new projects.

One example

A generic case, not a logged run: how the ladder grades one claim with /proof.

1 · Claim

"The fix is done and deployed"

An agent reports it. The proof skill atomizes it: the code was changed, the change is in the running process, the process does the work, the work has the intended effect.

2 · What was missing

Rung 4, Integrated

The diff is in place and the tests pass, but nothing shows the process was restarted. The new code may not be running. The skill names this false close: "It synced, therefore it runs."

3 · Evidence record

Verdict: partial

Reached: rung 2 Implemented and rung 3 Exercised. Not reached: rung 4 Integrated and rung 5 Observed. Green tests without a primary check are partial, never pass.

The evidence ladder

The newest piece, and the reason for release v0.2.0. An agent's most expensive habit is not being wrong. It's reporting done on work nothing actually checked, in the same tone it uses for work that was.

1 · Declared - someone said so. A comment, a commit message, an agent's own summary. Not evidence.
2 · Implemented - the code exists and parses.
3 · Exercised - it ran at least once. A test, a dry run.
4 · Integrated - it ran in the real assembly: the deployed process serves it.
5 · Observed - it produced the intended effect, visible afterwards and attributable to it.

passpartialfailskip

Climb only to the lowest rung that decides the claim: expense is not evidence. And keep skip distinct from fail - "we didn't look" and "we looked and it's broken" call for opposite actions.

Why curated

Vet before adopt. Every incoming tool runs one pipeline where security is a blocking gate, and the verdict lands in a registry so nothing is evaluated twice.
It evolves. Corrections become memory, findings become skills, but only past a promotion gate, so confident guesses never get enshrined as truth.
Cherry-pick over giant. A good idea inside a huge framework doesn't justify installing the framework.
No auto-hooks. The package installs no automatic hooks. That is a design decision, not an omission.

Install

No installer, and no curl | bash: this harness practices what it preaches. Copy what you want.

git clone https://github.com/Sanexxxx777/curated-claude-code
cd curated-claude-code

cp -r skills/proof ~/.claude/skills/
cp -r skills/vet   ~/.claude/skills/
cp agents/*.md      ~/.claude/agents/

Curated,not dumped