Redesign Pattern 9 from "forbid compression" to smart dual-layer approach: - Layer 1: TodoWrite in_progress phase preserved by compact, completed phases compressible - Layer 2: Compact sentinel in phase files as fallback re-read trigger Applied to 6 workflow skills: skill-designer, lite-plan, multi-cli-plan, plan, tdd, test-fix
19 KiB
name, description, allowed-tools
| name | description | allowed-tools |
|---|---|---|
| workflow-skill-designer | Meta-skill for designing orchestrator+phases structured workflow skills. Creates SKILL.md coordinator with progressive phase loading, TodoWrite patterns, and data flow. Triggers on "design workflow skill", "create workflow skill", "workflow skill designer". | Task, AskUserQuestion, TodoWrite, Read, Write, Edit, Bash, Glob, Grep |
Workflow Skill Designer
Meta-skill for creating structured workflow skills following the orchestrator + phases pattern. Generates complete skill packages with SKILL.md as coordinator and phases/ folder for execution details.
Architecture Overview
┌─────────────────────────────────────────────────────────────────┐
│ Workflow Skill Designer │
│ → Analyze requirements → Design orchestrator → Generate phases │
└───────────────┬─────────────────────────────────────────────────┘
│
┌───────────┼───────────┬───────────┐
↓ ↓ ↓ ↓
┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐
│ Phase 1 │ │ Phase 2 │ │ Phase 3 │ │ Phase 4 │
│ Require │ │ Orch │ │ Phases │ │ Valid │
│ Analysis│ │ Design │ │ Design │ │ & Integ │
└─────────┘ └─────────┘ └─────────┘ └─────────┘
↓ ↓ ↓ ↓
workflow SKILL.md phases/ Complete
config generated 0N-*.md skill pkg
Target Output Structure
The skill this meta-skill produces follows this structure:
.claude/skills/{skill-name}/
├── SKILL.md # Orchestrator: coordination, data flow, TodoWrite
├── phases/
│ ├── 01-{phase-name}.md # Phase execution detail (full content)
│ ├── 02-{phase-name}.md
│ ├── ...
│ └── 0N-{phase-name}.md
├── specs/ # [Optional] Domain specifications
└── templates/ # [Optional] Reusable templates
Core Design Patterns
Patterns extracted from successful workflow skill implementations (workflow-plan, project-analyze, etc.):
Pattern 1: Orchestrator + Progressive Loading
SKILL.md = Pure coordinator. Contains:
- Architecture diagram (ASCII)
- Execution flow with
Ref: phases/0N-xxx.mdmarkers - Phase Reference Documents table (read on-demand)
- Data flow between phases
- Core rules and error handling
Phase files = Full execution detail. Contains:
- Complete agent prompts, bash commands, code implementations
- Validation checklists, error handling
- Input/Output specification
- Next Phase link
Key Rule: SKILL.md references phase docs via Ref: markers. Phase docs are read only when that phase executes, not all at once.
Pattern 2: TodoWrite Attachment/Collapse
Phase starts:
→ Sub-tasks ATTACHED to TodoWrite (in_progress + pending)
→ Orchestrator executes sub-tasks sequentially
Phase ends:
→ Sub-tasks COLLAPSED back to high-level summary (completed)
→ Next phase begins
Pattern 3: Inter-Phase Data Flow
Phase N output → stored in memory/variable → Phase N+1 input
└─ or written to session file for persistence
Each phase receives outputs from prior phases via:
- In-memory variables (sessionId, contextPath, etc.)
- Session directory files (.workflow/active/{sessionId}/...)
- Planning notes (accumulated constraints document)
Pattern 4: Conditional Phase Execution
Phase N output contains condition flag
├─ condition met → Execute Phase N+1
└─ condition not met → Skip to Phase N+2
Pattern 5: Input Structuring
User input (free text) → Structured format before Phase 1:
GOAL: [objective]
SCOPE: [boundaries]
CONTEXT: [background/constraints]
Pattern 6: Interactive Preference Collection (SKILL.md Responsibility)
Workflow preferences (auto mode, force explore, etc.) MUST be collected via AskUserQuestion in SKILL.md before dispatching to phases. Phases reference these as workflowPreferences.{key} context variables.
Anti-Pattern: Command-line flags (--yes, -e, --explore) parsed within phase files via $ARGUMENTS.includes(...).
// CORRECT: In SKILL.md (before phase dispatch)
const prefResponse = AskUserQuestion({
questions: [
{ question: "是否跳过确认?", header: "Auto Mode", options: [
{ label: "Interactive (Recommended)", description: "交互模式" },
{ label: "Auto", description: "跳过所有确认" }
]}
]
})
workflowPreferences = { autoYes: prefResponse.autoMode === 'Auto' }
// CORRECT: In phase files (reference only)
const autoYes = workflowPreferences.autoYes
// WRONG: In phase files (flag parsing)
const autoYes = $ARGUMENTS.includes('--yes') || $ARGUMENTS.includes('-y')
Pattern 7: Direct Phase Handoff
When one phase needs to invoke another phase within the same skill, read and execute the phase document directly. Do NOT use Skill() routing back through SKILL.md.
// CORRECT: Direct handoff (executionContext already set)
Read("phases/02-lite-execute.md")
// Execute with executionContext (Mode 1)
// WRONG: Skill routing (unnecessary round-trip)
Skill(skill="workflow-lite-plan", args="--in-memory")
Pattern 8: Phase File Hygiene
Phase files are internal execution documents. They MUST NOT contain:
| Prohibited | Reason | Correct Location |
|---|---|---|
Flag parsing ($ARGUMENTS.includes(...)) |
Preferences collected in SKILL.md | SKILL.md via AskUserQuestion |
Invocation syntax (/skill-name "...") |
Not user-facing docs | Removed or SKILL.md only |
Conversion provenance (Source: Converted from...) |
Implementation detail | Removed |
Skill routing for inter-phase (Skill(skill="...")) |
Use direct phase read | Direct Read("phases/...") |
Pattern 9: Compact Recovery (Phase Persistence)
Multi-phase workflows span long conversations. Context compression (compact) will naturally summarize earlier phase documents. The strategy uses 双重保险: TodoWrite 跟踪 active phase 保护其不被压缩,sentinel 作为兜底在压缩发生时触发恢复。
Design principle: TodoWrite in_progress = active phase → protect from compact | Sentinel = re-read fallback if protection fails.
Double insurance mechanism:
| Layer | Role | Mechanism |
|---|---|---|
| 1. Active Phase Protection | 预防 | TodoWrite in_progress 标记当前 phase → compact 时保留该 phase 完整内容,不压缩 |
| 2. Sentinel Re-read | 恢复 | Phase 文件嵌入标识符 → 若仍被压缩,检测到 sentinel-only 状态时立即 re-read |
When to apply: 任何通过 direct handoff (Pattern 7) 跨 phase 执行的场景,尤其是后续 phase 包含复杂执行协议(多 Step、agent 调度、CLI 编排)时。
Layer 1: Active Phase Protection (TodoWrite 联动)
TodoWrite 已经跟踪每个 phase 的执行状态。利用此信息驱动 compact 策略:
TodoWrite status → Compact behavior:
┌─ completed → ✅ 可压缩(已完成,不再需要完整内容)
├─ in_progress → 🛡️ 禁止压缩(正在执行,必须保留完整协议)
└─ pending → ✅ 可压缩(尚未加载,无内容)
SKILL.md Execution Flow 中标注 compact 联动:
## Execution Flow
> **⚠️ COMPACT DIRECTIVE**: Context compression MUST check TodoWrite phase status.
> The phase currently marked `in_progress` is the active execution phase — preserve its FULL content.
> Only compress phases marked `completed` or `pending`.
Phase 1: Requirements Analysis ← TodoWrite tracks status
└─ Ref: phases/01-xxx.md
Phase 2: Execution Engine ← TodoWrite tracks status
└─ Ref: phases/02-xxx.md
...
TodoWrite 状态转换 时自动更新 compact 保护范围:
Phase 1: in_progress 🛡️ → completed ✅ (compact 可压缩 Phase 1)
Phase 2: pending ✅ → in_progress 🛡️ (compact 保护 Phase 2)
Layer 2: Sentinel Re-read (兜底恢复)
即使有 Layer 1 保护,compact 仍可能在极端场景(超长上下文、多轮 agent 调度)下压缩 active phase。Sentinel 确保恢复能力:
Phase 文件顶部嵌入 sentinel:
> **📌 COMPACT SENTINEL [Phase N: {phase-name}]**
> This phase contains {M} execution steps (Step N.1 — N.{M}).
> If you can read this sentinel but cannot find the full Step protocol below, context has been compressed.
> Recovery: `Read("phases/0N-xxx.md")`
Sentinel 设计特点:
- 结构化、醒目 → compact 会将其作为关键信息保留在摘要中
- 包含 step 数量 → 提供自检依据("应该看到 M 个 Step,但只有摘要")
- 包含 re-read 路径 → 无需查表即可恢复
Phase Reference Table (整合双重保险)
| Phase | Document | Purpose | Compact |
|-------|----------|---------|---------|
| 1 | phases/01-xxx.md | Planning | TodoWrite 驱动 |
| 2 | phases/02-xxx.md | Execution | TodoWrite 驱动 + 🔄 sentinel |
**Compact Rules**:
1. **TodoWrite `in_progress`** → 保留完整内容,禁止压缩
2. **TodoWrite `completed`** → 可压缩为摘要
3. **🔄 sentinel fallback** → 带此标记的 phase 包含 compact sentinel;若 compact 后仅存 sentinel 而无完整 Step 协议,**必须立即 `Read("phases/0N-xxx.md")` 恢复后再继续**
Checkpoint (执行步骤前双重验证)
> **⚠️ CHECKPOINT**: Before proceeding, verify:
> 1. This phase is TodoWrite `in_progress` (active phase protection)
> 2. Full protocol (Step N.X — N.{M}) is in active memory, not just sentinel
> If only sentinel remains → `Read("phases/0N-xxx.md")` now.
Handoff 注释
// Phase N is tracked by TodoWrite — active phase protection applies.
// Sentinel fallback: if compressed despite protection, re-read triggers automatically.
Read("phases/0N-xxx.md")
Execution Flow
Phase 1: Requirements Analysis
└─ Ref: phases/01-requirements-analysis.md
├─ Input source: commands, descriptions, user interaction
└─ Output: workflowConfig (phases, data flow, agents, conditions)
Phase 2: Orchestrator Design (SKILL.md)
└─ Ref: phases/02-orchestrator-design.md
├─ Input: workflowConfig
└─ Output: .claude/skills/{name}/SKILL.md
Phase 3: Phase Files Design
└─ Ref: phases/03-phase-design.md
├─ Input: workflowConfig + source content
└─ Output: .claude/skills/{name}/phases/0N-*.md
Phase 4: Validation & Integration
└─ Ref: phases/04-validation.md
└─ Output: Validated skill package
Phase Reference Documents (read on-demand):
| Phase | Document | Purpose |
|---|---|---|
| 1 | phases/01-requirements-analysis.md | Analyze workflow requirements from various sources |
| 2 | phases/02-orchestrator-design.md | Generate SKILL.md with orchestration patterns |
| 3 | phases/03-phase-design.md | Generate phase files preserving full execution detail |
| 4 | phases/04-validation.md | Validate structure, references, and integration |
Input Sources
This meta-skill accepts workflow definitions from multiple sources:
| Source | Description | Example |
|---|---|---|
| Existing commands | Convert .claude/commands/ orchestrator + sub-commands |
plan.md + session/start.md + tools/*.md |
| Text description | User describes workflow in natural language | "Create a 3-phase code review workflow" |
| Requirements doc | Structured requirements file | requirements.md with phases/agents/outputs |
| Existing skill | Refactor/redesign an existing skill | Restructure a flat skill into phases |
Frontmatter Conversion Rules
When converting from command format to skill format:
| Command Field | Skill Field | Transformation |
|---|---|---|
name |
name |
Prefix with group: plan → workflow-plan |
description |
description |
Append trigger phrase: Triggers on "xxx" |
argument-hint |
(removed) | Arguments handled in Input Processing section |
examples |
(removed) | Examples moved to inline documentation |
allowed-tools |
allowed-tools |
Expand wildcards: Skill(*) → Skill, add commonly needed tools |
group |
(removed) | Embedded in name prefix |
Orchestrator Content Mapping
What goes into SKILL.md vs what goes into phase files:
SKILL.md (Coordinator)
| Section | Content | Source |
|---|---|---|
| Frontmatter | name, description, allowed-tools | Command frontmatter (converted) |
| Architecture Overview | ASCII diagram of phase flow | Derived from execution structure |
| Key Design Principles | Coordination rules | Extracted from command coordinator role |
| Execution Flow | Phase sequence with Ref: markers + Phase Reference table |
Command execution process |
| Core Rules | Orchestration constraints | Command core rules |
| Input Processing | Structured format conversion | Command input processing |
| Data Flow | Inter-phase data passing | Command data flow |
| TodoWrite Pattern | Attachment/collapse lifecycle | Command TodoWrite sections |
| Post-Phase Updates | Planning notes / state updates between phases | Command inter-phase update code |
| Error Handling | Failure recovery | Command error handling |
| Coordinator Checklist | Pre/post phase actions | Command coordinator checklist |
| Related Commands | Prerequisites and follow-ups | Command related commands |
Phase Files (Execution Detail)
| Content | Rule |
|---|---|
| Full agent prompts | Preserve verbatim from source command |
| Bash command blocks | Preserve verbatim |
| Code implementations | Preserve verbatim |
| Validation checklists | Preserve verbatim |
| Error handling details | Preserve verbatim |
| Input/Output spec | Add if not present in source |
| Phase header | Add # Phase N: {Name} |
| Objective section | Add ## Objective with bullet points |
| Next Phase link | Add ## Next Phase with link to next |
Critical Rule: Phase files must be content-faithful to their source. Do NOT summarize, abbreviate, or simplify. The phase file IS the execution instruction - every bash command, every agent prompt, every validation step must be preserved.
SKILL.md Template
---
name: {skill-name}
description: {description}. Triggers on "{trigger1}", "{trigger2}".
allowed-tools: {tools}
---
# {Title}
{One-paragraph description of what this skill does and what it produces.}
## Architecture Overview
{ASCII diagram showing phases and data flow}
## Key Design Principles
1. **{Principle}**: {Description}
...
## Interactive Preference Collection
Collect workflow preferences via AskUserQuestion before dispatching to phases:
{AskUserQuestion code with preference derivation → workflowPreferences}
## Auto Mode Defaults
When `workflowPreferences.autoYes === true`: {auto-mode behavior}.
## Execution Flow
{Phase sequence with Ref: markers}
**Phase Reference Documents** (read on-demand when phase executes):
| Phase | Document | Purpose | Compact |
|-------|----------|---------|---------|
| 1 | [phases/01-xxx.md](phases/01-xxx.md) | ... | TodoWrite 驱动 |
| N | [phases/0N-xxx.md](phases/0N-xxx.md) | ... | TodoWrite 驱动 + 🔄 sentinel |
...
**Compact Rules**:
1. **TodoWrite `in_progress`** → 保留完整内容,禁止压缩
2. **TodoWrite `completed`** → 可压缩为摘要
3. **🔄 sentinel fallback** → 带此标记的 phase 包含 compact sentinel;若 compact 后仅存 sentinel 而无完整 Step 协议,必须立即 `Read()` 恢复
## Core Rules
1. {Rule}
...
## Input Processing
{How user input is converted to structured format}
## Data Flow
{Inter-phase data passing diagram}
## TodoWrite Pattern
{Attachment/collapse lifecycle description with examples}
## Post-Phase Updates
{State updates between phases}
## Error Handling
{Failure recovery rules}
## Coordinator Checklist
{Pre/post phase action list}
## Related Commands
{Prerequisites and follow-ups}
Phase File Template
# Phase N: {Phase Name}
> **📌 COMPACT SENTINEL [Phase N: {phase-name}]**
> This phase contains {M} execution steps (Step N.1 — N.{M}).
> If you can read this sentinel but cannot find the full Step protocol below, context has been compressed.
> Recovery: `Read("phases/0N-xxx.md")`
> _(Include for phases marked 🔄 in SKILL.md Phase Reference table — see Pattern 9)_
{One-sentence description of this phase's goal.}
## Objective
- {Goal 1}
- {Goal 2}
## Execution
### Step N.1: {Step Name}
{Full execution detail: commands, agent prompts, code}
### Step N.2: {Step Name}
> **⚠️ CHECKPOINT**: Before proceeding, verify:
> 1. This phase is TodoWrite `in_progress` (active phase protection)
> 2. Full protocol (Step N.X — N.{M}) is in active memory, not just sentinel
> If only sentinel remains → `Read("phases/0N-xxx.md")` now.
> _(Add checkpoints before critical execution steps: agent dispatch, CLI launch, review — see Pattern 9)_
{Full execution detail}
## Output
- **Variable**: `{variableName}` (e.g., `sessionId`)
- **File**: `{output file path}`
- **TodoWrite**: Mark Phase N completed, Phase N+1 in_progress
## Next Phase
Return to orchestrator, then auto-continue to [Phase N+1: xxx](0N+1-xxx.md).
Design Decision Framework
When designing a new workflow skill, answer these questions:
| Question | Impact | Example |
|---|---|---|
| How many phases? | Directory structure | 3-7 phases typical |
| Which phases are conditional? | Orchestrator logic | "Phase 3 only if conflict_risk >= medium" |
| What data flows between phases? | Data Flow section | sessionId, contextPath, configFlags |
| Which phases use agents? | Phase file complexity | Agent prompts need verbatim preservation |
| What's the TodoWrite granularity? | TodoWrite Pattern | Some phases have sub-tasks, others are atomic |
| Is there a planning notes pattern? | Post-Phase Updates | Accumulated state document across phases |
| What's the error recovery? | Error Handling | Retry once then report, vs rollback |
| Does it need preference collection? | Interactive Preference Collection | Collect via AskUserQuestion in SKILL.md, pass as workflowPreferences |
| Does phase N hand off to phase M? | Direct Phase Handoff (Pattern 7) | Read phase doc directly, not Skill() routing |
| Will later phases run after long context? | Compact Recovery (Pattern 9) | Add sentinel + checkpoints, mark 🔄 in Phase Reference table |