<plugin_mode>$0</plugin_mode> <plugin_target>$1</plugin_target> <invocation_args>$ARGUMENTS</invocation_args>

When editing files in plugins/, .claude/, AGENTS.md, or CLAUDE.md — delegate to subagent_type="plugin-creator:contextual-ai-documentation-optimizer".

[!IMPORTANT] When provided a process map or Mermaid diagram, treat it as the authoritative procedure. Execute steps in the exact order shown, including branches, decision points, and stop conditions. A Mermaid process diagram is an executable instruction set. Follow it exactly as written: respect sequence, conditions, loops, parallel paths, and terminal states. Do not improvise, reorder, or skip steps. If any node is ambiguous or missing required detail, pause and ask a clarifying question before continuing. When interacting with a user, report before acting the interpreted path you will follow from the diagram, then execute.

Plugin Lifecycle Orchestration

Orchestrate plugin development through seven phases. This skill composes existing plugin-creator skills and agents — it does not re-implement their logic.

Arguments: <invocation_args/>

• new <concept> — Create a plugin from scratch. Enters at Phase 0 (RT-ICA Prerequisite Check).
• existing <plugin-path> — Improve an existing plugin. Enters at Phase 1 (Assess).

Domain Knowledge Prerequisites

Load these skills at session start before executing any phase. Full skill descriptions and what each provides: references/domain-knowledge-prerequisites.md.

Required — load at session start:

1. Skill(skill="plugin-creator:claude-plugins-reference-2026") — plugin.json schema, component types, environment variables, installation scopes, path rules
2. Skill(skill="plugin-creator:claude-skills-overview-2026") — SKILL.md format, all 14 frontmatter fields, YAML multiline bug, allowed-tools string format, context fork behavior

Required for phases involving hooks (Phase 4: Create, Phase 5: Debug):

3. Skill(skill="plugin-creator:hooks-guide") — 13 hook event types, exit codes, tool denial mechanisms, agent frontmatter fields

Recommended for component selection and plugin configuration decisions:

4. Skill(skill="plugin-creator:component-patterns") — component lifecycle, discovery and activation phases, decision framework for choosing commands vs skills vs agents vs hooks vs MCP servers
5. Skill(skill="plugin-creator:plugin-settings") — .local.md per-project configuration pattern, YAML frontmatter parsing from hooks, configuration-driven behavior

Workflow Overview

The following diagram is the authoritative procedure for plugin lifecycle routing. Execute steps in the exact order shown, including branches, decision points, and stop conditions.

flowchart TD
    Start(["/plugin-lifecycle <invocation_args/>"]) --> Q1{"First argument is?"}
    Q1 -->|"new — create from scratch"| RTICA["Phase 0 — RT-ICA Prerequisite Check"]
    Q1 -->|"existing — improve existing plugin"| Assess["Phase 1 — Assess"]

    %% New path: RT-ICA gate
    RTICA --> RTICAGate{"RT-ICA decision?"}
    RTICAGate -->|"BLOCKED — one or more conditions MISSING"| RTICABlock(["STOP — present missing inputs to user<br>Do not proceed until resolved"])
    RTICAGate -->|"APPROVED — all conditions available or derivable"| Discuss["Phase 0.5 — Discussion"]

    %% New path: Discussion gate — file must exist before Research
    Discuss --> DiscussGate{"File .claude/plan/NAME/discuss-CONTEXT.md<br>exists and is non-empty?"}
    DiscussGate -->|"Yes — preferences captured"| Mission["Phase 0.6 — Draft Mission Statement"]
    Mission --> Research["Phase 2 — Research"]
    DiscussGate -->|"No — file absent or empty"| Discuss

    %% Existing path: Assess then validator
    Assess --> AssessFile{"File .claude/plan/NAME/assessment-REPORT.md<br>exists and is non-empty?"}
    AssessFile -->|"No — assessor did not complete"| Assess
    AssessFile -->|"Yes — assessment written"| AssessGate{"Run: uvx skilllint@latest check PATH<br>Exit code?"}
    AssessGate -->|"0 — no validation errors"| Optimize["Phase 6 — Optimize"]
    AssessGate -->|"non-zero — errors found"| Debug["Phase 5 — Debug"]

    %% New path: Research gate
    Research --> ResearchGate{"File .claude/plan/NAME/research-FINDINGS.md<br>exists and is non-empty?"}
    ResearchGate -->|"Yes — all 4 researcher outputs merged"| Design["Phase 3 — Design"]
    ResearchGate -->|"No — merge incomplete or file absent"| Research

    %% New path: Design gate with iteration limit
    Design --> DesignGate{"design-PLAN.md exists<br>AND plan-checker returns PASS?"}
    DesignGate -->|"PASS — plan complete and verified"| Create["Phase 4 — Create"]
    DesignGate -->|"FAIL — iteration count < 3"| Design
    DesignGate -->|"FAIL — iteration count = 3 (limit reached)"| DesignEscalate(["STOP — escalate to user<br>Plan checker has failed 3 times"])

    %% New path: Create gate
    Create --> CreateGate{"All files listed in design-PLAN.md<br>exist at their specified paths?"}
    CreateGate -->|"Yes — all components created"| Debug
    CreateGate -->|"No — one or more files missing"| Create

    %% Shared Debug phase (both paths converge here)
    Debug --> DebugGate{"Run: uvx skilllint@latest check PATH<br>Exit code 0 AND 0 errors?<br>(warnings acceptable)"}
    DebugGate -->|"Yes — 0 errors, validation passes"| Optimize
    DebugGate -->|"No — errors remain"| Debug

    %% Optimize gate
    Optimize --> OptGate{"Run: uvx skilllint@latest check PATH<br>Output contains 'Score:' line?"}
    OptGate -->|"Score >= 80 — quality target met"| Docs["Phase 6.5 — Documentation"]
    OptGate -->|"Score < 80 — quality below target"| Optimize
    OptGate -->|"No score in output — user acceptance required"| OptUser{"User accepts current quality?"}
    OptUser -->|"Yes — user accepts"| Docs
    OptUser -->|"No — continue improving"| Optimize

    %% Documentation gate
    Docs --> DocsGate{"File {plugin-path}/README.md<br>exists and is non-empty?"}
    DocsGate -->|"Yes — documentation complete"| Verify["Phase 7 — Verify"]
    DocsGate -->|"No — README.md absent or empty"| Docs

    %% Verify: 4 discrete layers
    Verify --> VL1{"Layer 1 — Run: uvx skilllint@latest check PATH<br>Exit code 0?"}
    VL1 -->|"Yes — structural validation passes"| VL2{"Layer 2 — Run: claude plugin validate PATH<br>Exit code 0?"}
    VL1 -->|"No — structural errors found"| VerifyFail["Return to Phase 5 — Debug<br>with Layer 1 error details"]
    VL2 -->|"Yes — runtime validation passes"| VL3{"Layer 3 — skilllint output<br>contains SK006 or SK007 for any skill?"}
    VL2 -->|"No — runtime validation fails"| VerifyFail
    VL3 -->|"No SK006/SK007 — all skills within token limits"| VL4{"Layer 4 — all internal links resolve,<br>all plugin.json skill paths exist,<br>all agent references point to existing files?"}
    VL3 -->|"Yes — SK006 or SK007 present"| VerifyFail
    VL4 -->|"Yes — cross-reference integrity confirmed"| Done(["Write .claude/plan/NAME/SUMMARY.md<br>Plugin is marketplace-ready"])
    VL4 -->|"No — broken cross-references found"| VerifyFail
    VerifyFail --> Debug

Artifact System

All work artifacts are stored in .claude/plan/{plugin-name}/:

.claude/plan/{plugin-name}/
├── PROJECT.md                # Vision and goals
├── STATE.md                  # Current phase, decisions, blockers
├── discuss-CONTEXT.md        # Phase 0.5 output — user preferences (new path only)
├── research-FINDINGS.md      # Phase 2 output (new path only)
├── design-PLAN.md            # Phase 3 output (new path only)
├── assessment-REPORT.md      # Phase 1 output (existing path only)
├── validation-REPORT.md      # Phase 7 output
└── SUMMARY.md                # Completion record

{plugin-path}/mission.json — Phase 0.6 output — plugin mission statement with status: "draft" (new path); created by mission-statement skill at the plugin root (not inside .claude/plan/).

Before starting any phase, read STATE.md if it exists to determine current progress. After completing each phase, update STATE.md with the phase completed and any decisions made.

---

Phase 0: RT-ICA Prerequisite Check (New Plugin Only)

Entry condition: User provides new <concept>.

Before creating any plugin, verify all prerequisites are in place. Perform this RT-ICA assessment:

RT-ICA SUMMARY

Goal:
- Create a Claude Code plugin for [purpose]

Success Output:
- Functional plugin that [specific outcome]

Conditions (reverse prerequisites):
1. Purpose clarity     | Requires: Clear problem statement   | Why: Determines plugin scope
2. Target users        | Requires: Who will use this         | Why: Shapes UX decisions
3. Component selection | Requires: Skills vs Agents vs Hooks | Why: Architecture
4. Existing solutions  | Requires: Check for similar plugins | Why: Avoid duplication
5. Source material     | Requires: Documentation/APIs to encode | Why: Content accuracy
6. Verification method | Requires: How to test the plugin works | Why: Quality gate

Verification:
- [Check each condition: AVAILABLE / DERIVABLE / MISSING]

Decision:
- [APPROVED / BLOCKED]

The following diagram is the authoritative procedure for Phase 0 RT-ICA decision gate. Execute steps in the exact order shown, including branches, decision points, and stop conditions.

flowchart TD
    Q{"RT-ICA decision?"}
    Q -->|"APPROVED — all 6 conditions available or derivable"| Next["Proceed to Phase 0.5 — Discussion"]
    Q -->|"BLOCKED — one or more conditions MISSING"| Block(["STOP — present missing conditions to user<br>Do not proceed to Phase 0.5 or Phase 2<br>until all conditions are AVAILABLE or DERIVABLE"])

---

Phase 0.5: Discussion — Capture User Preferences (New Plugin Only)

Entry condition: RT-ICA gate returned APPROVED.

Before research, identify gray areas and capture user preferences to guide all subsequent phases.

Ask targeted questions to eliminate ambiguity:

For skill-focused plugins:

• Activation triggers: When should Claude auto-load vs user-invoke?
• Tool restrictions: Full access or limited tools?
• Output format: Verbose explanations or terse instructions?
• Reference structure: Inline content or progressive disclosure?

For agent-focused plugins:

• Delegation scope: What tasks should agents handle?
• Return format: Summaries or detailed reports?
• Error handling: Retry, escalate, or fail fast?

For hook-focused plugins:

• Trigger events: Which tool/session events matter?
• Hook type: Command, prompt, or agent verification?
• Timeout handling: Fail silently or block?

Save preferences to .claude/plan/{plugin-name}/discuss-CONTEXT.md:

# Plugin Discussion: {plugin-name}
Date: {ISO timestamp}

## Scope Decisions
- {question}: {user preference}

## UX Preferences
- Invocation: {user-invoked | model-invoked | both}
- Verbosity: {terse | balanced | verbose}

## Technical Choices
- {choice}: {preference with rationale}

These preferences guide all subsequent research and planning phases.

The following diagram is the authoritative procedure for Phase 0.5 discussion completion gate. Execute steps in the exact order shown, including branches, decision points, and stop conditions.

flowchart TD
    Q{"File .claude/plan/NAME/discuss-CONTEXT.md<br>exists and is non-empty?"}
    Q -->|"Yes — user preferences captured and written"| Next["Proceed to Phase 2 — Research"]
    Q -->|"No — file absent or empty"| Retry["Re-run Phase 0.5 discussion<br>Ask targeted questions again<br>Write preferences to discuss-CONTEXT.md"]
    Retry --> Q

---

Phase 1: Assess (Existing Plugin Only)

Entry condition: User provides existing <plugin-path>.

1. Task is plugin assessment with Skill(skill="plugin-creator:assessor") Context to include in the prompt: plugin directory path from <plugin_target/> Output: .claude/plan/{plugin-name}/assessment-REPORT.md — assessment report with design map and task file

The following diagram is the authoritative procedure for Phase 1 Assess decision gate. Execute steps in the exact order shown, including branches, decision points, and stop conditions.

flowchart TD
    %% Gate 1: assessor output must exist before validator can be meaningful
    AssessFile{"File .claude/plan/NAME/assessment-REPORT.md<br>exists and is non-empty?"}
    AssessFile -->|"No — assessor did not complete"| RetryAssess["Re-run assessor skill<br>with plugin directory path"]
    RetryAssess --> AssessFile
    AssessFile -->|"Yes — assessment written"| ValidatorGate{"Run: uvx skilllint@latest check PATH<br>Exit code?"}
    ValidatorGate -->|"0 — no validation errors"| Skip["Proceed to Phase 6 — Optimize"]
    ValidatorGate -->|"non-zero — errors found"| Next["Proceed to Phase 5 — Debug"]

---

Phase 0.6: Mission Statement Draft (New Plugin Only)

Entry condition: Discussion phase completed and discuss-CONTEXT.md written.

Before research begins, draft an initial mission statement for the plugin. This anchors all subsequent phases to the plugin's purpose and values and creates a backlog interview task for async human refinement.

1. Task is mission statement drafting with Skill(skill="plugin-creator:mission-statement") Context to include in the prompt: plugin concept from <plugin_target/>, path to discuss-CONTEXT.md Output: {plugin-path}/mission.json with status: "draft" — a GitHub backlog interview task is created automatically by the skill

The mission statement is never a blocker. Research and all subsequent phases proceed without waiting for the interview. The [draft] status on mission.json signals this is a hypothesis, not a decision.

The following diagram is the authoritative procedure for Phase 0.6 completion gate.

flowchart TD
    Q{"File {plugin-path}/mission.json<br>exists and status field is present?"}
    Q -->|"Yes — draft mission written"| Next["Proceed to Phase 2 — Research"]
    Q -->|"No — mission.json absent"| Retry["Re-run mission-statement skill<br>with plugin concept and discuss-CONTEXT.md path"]
    Retry --> Q

---

Phase 2: Research (New Plugin Only)

Entry condition: Discussion phase completed and discuss-CONTEXT.md written.

Spawn all four researchers in a single message to run concurrently. Merge results into research-FINDINGS.md before proceeding to Design.

1. Task is feature discovery with Skill(skill="plugin-creator:feature-discovery") Context to include in the prompt: plugin concept from <plugin_target/> (everything after "new"), discuss-CONTEXT.md Output: .claude/plan/{plugin-name}/feature-context-{slug}.md — feature context document

2. Task is existing solutions research with subagent_type="plugin-creator:plugin-assessor" Context to include in the prompt: plugin concept, feature context from step 1 Prompt for researcher: Search plugins/ and ~/.claude/skills/ for similar functionality. Report what exists, gaps to fill, patterns to follow or avoid. Output: .claude/plan/{plugin-name}/research-1-existing.md

3. Task is Claude Code features research with subagent_type="plugin-creator:plugin-assessor" Context to include in the prompt: plugin concept, feature context from step 1 Prompt for researcher: What capabilities should this plugin use — dynamic context injection (!command), subagent execution (context: fork), hooks (which events?), MCP/LSP integration opportunities? Report recommended features with rationale. Output: .claude/plan/{plugin-name}/research-2-features.md

4. Task is architecture patterns research with subagent_type="plugin-creator:plugin-assessor" Context to include in the prompt: plugin concept, feature context from step 1 Prompt for researcher: How do well-structured plugins organize — skill directory structure, reference file patterns, agent definitions, hook configurations? Report recommended structure based on similar plugins. Output: .claude/plan/{plugin-name}/research-3-architecture.md

5. Task is pitfalls and official docs research with subagent_type="general-purpose" Context to include in the prompt: plugin concept, feature context from step 1 Prompt for researcher: Fetch https://code.claude.com/docs/en/plugins-reference.md and https://code.claude.com/docs/en/skills.md. Identify schema requirements (comma-separated strings NOT arrays), common mistakes, deprecations or new features. Report gotchas to avoid. Output: .claude/plan/{plugin-name}/research-4-pitfalls.md

After all four researchers complete, consolidate into research-FINDINGS.md:

# Research Findings: {plugin-name}
Date: {ISO timestamp}

## 1. Existing Solutions
{Researcher 1 findings}

## 2. Recommended Features
{Researcher 2 findings}

## 3. Architecture Patterns
{Researcher 3 findings}

## 4. Pitfalls & Requirements
{Researcher 4 findings}

## Synthesis
- Key insights: {combined learnings}
- Recommended approach: {synthesis}

The following diagram is the authoritative procedure for Phase 2 Research decision gate. Execute steps in the exact order shown, including branches, decision points, and stop conditions.

flowchart TD
    %% All 4 individual research files must exist before merge is valid
    R1{"File .claude/plan/NAME/research-1-existing.md<br>exists and is non-empty?"}
    R1 -->|"No — researcher 1 (existing solutions) failed"| Retry1["Re-spawn researcher 1<br>with more specific prompt"]
    Retry1 --> R1
    R1 -->|"Yes"| R2{"File .claude/plan/NAME/research-2-features.md<br>exists and is non-empty?"}
    R2 -->|"No — researcher 2 (Claude Code features) failed"| Retry2["Re-spawn researcher 2<br>with more specific prompt"]
    Retry2 --> R2
    R2 -->|"Yes"| R3{"File .claude/plan/NAME/research-3-architecture.md<br>exists and is non-empty?"}
    R3 -->|"No — researcher 3 (architecture patterns) failed"| Retry3["Re-spawn researcher 3<br>with more specific prompt"]
    Retry3 --> R3
    R3 -->|"Yes"| R4{"File .claude/plan/NAME/research-4-pitfalls.md<br>exists and is non-empty?"}
    R4 -->|"No — researcher 4 (pitfalls/docs) failed"| Retry4["Re-spawn researcher 4<br>with more specific prompt"]
    Retry4 --> R4
    R4 -->|"Yes — all 4 researcher outputs exist"| Merge{"File .claude/plan/NAME/research-FINDINGS.md<br>exists and is non-empty?"}
    Merge -->|"Yes — merge complete"| Next["Proceed to Phase 3 — Design"]
    Merge -->|"No — merge not yet written"| DoMerge["Consolidate all 4 research files<br>into research-FINDINGS.md"]
    DoMerge --> Merge

---

Phase 3: Design (New Plugin Only)

Entry condition: Research gate passed.

1. Task is prerequisite check with Skill(skill="plugin-creator:rt-ica") Context to include in the prompt: research-FINDINGS.md, plugin concept, user requirements from discuss-CONTEXT.md Output: APPROVED or BLOCKED verdict — if BLOCKED, resolve blockers before proceeding

2. Task is design plan creation with subagent_type="general-purpose" Context to include in the prompt: research-FINDINGS.md, rt-ica output, discuss-CONTEXT.md Output: .claude/plan/{plugin-name}/design-PLAN.md — design plan with XML task specs defining every skill, agent, and hook to create. Each task must have: single responsibility, testable <verify> command, clear <done> criteria.

3. Task is plan verification with subagent_type="general-purpose" Context to include in the prompt: design-PLAN.md, discuss-CONTEXT.md, research-FINDINGS.md key sections Prompt: Verify this plan achieves the plugin goals. Check: (1) do tasks cover all required components? (2) are tasks truly atomic? (3) are <verify> commands testable? (4) are there gaps between tasks? (5) does sequence respect dependencies? Return PASS or FAIL with specific issues. Output: PASS verdict (proceed) or FAIL with feedback (return to step 2)

The following diagram is the authoritative procedure for Phase 3 Design decision gate. Execute steps in the exact order shown, including branches, decision points, and stop conditions.

flowchart TD
    %% Track iteration count to enforce the 3-attempt limit from Error Handling
    IterCheck{"Current plan-checker iteration count<br>(track in STATE.md — how many FAIL verdicts so far)?"}
    IterCheck -->|"Count = 3 — limit reached"| Escalate(["STOP — escalate to user<br>Plan checker has returned FAIL 3 times<br>Present FAIL feedback and await direction"])
    IterCheck -->|"Count < 3 — iterations remain"| PlanCheck{"design-PLAN.md exists<br>AND plan-checker returns PASS?"}
    PlanCheck -->|"PASS — plan complete and verified"| Next["Proceed to Phase 4 — Create"]
    PlanCheck -->|"FAIL — plan incomplete or unverifiable"| Revise["Increment iteration count in STATE.md<br>Pass design-PLAN.md and FAIL feedback<br>to planner for revision"]
    Revise --> IterCheck

---

Phase 4: Create (New Plugin Only)

Entry condition: Design gate passed.

For each component defined in design-PLAN.md, invoke the appropriate creator skill:

1. Task is skill creation with Skill(skill="plugin-creator:skill-creator") Context to include in the prompt: design-PLAN.md task spec for this skill, plugin path Output: {plugin-path}/skills/{skill-name}/SKILL.md and any bundled resources

2. Task is agent creation with Skill(skill="plugin-creator:agent-creator") Context to include in the prompt: design-PLAN.md task spec for this agent, plugin path Output: {plugin-path}/agents/{agent-name}.md

3. Task is hook creation with Skill(skill="plugin-creator:hook-creator") Context to include in the prompt: design-PLAN.md task spec for this hook, plugin path Output: hook scripts and hooks.json configuration

Repeat for each planned component. Create plugin.json via uv run plugins/plugin-creator/scripts/create_plugin.py if it does not exist.

The following diagram is the authoritative procedure for Phase 4 Create decision gate. Execute steps in the exact order shown, including branches, decision points, and stop conditions.

flowchart TD
    Q{"All file paths listed in design-PLAN.md<br>exist on disk at their specified paths?"}
    Q -->|"Yes — all components created"| Next["Proceed to Phase 5 — Debug"]
    Q -->|"No — one or more planned files are absent"| Retry["Identify which planned files are missing<br>Create remaining components using the appropriate creator skill<br>Return to gate check"]
    Retry --> Q

---

Phase 5: Debug (Both Paths)

Entry condition: Create gate passed (new path) OR Assess gate failed (existing path).

Debug fixes validation errors. Run the validator first to identify issues:

uvx skilllint@latest check <plugin-path>

The following diagram is the authoritative procedure for Phase 5 Debug error routing and completion gate. Execute steps in the exact order shown, including branches, decision points, and stop conditions.

flowchart TD
    %% Entry: run validator to get current error list
    RunValidator["Run: uvx skilllint@latest check PATH<br>Capture full output"] --> HasErrors{"Exit code 0<br>AND 0 errors in output?<br>(warnings are acceptable)"}
    HasErrors -->|"Yes — 0 errors, validation passes"| Next["Proceed to Phase 6 — Optimize"]
    HasErrors -->|"No — errors remain"| Q{"Error type in validator output?"}

    %% Route each error type to its fix, then loop back to re-validate
    Q -->|"SK007 — skill exceeds token limit (hard error)"| Split["Invoke: Skill(skill='plugin-creator:refactor-skill')<br>Context = oversized SKILL.md path<br>Output = split skill files at same plugin path"]
    Q -->|"SK006 — skill approaching token limit (warning)"| Extract["Extract content to references/ directory<br>Update SKILL.md to reference extracted files<br>Output = reduced SKILL.md + new references/ file"]
    Q -->|"Broken link error (LINK01 or similar)"| Links["Read the file containing the broken link<br>Verify the target path exists on disk<br>Fix with Edit tool — update or remove the broken reference"]
    Q -->|"Frontmatter issues (FM-series errors)"| Lint["Invoke: Skill(skill='plugin-creator:lint', args='--fix PATH')<br>Context = file path + validator output<br>Output = corrected frontmatter in the file"]
    Q -->|"Tool format issues (array instead of string)"| Tools["Invoke: Skill(skill='plugin-creator:lint', args='--fix PATH')<br>Output = fixed comma-separated string in frontmatter"]
    Q -->|"Other structural errors"| Manual["Read the validator error message<br>Identify the file and line referenced<br>Apply Edit fix directly to that file<br>Verify fix is consistent with plugin schema"]

    %% All fix paths loop back to re-validate
    Split --> RunValidator
    Extract --> RunValidator
    Links --> RunValidator
    Lint --> RunValidator
    Tools --> RunValidator
    Manual --> RunValidator

---

Phase 6: Optimize (Both Paths)

Entry condition: Debug gate passed OR Assess gate passed with no errors.

Optimize improves quality — descriptions, progressive disclosure, agent prompts, documentation. This phase is not about fixing errors (that is Debug) but about raising quality.

1. Task is structural plugin improvement with Skill(skill="plugin-creator:refactor-plugin") Context to include in the prompt: plugin path, assessment-REPORT.md (if available from Phase 1) Output: improved plugin structure, updated SKILL.md files, better progressive disclosure

2. Task is content quality optimization with subagent_type="plugin-creator:contextual-ai-documentation-optimizer" Context to include in the prompt: SKILL.md or CLAUDE.md files needing improvement, assessment findings Output: optimized documentation with better Claude comprehension

3. Task is agent prompt optimization with subagent_type="plugin-creator:subagent-refactorer" Context to include in the prompt: agent .md files needing improvement Output: optimized agent prompts using Anthropic best practices

The following diagram is the authoritative procedure for Phase 6 Optimize completion gate. Execute steps in the exact order shown, including branches, decision points, and stop conditions.

flowchart TD
    %% Score line presence determines which branch to take
    ScoreCheck{"Run: uvx skilllint@latest check PATH<br>Does output contain a 'Score:' line?"}
    ScoreCheck -->|"Yes — score present in output"| ScoreVal{"Score value >= 80?"}
    ScoreCheck -->|"No — validator produces no score"| UserAccept{"Ask user — accept current quality<br>and proceed to documentation?"}

    ScoreVal -->|"Yes — score >= 80, quality target met"| Next["Proceed to Phase 6.5 — Documentation"]
    ScoreVal -->|"No — score < 80, quality below target"| Retry["Identify lowest-scoring components<br>Re-run optimization steps targeting those components"]
    Retry --> ScoreCheck

    UserAccept -->|"Yes — user accepts current quality"| Next
    UserAccept -->|"No — user wants more improvement"| Retry

---

Phase 6.5: Documentation (Both Paths)

Entry condition: Optimize phase complete.

Generate comprehensive documentation for the plugin:

1. Task is plugin documentation generation with subagent_type="plugin-creator:plugin-assessor" Context to include in the prompt: plugin path, all SKILL.md files, agent files, plugin.json, assess-REPORT.md or design-PLAN.md (whichever is available) Prompt: Generate comprehensive documentation. Create: README.md with installation, usage, and examples; docs/skills.md if multiple skills exist; configuration guide if hooks or MCP servers are included. Ensure all features are documented, installation instructions are accurate, and examples are runnable. Output: {plugin-path}/README.md and any additional documentation files

The following diagram is the authoritative procedure for Phase 6.5 Documentation completion gate. Execute steps in the exact order shown, including branches, decision points, and stop conditions.

flowchart TD
    Q{"File {plugin-path}/README.md<br>exists and is non-empty?"}
    Q -->|"Yes — documentation generated"| Next["Proceed to Phase 7 — Verify"]
    Q -->|"No — README.md absent or empty"| Retry["Re-run documentation task<br>with explicit instruction to create README.md<br>at {plugin-path}/README.md"]
    Retry --> Q

---

Phase 7: Verify (Both Paths)

Entry condition: Documentation phase complete.

Run multi-layer validation:

1. Task is recursive validation with Skill(skill="plugin-creator:ensure-complete") Context to include in the prompt: plugin path, task file (if applicable) Output: .claude/plan/{plugin-name}/validation-REPORT.md

2. Layer 1 — Structural validation:

   uvx skilllint@latest check <plugin-path>
   

3. Layer 2 — Runtime validation:

   claude plugin validate <plugin-path>
   

4. Layer 3 — Token complexity: Check skilllint output for SK006/SK007 warnings on all skills.

5. Layer 4 — Cross-reference integrity: Verify all internal links resolve, all skills referenced in plugin.json exist, all agent references in skills point to existing agent files.

The following diagram is the authoritative procedure for Phase 7 Verify 4-layer validation gate. Execute steps in the exact order shown, including branches, decision points, and stop conditions.

flowchart TD
    %% Layer 1: structural validator
    VL1{"Layer 1 — Run: uvx skilllint@latest check PATH<br>Exit code 0 AND 0 errors in output?"}
    VL1 -->|"Yes — structural validation passes"| VL2{"Layer 2 — Run: claude plugin validate PATH<br>Exit code 0?"}
    VL1 -->|"No — structural errors found"| Fail1["Capture Layer 1 error details<br>Proceed to Phase 5 — Debug with these errors"]

    %% Layer 2: runtime validator
    VL2 -->|"Yes — runtime validation passes"| VL3{"Layer 3 — Does skilllint output<br>contain SK006 or SK007 for any skill?"}
    VL2 -->|"No — runtime validation fails"| Fail2["Capture Layer 2 error details<br>Check .claude-plugin/plugin.json exists<br>Check all paths start with ./<br>Proceed to Phase 5 — Debug with these errors"]

    %% Layer 3: token complexity
    VL3 -->|"No SK006/SK007 — all skills within token limits"| VL4{"Layer 4 — For every internal link in all SKILL.md and agent files:<br>does the target file exist on disk?<br>For every skill in plugin.json: does the SKILL.md exist?<br>For every agent reference in skills: does the agent .md exist?"}
    VL3 -->|"Yes — SK006 or SK007 present"| Fail3["Identify which skills triggered SK006/SK007<br>Proceed to Phase 5 — Debug targeting those skills"]

    %% Layer 4: cross-reference integrity — attempt inline fix before returning to Debug
    VL4 -->|"Yes — all cross-references resolve"| Done(["Write .claude/plan/NAME/SUMMARY.md<br>Plugin is marketplace-ready"])
    VL4 -->|"No — one or more broken cross-references"| Fail4["List each broken reference with file path and line<br>Fix with Edit tool directly<br>Re-run Layer 4 check"]
    Fail4 --> VL4

    %% Layers 1-3 failures route to Phase 5
    Fail1 --> DebugReturn["Proceed to Phase 5 — Debug"]
    Fail2 --> DebugReturn
    Fail3 --> DebugReturn

---

Phase-to-Skill Mapping

Full lookup table with exact invocation syntax for all 18 phase-skill pairings: references/phase-skill-mapping.md.

Key invocations:

• Phase 1: Skill(skill="plugin-creator:assessor")
• Phase 2: Skill(skill="plugin-creator:feature-discovery") + 4-way parallel researchers via subagent_type
• Phase 4: skill-creator, agent-creator, hook-creator (one Skill call per component type)
• Phase 5: lint, refactor-skill (one Skill call per error type)
• Phase 7: Skill(skill="plugin-creator:ensure-complete")

---

Error Handling

14 failure modes with recovery actions: references/error-handling.md.

Key rules:

• SK007 (token limit exceeded) — run /plugin-creator:refactor-skill; editing alone is not sufficient
• SK006 (approaching limit) — extract content to references/ and re-validate
• RT-ICA BLOCKED — do not proceed to Discussion or Research until all conditions resolve
• STATE.md absent — read all .claude/plan/{plugin-name}/ artifacts to reconstruct phase

---

Example Sessions

Two complete walkthroughs (new plugin full lifecycle + existing plugin with validation errors): references/example-sessions.md.

---

Sources

• Architecture spec: ../../../../plan/architect-plugin-lifecycle.md
• Feature context: ../../../../plan/feature-context-plugin-lifecycle.md
• Plugin-creator CLAUDE.md: ../../CLAUDE.md
• GitHub Issue: #427