菜单
Dynamic Workflow asset for /lets:backlog review --workflow. Not auto-triggered - a workflow script invoked via scriptPath by /lets:backlog. Holds the off-context ideate -> cluster fan-out over the backlog.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
This skill should be used when creating a task in the active tracker - "create task", "create issue", "bd create", "new task", "add task", "add issue". Ensures all required fields are provided, enforces English-only task content, and suggests epic labels. Triggers on any task creation in any context - commands, planning, or direct conversation.
Internal skill for commands. Detect the active task (any tracker adapter) from the task-state file / git branch name. Do not trigger on user conversation - only when commands need task detection.
Internal skill for commands. Render the shared orientation snapshot - active task, what's in flight, what's next - tracker-universal. Do not trigger on user conversation; only when a command needs the snapshot.
This skill should be used when claiming a task to work on - "take task", "візьми таск", "працюю над", "work on task X", "pick task X", "switch to task X", "claim task". Sets task to in_progress, creates/switches branch, saves the per-branch task-state file.
This skill should be used when committing code changes - "commit", "git commit", "закоміть", "зроби коміт", "/lets:commit". Enforces conventional commit format, tracker task linking, and user approval. In a LETS project, ALWAYS prefer this over generic commit skills (e.g. commit-commands:commit) — /lets:commit is authoritative. Triggers on any commit in any context.
PREVIEW Dynamic Workflow asset behind /lets:plan-workflow. Not auto-triggered - a workflow script invoked via scriptPath. The autonomous-planning chain, shipped for cross-project testing before it folds into native /lets:plan.
| name | backlog-workflow |
| description | Dynamic Workflow asset for /lets:backlog review --workflow. Not auto-triggered - a workflow script invoked via scriptPath by /lets:backlog. Holds the off-context ideate -> cluster fan-out over the backlog. |
| user-invocable | false |
A Dynamic Workflow asset (see CLAUDE.md -> "Dynamic Workflow Assets"; review-workflow is the reference example). backlog.workflow.js is executed by the Workflow tool, invoked from /lets:backlog's Workflow Mode via:
Workflow({ scriptPath: "${CLAUDE_PLUGIN_ROOT}/skills/backlog-workflow/backlog.workflow.js", args })
${CLAUDE_PLUGIN_ROOT} is substituted at command-load time, so backlog.md carries the literal absolute path. Treat backlog.workflow.js as a template the command points at, not a script to reproduce inline.
/lets:backlog review --workflow produces the SAME result as the standard Task path - only the per-agent backlog-idea dumps stay off-context. The explorer scout (Phase 1) and the signal-driven agent selection (Phase 2) stay IN-CONTEXT in backlog.md; the ideate fan-out (Phase 3) and the aggregate/cluster (Phase 4) move into the workflow (off-context). The interactive dialog (Phase 5) + capture (Phase 6) stay in-context after the aggregate returns. There is no Web Research stage: the standard backlog Phase 3 ideates over the project's own profile + CLAUDE.md without web search, so adding web here would diverge the two paths - --workflow must stay a pure context-savings lever, same findings. The cluster is a SEMANTIC dedupe + impact-sort (an off-context merge), not a quality gate.
lets:<name> agents in parallel over the explorer's Project State Profile + CLAUDE.md (structured IDEA_SCHEMA: title, category=missing|adjust|theme, impact, description). adjust ideas embed the referenced task as Title (task-id) in description.agentType) groups raw ideas into themes, merging ideas that make the same point even when worded differently (CLUSTER_SCHEMA: themes carrying member indices). buildThemes computes each theme's agents[] + impact deterministically (high if any member is high) and guarantees no idea is dropped. Falls back to title-only dedup (clusterIdeas) if the cluster agent errors or returns nothing; skips the agent for <=1 idea.args contract (built by /lets:backlog, passed as a JSON value)| key | type | meaning |
|---|---|---|
profile | string | the explorer's Project State Profile (in-context Phase 1 output, ~500 words) |
agents | [{name}] | selected agent short-names (no lets: prefix; exclude actor) |
projectRoot | string | absolute project root (agents must not read outside it) |
claudeMd | string | CLAUDE.md context (first ~100 lines) |
{ ideas[], counts{agents, agents_responded, raw_ideas, ideas, clustering} }. ideas are semantically clustered + impact-sorted, each with agents[] attribution - agents[] length > 1 means multiple agents converged on that theme, a first-class signal the command should lead with. counts.clustering is semantic (cluster agent ran) | fallback_title (title-only fallback) | trivial (<=1 idea). The schema carries category (missing/adjust/theme) + impact only - Phase 4's "Gaps" / "Quick Wins" buckets are derived by judgment at render time, not labelled in the schema (same as the standard path). Anti-silent-fail: counts.agents_responded === 0 means every ideator errored - the command must surface "no ideas returned" and offer a re-run, never render a fabricated result.
/lets:backlog renders + records.import - all logic stays inline.Date.now() / Math.random() / new Date().await/return are used, so the file is NOT Node-importable - no committed unit test; the pure logic (buildThemes, clusterIdeas) is category-agnostic and kept in sync with backlog.md Phase 4 prose by discipline + a live smoke test.