tools-only Research Agent Assessment — Phase 1 Plugin Creation
Assessed: 2026-02-26
Scope: All agents in plugins/plugin-creator/agents/
Purpose: Identify best agent for Phase 1 research tasks in plugin creation workflows
Research Tasks Being Evaluated
- Search existing plugins/skills for similar functionality (code discovery)
- Gather domain knowledge about plugin structure, conventions, frontmatter rules
- Identify architecture patterns from well-structured plugins
Agent Inventory
1. plugin-assessor
File: plugins/plugin-creator/agents/plugin-assessor.md
Skills loaded:
claude-skills-overview-2026claude-plugins-reference-2026claude-hooks-reference-2026
Domain knowledge provided by skills:
claude-skills-overview-2026— Complete reference for Claude Code skills system: frontmatter schema, YAML syntax rules, multiline indicator prohibitions, tool/allowed-tools field format, progressive disclosure patterns,name:field requirements per agentskills.io specclaude-plugins-reference-2026— Complete reference for plugin system: plugin.json schema, installation scopes, component path fields, path behavior rules, environment variables, LSP server integration, CLI commandsclaude-hooks-reference-2026— Complete reference for hooks system: all hook events, configuration fields, exit codes, matcher syntax
Model: sonnet
Research task fitness:
| Task | Fitness | Evidence |
|---|---|---|
| Search existing plugins/skills for similar functionality | HIGH — Phase 1 "Discovery" workflow explicitly Globs for all capability files | Lines 32-44: Verify plugin.json, Glob for skills/*/SKILL.md, commands/*.md, agents/*.md |
| Gather domain knowledge about plugin structure, conventions, frontmatter rules | HIGH — three reference skills cover all schema, YAML rules, and conventions | Loads 3 reference skills; Rules 11-12 require consulting loaded skills before making technical claims |
| Identify architecture patterns from well-structured plugins | HIGH — Phase 3 Reference File Audit + Phase 8 Cross-Reference Analysis build structural maps | Lines 100-158 for reference classification; lines 280-310 for cross-reference graph |
Assessment: plugin-assessor is the strongest candidate. It performs structural discovery as its primary workflow, carries all three authoritative domain-knowledge skills, and has explicit rules requiring it to cite those skills rather than guess.
2. refactor-planner
File: plugins/plugin-creator/agents/refactor-planner.md
Skills loaded: none (no skills field in frontmatter)
Domain knowledge provided by skills: None — no skills field. All domain knowledge must come from the model's training data.
Model: sonnet
Research task fitness:
| Task | Fitness | Evidence |
|---|---|---|
| Search existing plugins/skills for similar functionality | MEDIUM — Discovery Phase reads plugin.json and Globs component directories | Lines 21-26 perform file discovery |
| Gather domain knowledge about plugin structure, conventions, frontmatter rules | LOW — no reference skills loaded; relies on training data | No skills field; no authoritative source for schema claims |
| Identify architecture patterns from well-structured plugins | MEDIUM — Skill Analysis phase reads SKILL.md files and evaluates domains | Lines 27-35 evaluate each skill for domains and split candidates |
Assessment: Useful for planning after discovery is complete, but carries no authoritative schema references. Would need to guess at conventions rather than cite them.
3. subagent-refactorer
File: plugins/plugin-creator/agents/subagent-refactorer.md
Skills loaded:
write-frontmatter-description
Domain knowledge provided by skills:
write-frontmatter-description— Formatting rules for frontmatter description fields: single-line only, no colons, front-load critical info, 1024-char limit
Model: sonnet
Tools: Extensive — includes WebFetch, WebSearch, MCP Ref tools, context7, exa, github, sequential-thinking, episodic-memory
Research task fitness:
| Task | Fitness | Evidence |
|---|---|---|
| Search existing plugins/skills for similar functionality | LOW — designed for prompt engineering research, not codebase discovery | Phase 1 focuses on Anthropic documentation URLs, not local filesystem search |
| Gather domain knowledge about plugin structure, conventions, frontmatter rules | MEDIUM — can research Anthropic docs via MCP tools but lacks the three plugin reference skills | Phase 1 lists specific Anthropic doc URLs to consult |
| Identify architecture patterns from well-structured plugins | LOW — targets agent refactoring patterns, not plugin structural patterns | Focus is prompt engineering methodology, not plugin architecture |
Assessment: Wrong specialization for Phase 1 plugin research. Strong for improving agent prompts; weak for discovering local plugin conventions.
4. agent-creator
File: plugins/plugin-creator/agents/agent-creator.md
Skills loaded:
plugin-creator:claude-plugins-reference-2026plugin-creator:claude-hooks-reference-2026plugin-creator:claude-skills-overview-2026plugin-creator:agent-creator
Domain knowledge provided by skills:
- Three authoritative references identical to plugin-assessor (skills, plugins, hooks)
- Plus
agent-creatorskill (creation workflow guidance)
Model: sonnet
Tools: Read, Write, Edit, Grep, Glob, Bash, Task
Research task fitness:
| Task | Fitness | Evidence |
|---|---|---|
| Search existing plugins/skills for similar functionality | HIGH — Phase 1 Discovery explicitly Globs agents/*.md across project and plugin directories |
Lines 33-36: Glob("agents/*.md", ".claude/") and Glob("plugins/*/agents/*.md") |
| Gather domain knowledge about plugin structure, conventions, frontmatter rules | HIGH — three reference skills provide complete authoritative coverage | Frontmatter Constraints section cites exact schema requirements backed by loaded skills |
| Identify architecture patterns from well-structured plugins | HIGH — discovery reads existing agents and identifies patterns for template selection | Phase 3 Template Selection reads existing agents as templates |
Assessment: Strong research candidate. Carries the same three reference skills as plugin-assessor. Discovery workflow is narrower (agents only vs. all components), but it explicitly reads existing files to identify patterns.
5. refactor-validator
File: plugins/plugin-creator/agents/refactor-validator.md
Skills loaded: none (no skills field in frontmatter)
Domain knowledge provided by skills: None.
Model: sonnet
Research task fitness:
| Task | Fitness | Evidence |
|---|---|---|
| Search existing plugins/skills for similar functionality | LOW — validates completed refactoring, does not discover existing work | Workflow is task completion check + structure validation, not discovery |
| Gather domain knowledge about plugin structure, conventions, frontmatter rules | LOW — no reference skills; validates against hardcoded criteria | Lines 60-86 list criteria but without authoritative skill backing |
| Identify architecture patterns from well-structured plugins | LOW — checks for regressions, does not identify patterns | Phase 4 regression detection, not pattern research |
Assessment: Post-implementation validator. Not suited for Phase 1 research.
6. hook-creator
File: plugins/plugin-creator/agents/hook-creator.md
Skills loaded:
plugin-creator:claude-hooks-reference-2026plugin-creator:hooks-core-referenceplugin-creator:hooks-io-apiplugin-creator:hooks-patternsplugin-creator:hook-creator
Domain knowledge provided by skills:
claude-hooks-reference-2026— Complete hooks referencehooks-core-reference— Core hooks documentationhooks-io-api— Hooks I/O API spechooks-patterns— Reusable hook patternshook-creator— Creation workflow
Model: sonnet
Research task fitness:
| Task | Fitness | Evidence |
|---|---|---|
| Search existing plugins/skills for similar functionality | LOW — domain is hook scripts, not general plugin discovery | Workflow reads hooks.json and hook scripts, not skills/agents |
| Gather domain knowledge about plugin structure, conventions, frontmatter rules | LOW — authoritative only on hook domain; no plugin/skill reference skills | No claude-skills-overview or claude-plugins-reference loaded |
| Identify architecture patterns from well-structured plugins | LOW — hook engineering specialist, not plugin architecture analyst | Pattern library covers hook patterns only |
Assessment: Excellent for hook work; wrong specialization for Phase 1 plugin research.
7. refactor-executor
File: plugins/plugin-creator/agents/refactor-executor.md
Skills loaded: none (no skills field in frontmatter)
Domain knowledge provided by skills: None.
Model: sonnet
Research task fitness:
All three tasks: LOW — this is an execution orchestrator that runs pre-planned tasks. It does not perform discovery, gather domain knowledge, or identify patterns.
Assessment: Implementation agent; does not serve research purposes.
8. contextual-ai-documentation-optimizer
File: plugins/plugin-creator/agents/contextual-ai-documentation-optimizer.md
Skills loaded:
prompt-optimization-claude-45write-frontmatter-descriptionsubagent-contractplugin-creator:audit-skill-completeness
Domain knowledge provided by skills:
prompt-optimization-claude-45— Prompt optimization principles for Claude modelswrite-frontmatter-description— Frontmatter description formatting rulessubagent-contract— Subagent behavioral disciplineaudit-skill-completeness— 8-category skill quality evaluation
Model: sonnet
Research task fitness:
| Task | Fitness | Evidence |
|---|---|---|
| Search existing plugins/skills for similar functionality | LOW — designed for documentation optimization, not codebase discovery | Workflow is RT-ICA → analyze → optimize → verify |
| Gather domain knowledge about plugin structure, conventions, frontmatter rules | MEDIUM — knows frontmatter formatting rules and skill completeness criteria, but lacks the three full reference skills | Has write-frontmatter-description and audit-skill-completeness; lacks claude-skills-overview-2026 and claude-plugins-reference-2026 |
| Identify architecture patterns from well-structured plugins | LOW — evaluates documentation quality of a given file, does not survey the plugin landscape | File-type-specific strategies are for optimizing content, not discovering patterns |
Assessment: Excellent for optimizing documentation after research is complete; not suited for research itself.
Comparison Matrix
| Agent | Skills Covering Schema | Discovery Workflow | Pattern Identification | Authoritative Claims Rule |
|---|---|---|---|---|
| plugin-assessor | 3/3 (skills, plugins, hooks) | Full multi-component Glob | Phase 8 Cross-Reference Analysis | Explicit (Rules 11-13) |
| agent-creator | 3/3 (skills, plugins, hooks) | Agent-focused Glob | Template pattern reading | Implicit via loaded skills |
| refactor-planner | 0/3 | Reads plugin.json + Globs | Skill domain analysis | None |
| subagent-refactorer | 0/3 (description rules only) | Anthropic docs only | None | Research-first mandate |
| contextual-ai-documentation-optimizer | 0/3 (formatting + completeness only) | None | None | CoVe post-check |
| hook-creator | 1/3 (hooks only) | Hook scripts only | Hook patterns only | None |
| refactor-validator | 0/3 | None | None | None |
| refactor-executor | 0/3 | None | None | None |
Recommendation
Primary recommendation: plugin-creator:plugin-assessor
Rationale:
- Loads all three authoritative reference skills (
claude-skills-overview-2026,claude-plugins-reference-2026,claude-hooks-reference-2026) providing complete schema, convention, and frontmatter knowledge - Phase 1 Discovery workflow explicitly designed for comprehensive filesystem scanning across all component types (skills, commands, agents, config files)
- Phase 8 Cross-Reference Analysis builds a structural link graph — directly addresses "identify architecture patterns"
- Rules 11-13 explicitly require consulting loaded skills before making claims and citing authoritative sources — reduces hallucination risk on schema questions
- Report output documents findings with file:line citations, making research artifacts usable in subsequent phases
Secondary option: plugin-creator:agent-creator
Use agent-creator when Phase 1 research scope is limited to agents specifically (not full plugin components). It carries the same three reference skills and explicitly reads existing agents to identify templates and patterns.
Do not use for Phase 1:
refactor-planner— no reference skills; guesses at conventionssubagent-refactorer— wrong domain (prompt engineering research, not local discovery)contextual-ai-documentation-optimizer— optimization specialist, not researcherhook-creator— hook-domain onlyrefactor-validator— post-implementation validatorrefactor-executor— execution orchestrator
Delegation Format for SKILL.md Workflows
When referencing Phase 1 research in a plugin-creation SKILL.md workflow:
1. Task is domain research and code discovery with subagent_type="plugin-creator:plugin-assessor"
Context to include in the prompt: target plugin directory path, list of component types to survey,
any specific conventions or patterns to look for
Output: .claude/audits/plugin-assessment-{plugin-slug}.md — structural assessment with
file:line references for all discovered patterns, conventions, and similar functionality
If agent-scope is limited to agents only:
1. Task is agent pattern discovery with subagent_type="plugin-creator:agent-creator"
Context to include in the prompt: plugin directory path, description of agent functionality needed
Output: summary of existing agent patterns and recommended template approach