CLAUDE.md@claude git:20260916.a87d7deA

Immutable. This exact content is served forever at /api/v1/blob/aef4581de1fe09c5.

# Global instructions

These instructions come from an `agent-harness` checkout, installed at user level, and load in
every session. Topic rules live in `~/.claude/rules/harness/` and the chosen preference variants
in `~/.claude/rules/harness-stances/`; both load automatically. This file holds only what belongs
to no topic.

**Standing instruction on delegation.** Use subagents freely for information gathering —
locating files, reading, grepping, extracting, summarizing — without waiting to be asked, in
every repo and every mode. `delegation.md` carries the rules and the chosen `delegation` stance
carries the model tiers. Gather with delegation; keep judgment in the session.

| Rule file | Covers |
| --- | --- |
| `working-style.md` | Standing patterns: honesty over polish, verify before claiming, immutable dated artifacts |
| `decisions-and-plans.md` | How to present decisions, and what a pre-approved plan licenses |
| `delegation.md` | Delegate information gathering, bound what comes back, four prohibitions |
| `transcript-hygiene.md` | Read narrowly; cap and never reprint subagent output |
| `voice-and-format.md` | Output shape for subagents and deliverables; posts in the user's name |
| `research-and-verification.md` | Search budgets, and how to prove an API answer is real |
| `conciseness.md` | Explain a rationale once; never narrate code |
| `verification.md` | Run the gates locally; never open a PR on unverified work |
| `secrets.md` | No credentials in tracked files; stop and rotate on discovery |

**Stances** are preferences a reasonable user might hold differently. One variant of each is
linked in from `~/.config/agent-harness/config.json`: `licensing`, `build-vs-buy`, `commits`,
`plan-ceremony`, `delegation`, `testing`, `autonomy`. Read the linked variant, not this list.

**Skills** carry procedures and load only when invoked: `plan-authoring`, `delegation-tiering`,
`licensing-review`, `design-loop`, `workflow-status`, `harness-authoring`, `migration-safety`,
`upstream-contribution`, `worktree-per-agent`, `spike-contract`.

## Where an instruction belongs

Place by how widely it applies, most general first.

1. **How agents should behave anywhere** → a harness rule (generic) or a personal file in
   `~/.claude/rules/` (true only of you).
2. **A preference reasonable people differ on** → a stance variant.
3. **A family of repos** → a `CLAUDE.md` in the directory above them.
4. **One repo** → that repo's `AGENTS.md`, with `CLAUDE.md` symlinked to it.
5. **One kind of file** → a rule in that repo's `.claude/rules/` with a `paths:` glob.
6. **A procedure with steps** → a skill.
7. **Something that must run at a lifecycle point regardless of judgment** → a hook.
8. **A record of what happened** → auto memory.

A rule loads on every turn, so it stays short. A skill costs one line of description until it
fires, so it can be long. When a rule starts growing steps, it wanted to be a skill. The
`harness-authoring` skill applies this ladder whenever you are asked to add or change one.

@~/.claude/CLAUDE.personal.md