using-orc
orc's iron rules, instruction priority, and skill-routing discipline. Injected at every SessionStart; required reading before your first action.
- License 许可
- MIT
- 所属仓库
- HigorAlves/orc
- 最近更新
- 2026-11-07
- Stars
- ★ 7
- Forks
- 0
01 前言 Front Matter
- name
- using-orc
- description
- orc's iron rules, instruction priority, and skill-routing discipline. Injected at every SessionStart; required reading before your first action.
- user-invocable
- false
- disable-model-invocation
- true
03 SKILL.md 正文
<SUBAGENT-STOP> If you were dispatched as a subagent to execute a specific task, skip this skill. </SUBAGENT-STOP>
<EXTREMELY-IMPORTANT>
If there is even a 1% chance an orc skill applies, you MUST invoke it with the Skill tool — not optional, and you cannot rationalize past it. Check for a relevant skill BEFORE responding, before clarifying questions, and before exploring the codebase: "it's just a quick question" and "let me explore first" are rationalizations — questions are tasks. Announce "Using [skill] to [purpose]", then follow it. If an invoked skill turns out wrong for the situation, you don't have to use it.
</EXTREMELY-IMPORTANT>
Instruction priority
- User instructions (CLAUDE.md, direct requests) — always win.
- orc skills — override default system behavior where they conflict.
- Default system prompt — lowest.
Instructions say WHAT, not HOW: "Add X" / "Fix Y" never means skip the workflow.
Accessing skills
Use the Skill tool — never Read a skill file. Every skill's description is already loaded; invoke the one that fits.
Iron rules
- No commits to main/master/develop — branch first; check the branch before any commit. The PreToolUse hook downgrades protected-branch commits/pushes to a confirm prompt — approve it only with explicit user consent.
- No code without a failing test — write it, watch it fail, then implement. →
orc:tdd - No claim without verification — run it, read the output, then claim the result. →
orc:verification-before-completion - No fix without root cause — find why it's broken before changing code. →
orc:systematic-debugging - No AI attribution — never reference Claude/AI/automation, and never add
Co-Authored-Bytrailers, in code, commits, or PRs. The PreToolUse hook blocks it. - No multi-phase work without
.orc/state — register and checkpoint viaorc-stateafter every phase (schema + rules:orc:state-protocol), so/orc:resumesurvives interruption. - No silent broadcast in workspace mode — when the SessionStart banner says
workspace[…], repo-touching commands need an explicit--repos/--repo/--all-repos/--this-repoor a confirmingAskUserQuestion. →orc:workspace-mode - No PR over the size budget without a recorded choice — default 300 LOC (additions+deletions, post-exclusion); over budget → stack / open-big with a
Size-budget-override:trailer / abort. →orc:pr-size-budget. A web change → browser QA with.orc/<branch>/files/qa/artifacts, no exceptions. →orc:qa(drivesorc:agent-browser)
Gate taxonomy (interaction policy)
Gates come in three classes. Hard-outward — anything writing to a tracker, posting a PR review, or publishing evidence: always asks; no flag or policy skips it. Soft-inward — decisions about the user's own working tree (plan approval, driver choice, PR compose): convertible to pre-approved defaults by the interaction_policy autopilot ladder (manual | guided | auto, via userConfig or --auto; banner shows the resolved level). Escalation-only — never asked proactively; always stop the run when triggered (implementer escalations, env failures, criteria misses), at every policy level. How every gate is rendered (header chips, Recommended-first, preview, multiSelect chunking): → orc:gates
Skill routing
- Process skills first —
orc:systematic-debugging,orc:tdd,orc:verification-before-completiondecide HOW to approach a task; implementation skills come second. - Rigid skills (tdd, systematic-debugging, verification) — follow exactly. Flexible skills — adapt the principles. The skill itself tells you which.
- Stack packs (skills auto-surface via their own descriptions; these just group them by stack, load on demand): web-react · backend · ios · workflow-extras.
Insights & callouts
When writing substantive code, surface 2–3 brief, codebase-specific notes as emoji-header blockquotes — > **💡 Insight** / > **⚠️ Caution**. → orc:callouts
Same palette for must-not-miss command output: 🛑 danger, ⛔ gate, 📋 preview, ➡️ next. Terminal output carries no [!TYPE] tag (it prints as junk in the TUI); add the GitHub tag ([!IMPORTANT]/[!WARNING]/[!CAUTION]/[!NOTE]/[!TIP]) only in GitHub-bound output — PR bodies, review comments, committed docs.
04 同仓库更多 Skills
writing-plans
Write a comprehensive implementation plan of bite-sized TDD tasks that an engineer with zero codebase context could execute — exact files, code, tests, and verification steps
writing-for-agents
orc's doctrine layer for writing any document an agent consumes — skills, AGENTS.md/CLAUDE.md, reference docs
workspace-mode
Use when cwd is a parent dir holding multiple sibling git repos — defines repo auto-detection, --repos/--repo/--all-repos/--this-repo precedence, per-repo dispatch, and linked-PR mechanics.
wizard
Generate an interactive bash wizard walking a human through work only they can perform — credentials, secrets, one-off migrations
wait-what
Stop and re-pitch — the last message did not land; say it again simply, in the project's own vocabulary.
vitest
Vitest fast unit testing framework powered by Vite with Jest-compatible API