> **Created**: 2026-03-19 > **Status**: Candidate Complete > **Priority**: P1 > **Deep Research**: conversation 2026-03-19 (2 agents, score 88/100)
Scanned 9/4/2026
Install to Claude Code
npx -y skills add sd0xdev/sd0x-harness --skill requests --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Requests?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/sd0xdev-requests-76f1863b)More formats (shields.io, HTML) on the badges page.
# Lightweight Correction Skill (`/remind`)
> **Created**: 2026-03-19
> **Status**: Candidate Complete
> **Priority**: P1
> **Deep Research**: conversation 2026-03-19 (2 agents, score 88/100)
## Background
模型在長 session 中會忘記 CLAUDE.md 和 Rules 的指示——特別是 behavior-layer 規則(auto-loop、doc sync、adequacy gate 等)。現有 hooks(stop-guard、post-tool-review-state)可以 enforce **completeness**(是否完成了必要步驟),但無法 enforce **timing**(是否在同一 reply 中執行)或偵測 **reasoning-level** 違規("Declaring ≠ Executing")。
`/next-step` 已有 16 個 deterministic heuristics,但它的定位是 "what to do next"——不是 "what you forgot"。需要一個輕量級 correction skill 填補 reasoning-level 盲點。
## Requirements
| 需求 | 說明 |
|------|------|
| User-invoked correction | `/remind auto-loop` 告訴模型忘了什麼 |
| Smart detection | `/remind` (no args) 讀取 state + git + context 智慧判斷遺漏 |
| Lightweight | 不做 research,只做檢查(< 5s) |
| Reuse existing infra | 重用 stop-guard state parsing + next-step heuristics |
## Scope
| Scope | Description |
|-------|-------------|
| In | State-based detection (review/precommit/doc-review), git status checks, user-specified rule reminders, correction action output |
| Out | Transcript NLP analysis(v2), new hooks |
> **Note**: v1 實作包含 correction execution(偵測到違規時自動執行修正指令),原 scope 規劃為 v2,但實作過程中決定納入 v1(見 SKILL.md:13 execution mandate)。
>
> **Security**: `<rule-name>` 解析為 `rules/<rule>.md`(直接路徑建構)。安全邊界為 Claude 的 Read tool(僅允許讀取 repo 內檔案)+ Glob fallback 列出可用規則。未來可增加 slug 格式驗證(`^[a-z0-9-]+$`)作為額外防護。
## Related Files
| File | Action | Description |
|------|--------|-------------|
| `skills/remind/SKILL.md` | New | Skill definition(detection heuristics + correction output) |
| `skills/remind/references/detection-rules.md` | New | Detection rule table + state file parsing |
| `commands/remind.md` | New | Command entry point |
| `test/commands/remind.test.js` | New | Schema + content tests |
| `CLAUDE.template.md` | Modify | Command Quick Reference 加入 `/remind` |
| `CLAUDE.md` | Modify | Command Quick Reference 加入 `/remind` |
| `.claude/CLAUDE.md` | Modify | Command Quick Reference 加入 `/remind` |
## Acceptance Criteria
### AC1: User-Invoked Mode
- [x] `/remind auto-loop` 輸出 auto-loop rule 摘要 + 當前違規狀態
- [x] `/remind git-workflow` 輸出 git 規則摘要
- [x] `/remind <rule-name>` 支援所有 `rules/` 下的規則名稱(解析為 `rules/<rule>.md`,fallback Glob 列出可用規則)
- [x] 未知規則名稱回傳錯誤 + 可用規則列表
### AC2: Smart Detection Mode
- [x] `/remind` (no args) 讀取 `.claude_review_state.json`
- [x] 偵測:has_code_change + code_review.passed=false → 提醒跑 review
- [x] 偵測:has_doc_change + doc_review.passed=false → 提醒跑 doc review
- [x] 偵測:code_review.passed + precommit.passed=false → 提醒跑 precommit
- [x] 偵測:git branch = main → 建議建分支
- [x] 偵測:git status dirty + no review state → 提醒開始 review loop
### AC3: Output Format
- [x] 輸出格式:finding list (priority + rule + correction action)
- [x] 無遺漏時顯示 "All clear" 訊息
- [x] 提供 copy-pasteable 修正指令(如 `/codex-review-fast`)
### AC4: Infrastructure
- [x] Tests pass
- [x] skills-schema.test.js pass
- [x] CLAUDE.md command tables updated
## Progress
| Phase | Status | Note |
|-------|--------|------|
| Deep research | Done | 2 agents, score 88/100 |
| Tech spec | Done | `docs/features/remind/2-tech-spec.md` |
| Implementation | Done | SKILL.md (195 lines) + detection-rules.md + commands/remind.md |
| Testing | Done | 8 tests passing (`test/commands/remind.test.js`) |
| Acceptance | Done | Codex review + precommit pass |
Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!