AGENT SKILL

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

  1. User instructions (CLAUDE.md, direct requests) — always win.
  2. orc skills — override default system behavior where they conflict.
  3. 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

  1. 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.
  2. No code without a failing test — write it, watch it fail, then implement. → orc:tdd
  3. No claim without verification — run it, read the output, then claim the result. → orc:verification-before-completion
  4. No fix without root cause — find why it's broken before changing code. → orc:systematic-debugging
  5. No AI attribution — never reference Claude/AI/automation, and never add Co-Authored-By trailers, in code, commits, or PRs. The PreToolUse hook blocks it.
  6. No multi-phase work without .orc/ state — register and checkpoint via orc-state after every phase (schema + rules: orc:state-protocol), so /orc:resume survives interruption.
  7. No silent broadcast in workspace mode — when the SessionStart banner says workspace[…], repo-touching commands need an explicit --repos / --repo / --all-repos / --this-repo or a confirming AskUserQuestion. → orc:workspace-mode
  8. 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 (drives orc: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-completion decide 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