跳到正文
cccskills
免费在 GitHub 公开

quick

Tracked lightweight execution with composable rigor flags: --trivial, --discuss, --research, --full. Covers zero-ceremony inline fixes (typo, spelling fix, small mistake in a single file, ≤3 edits) through contained multi-file changes.

中文简介正在准备中。当前显示原文说明。

查看安装方法

包含的文件(2)

  • SKILL.md7.4 KB
  • references/templates.md2.6 KB

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

/quick

Make contained changes in one execution thread. Use /do for multiple components, architectural changes, or parallel work. Keep plans inline; do not create task_plan.md.

Flags

All flags default off. Combine discussion, research, and verification as needed.

FlagEffect
--trivialMechanical fix: at most 3 edits across at most 3 files; no displayed plan, task ID, or subagents.
--discussResolve independent ambiguities together before planning.
--interviewResolve dependent decisions one question at a time, with a recommendation.
--researchInvestigate unfamiliar code before planning.
--fullVerify the plan, then run affected tests, configured lint, and diff review.
--no-branchStay on the current branch if repository branch rules permit it.
--no-commitLeave edits uncommitted, including in trivial mode.

Deep References

WhenLoadContent
Emitting banners, commit format, STATE.md entriesreferences/templates.mdOutput and tracking contracts

Setup

Read repository CLAUDE.md unless already loaded and current. Parse flags; the remaining text is the task. Follow existing authorization and repository constraints throughout.

Check the branch during setup. Never let --trivial or --no-branch bypass branch safety. For trivial work on main/master, create quick/<brief-description>; for standard work, create quick/<task-id>-<brief-kebab-description> after assigning the ID in step 2 and before editing, unless --no-branch is allowed. Preserve unrelated changes.

Trivial mode

Use for --trivial or a clearly mechanical one-line change identified by the router.

  1. Read the targets and check scope. Investigation or unfamiliar behavior requires /quick --research; more than 3 files, new package imports, or dependency-file changes require standard /quick. For ambiguity, ask one clarifying question; if unresolved, use /quick --discuss.
  2. Edit directly and count edits. If more than 3 are needed, preserve completed work and continue standard /quick. Explain the scope change and carry forward the original request and completed edits.
  3. Review the diff and run applicable repository checks. Stage only intended files with git add <specific-files> and commit using references/templates.md, unless --no-commit applies.
  4. Emit the trivial summary and stop. Report an omitted commit as skipped, never as successful.

Standard procedure

1. Resolve decisions and investigate

Use discussion for --discuss or material uncertainty about the requested change, approach, or acceptance criteria. Batch independent questions using the DISCUSS template. Wait for answers needed to proceed; do not ask again about decisions or actions already authorized.

For --research or unfamiliar code, read relevant source, tests, and configuration. Establish current behavior, where the change fits, and what could break. Summarize findings and their effect on the plan in 3–5 lines.

2. Plan and assign an ID

Use YYMMDD-xxx, with a Base36 sequence: 001 through 009, 00a through 00z, then 010.

date_prefix=$(date +%y%m%d)

Increment today's highest sequence in root STATE.md, starting at 001 if absent. If corrupted, recover the sequence from git log entries matching Quick task YYMMDD-. Increment again on branch-name collision.

Display the inline plan from references/templates.md: intended edits, files, rationale, and estimated edit count. Create the branch under the setup rules. With --full, verify that the plan meets acceptance criteria before editing. Recommend --full for security, payments, or data migration.

3. Execute and watch scope

Make the planned edits and track their count. Above 15 estimated edits, suggest /do. Warn at 10 actual edits and reassess at 15. These standard-mode thresholds are advisory; continue within established authorization when the work remains contained. Ask only when a scope decision is unresolved. Trivial mode's 3-edit limit remains strict.

For base verification, run an appropriate syntax/build check, such as python3 -m py_compile <files>, go build ./..., or tsc --noEmit, plus required repository checks.

With --full, run tests for affected packages/modules, configured lint on changed files, and git diff review for unintended changes, missing error handling, and broken imports. Run the full suite when required by the repository or requested by the user. Fix relevant failures before claiming completion; report unrelated blockers accurately.

4. Commit and log

Unless --no-commit, stage specific intended files with git add <specific-files>, use the conventional commit format from references/templates.md, and include Quick task <task-id> in the body. Verify with git log -1 --oneline.

Create or append to root STATE.md using the reference schema. Use tier trivial->quick after escalation; otherwise quick. Record skipped commits explicitly. Emit the completion summary with changes, checks, commit or skipped status, branch, flags, and log location. Continue any already-authorized delivery steps.

Examples

Base mode: /quick add --verbose flag to the CLI -- Generate ID 260322-001, plan 3 edits (flag definition, handler, help text), create branch quick/260322-001-add-verbose-flag, execute, commit, log.

With research: /quick --research fix the timeout bug in auth middleware -- Read auth middleware first, trace call path, then plan and execute.

Escalated from trivial: --trivial hit 3-edit limit across 5 files. Quick picks up with context, plans remaining edits, commits all changes, logs as tier trivial->quick.

Full rigor: /quick --full update payment amount rounding logic -- Plan edit, execute, run tests + lint + diff review, commit.

Error Handling

ErrorCauseFix
Task ID collisionTwo tasks with same sequenceIncrement sequence. If STATE.md corrupted, scan git log for Quick task YYMMDD- to find next ID.
Scope exceeds quick tierTask grows beyond contained workSuggest /do for multi-component or architectural changes. >15 edits is advisory.
Test failure in --full modeQuality gate found issuesFix failing tests. If fix needs significant work, note in STATE.md and suggest follow-up /quick.
Branch conflictBranch quick/<id>-... existsIncrement task ID sequence and retry.

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

同一仓库的技能

accounts

免费

Manage multiple Claude Code accounts: add, list, check, launch, and install shell aliases for 10+ isolated CLAUDE_CONFIG_DIR profiles.

中文简介正在准备中。当前显示原文说明。

notque/vexjoy-agent4422026年10月17日 更新

Improve architecture across modules by deepening interfaces.

中文简介正在准备中。当前显示原文说明。

notque/vexjoy-agent4422026年10月17日 更新

Assessment: read-only inspection, codebase overview, value analysis, health checks, ADR consultation, decision analysis, multi-perspective critique.

中文简介正在准备中。当前显示原文说明。

notque/vexjoy-agent4422026年10月17日 更新

Background memory consolidation — overnight review, merge, and injection payload for memory files.

中文简介正在准备中。当前显示原文说明。

notque/vexjoy-agent4422026年10月17日 更新

Jev-driven browser automation: Jev picks operations, programs execute, a text model writes field values only when Jev cannot pick one from the goal.

中文简介正在准备中。当前显示原文说明。

notque/vexjoy-agent4422026年10月17日 更新

Write, compose, integrate, and improve programs that call Jev, TypeSafe's System One judgment model.

中文简介正在准备中。当前显示原文说明。

notque/vexjoy-agent4422026年10月17日 更新

查看 notque 的全部技能

报告此技能的问题