Assimilate Popular Workflows

Assimilate Popular Workflows

Search public GitHub repositories for SKILL.md files, classify each repo by archetype, and maintain structured research documents under docs/reference-repos/[org]/[repo-name]/. The goal is not to copy skills verbatim but to extract transferable value: processes for the babysitter process library, babysitter marketplace plugin ideas, and implicit procedural knowledge that can be codified into babysitter JS processes.

Process Library Placement Rules

Extracted processes go into the babysitter process library (library/). Placement depends on scope:

Important: Do NOT place domain-specific processes in methodologies/. Only full, generic development methodologies belong there. A "k8s security audit" is specializations/security-compliance/, not a methodology. A "deep research pipeline" is specializations/shared/ (cross-domain). A "TDD agent workflow" is methodologies/atdd-tdd/ (full dev methodology).

Plugin Ideas = Babysitter Marketplace Plugins

A babysitter plugin is a set of natural language instructions (markdown) or deterministic coded processes (JS) that an AI agent reads and executes to install a modular set of capabilities. A plugin contains at minimum install.md with instructions the AI agent follows to modify the user's project. See docs/plugins.md for the full specification.

CRITICAL DISTINCTION: Plugin ideas should ONLY be things that modify project setup, install external integrations, or enforce workflows beyond just adding processes. Do NOT suggest plugins for:

• Skill pack collections: If you mark processes for extraction, don't suggest a plugin that just bundles those processes
• Expert/Role plugins: ".NET Expert", "React Native Expert", "Vue Development Suite", "Security Expert" - these are just skill packs
• Domain suites: "Frontend Development Suite", "DevOps Toolkit", "Data Science Suite" - these bundle processes
• Orchestration patterns: Multi-agent coordination, session continuity, workflow orchestration belong in babysitter core or as processes
• Process repackaging: Any plugin that just wraps processes you already marked for extraction

Valid plugin ideas change the project or setup (may not install skills at all):

• Project configuration changes: Modify CLAUDE.md/AGENTS.md instructions, update settings, configure behaviors
• External service integrations: GitHub API, Slack API, database connections, CLI tools, MCP servers
• Project enforcement mechanisms: Git hooks, ESLint rules, pre-commit checks, CI/CD pipeline templates
• Infrastructure and deployment: Docker configs, cloud provider setup, deployment templates, containerization
• Memory and persistence systems: Context storage, session state, cross-run memory, caching layers
• Development environment changes: IDE integrations, build tool configs, linting setups, editor extensions
• Workflow enforcement: Harness hooks, commit policies, pipeline triggers, quality gates, approval workflows
• Project structure modifications: Directory layouts, file templates, scaffolding, boilerplate generation
• Additional project functionality: New capabilities, tool chains, automation layers, monitoring integration

Rule of thumb: If it teaches babysitter how to do something → process. If it changes the project, adds external connections, or modifies behavior → plugin.

Valid plugin use case categories (derived from the existing marketplace):

IMPORTANT DISTINCTION: Do NOT confuse babysitter marketplace plugins with harness assimilation:

• Babysitter marketplace plugins: Install INTO user projects via install.md to add capabilities
• Harness assimilation: Create plugins FOR other harnesses (like hermes-agent) that integrate babysitter INTO those harnesses

When to use

• User asks to discover what skills or workflows exist in popular repos.
• User asks to research a specific repo's skill ecosystem.
• User asks to extract processes or patterns from external skills.
• Periodic refresh to track the evolving skill landscape.

Phase 1 -- Discovery

Search GitHub for repositories containing SKILL.md files. Use multiple search strategies to cast a wide net:

# Primary: find SKILL.md files in public repos
gh search code "filename:SKILL.md" --json repository,path,url --limit 100

# Supplementary: search for skill frontmatter patterns
gh search code "description:" "filename:SKILL.md" --json repository,path,url --limit 100

# Claude Code plugin skills specifically
gh search code "plugin.json" "skills" --json repository,path,url --limit 100

Topic-based discovery

Search for repos tagged with relevant GitHub topics. These are high-signal candidates even without SKILL.md files:

# Search by topic tags (each is a separate query)
for topic in claude-code claude-skills mcp agentic-workflow agent-skills skills agent-harness ai-agents; do
  gh search repos --topic "$topic" --stars=">50" --sort stars --limit 50 --json fullName,stargazersCount,description
done

# Combined keyword + star searches for broader coverage
gh search repos "agent skill" --stars=">50" --sort stars --limit 50 --json fullName,stargazersCount,description
gh search repos "claude code skills" --stars=">100" --sort stars --limit 30 --json fullName,stargazersCount,description
gh search repos "workflow automation skill" --stars=">100" --sort stars --limit 30 --json fullName,stargazersCount,description

Topic-tagged repos that lack SKILL.md files may still contain extractable processes or plugin ideas if they implement multi-step workflows, domain pipelines, or tool integrations. Classify and research them using the same Phase 2/3 pipeline.

Marketplace/registry discovery

Browse public skill and plugin registries for high-download or featured entries. These surface popular repos that may not appear in GitHub search:

• ClawHub Skills: https://clawhub.ai/skills?sort=downloads -- browse top skills by download count. Each skill links to a GitHub repo. Extract repo URLs and cross-reference with the tracked set.
• ClawHub Plugins: https://clawhub.ai/plugins -- browse plugins by popularity. Each plugin links to a GitHub repo. Extract repo URLs and cross-reference.

Use a browser tool or curl to fetch these pages and extract GitHub repo links. For each new repo found, enrich and classify using the standard pipeline.

Filtering rules

1. Drop any hit from a5c-ai/babysitter (this repo).
2. Handle archived/moved repos. If a repo is archived, check for a successor/migration notice. If the archive points to a new location (e.g., "moved to org/new-repo"), skip the archived repo and evaluate the new location instead. Only track active, maintained repositories.
3. Drop repos without a permissive license. Only track repos with MIT, BSD (2-clause or 3-clause), or Apache-2.0 licenses. Drop repos with GPL, AGPL, CC-NC, CC-SA, proprietary, or no license specified. Check license.spdx_id during enrichment.
4. Dedupe by repository.nameWithOwner.
5. Group hits by repo -- one repo may contain many SKILL.md files.
6. Prefer repos with 50+ stars. Lower-star repos may be included only if they contain exceptionally novel processes not found elsewhere. Use gh search repos with --stars=">50" to find higher-quality repos.

Enrichment

For each surviving repo:

gh api repos/<owner>/<name> \
  --jq '{nameWithOwner, description, stargazerCount: .stargazers_count, pushedAt: .pushed_at, topics, license: .license.spdx_id}'

Record the list of SKILL.md paths found per repo.

Phase 2 -- Classification

For each repo, shallow-clone into .a5c/tmp/skill-discovery/ and investigate the structure. Classify into exactly one archetype:

Classification signals

Read the repo's top-level README, plugin.json (if present), directory structure, and a sample of SKILL.md files. Look for:

• mega-skill-pack: skills/ directory with 5+ subdirectories, no primary application code
• methodology-repo: Process/workflow documentation dominates, SKILL.md describes a methodology
• internal-maintenance: SKILL.md references only internal paths, CI pipelines, repo-specific tooling
• other-harness: Skill is for Codex, Cursor, or another non-Claude harness; or focuses on CLI orchestration / harness invocation patterns
• claude-plugin: .claude-plugin/plugin.json or plugin.json with skill registrations
• harness-framework: CLI executable for AI interaction (like opencode, antigravity), or Claude Code orchestration/TUI/hook improvements (workflow automation, delegation frameworks, status line enhancements)
• domain-skill-pack: Skills all relate to one domain; directory structure groups by topic
• utility-with-skill: Repo is primarily a library/tool; SKILL.md is usage documentation

Phase 3 -- Deep Research

For each non-skipped repo, produce a single research.md file containing overview, assessment, and extractable value.

Harness Capability Verification: For repos classified as harness-framework, verify three critical capabilities for babysitter integration:

1. Custom Tools/MCP: Can execute custom tools, MCP servers, or bash commands
2. Stop Hooks: Has stop-hooks or end-turn hooks to interrupt agent conversation for feedback
3. Plugin System: Plugin/extension system with manifests and optionally marketplace

Use WebSearch/WebFetch to research the harness documentation and verify these capabilities. Stop hooks are CRITICAL - without them, babysitter's orchestration loop cannot function (harness must be interruptible between iterations for feedback).

Directory layout

• GitHub-sourced repos: docs/reference-repos/[org]/[repo-name]/research.md
• ClawHub-sourced skills/plugins: docs/reference-repos/clawhub/[author]/[skill-name]/research.md

Each tracked repo gets exactly one file (research.md) in its directory. Do not split into multiple files (no separate index.md or extractable-value.md).

research.md -- Unified research document

# [org]/[repo-name]

- **Archetype**: mega-skill-pack | methodology-repo | claude-plugin | domain-skill-pack | utility-with-skill
- **Stars**: N
- **Last pushed**: YYYY-MM-DD
- **License**: MIT / Apache-2.0 / BSD-2-Clause / BSD-3-Clause
- **Discovered**: YYYY-MM-DD
- **Source**: gh-search | clawhub-skills | clawhub-plugins | topic:X
- **Skills found**: N

## Summary
<2-3 sentences on what the repo provides and why it's interesting>

## Assessment
<What is transferable? What is repo-specific? Quality of skill design?
Look beyond methodologies -- domain-specific skills (DevOps, security, frontend, data, etc.)
often contain multi-step processes extractable as specializations/<domain>/ entries.
A "kubernetes-specialist" skill may encode a k8s deployment audit process.
A "debugging-wizard" may encode a systematic debugging process.
For harness-framework repos, assess: TUI/orchestration improvements for our internal agent harness,
CLI patterns for new harness adapter creation, and workflow automation patterns.
Assess each skill for procedural content, not just methodology content.>

## Extraction Priority
- High / Medium / Low
- Rationale: <why>

## Skills Inventory

| Skill | Path | Domain | Transferable? | Notes |
|-------|------|--------|---------------|-------|
| skill-name | skills/foo/SKILL.md | DevOps | Yes - pattern | Describes a CI/CD workflow |

## Processes
<Workflows that can be codified as babysitter JS processes.
Domain-specific skills are prime extraction targets -- a "react-expert" skill may contain
a component architecture review process (specializations/frontend/), a "terraform-engineer"
may contain an IaC audit process (specializations/devops-sre-platform/), etc.
Don't dismiss domain skills as "just expert personas" -- read them for procedural content.>
- **Process name**: Description of what it does
  - Source: path/to/SKILL.md (lines N-M)
  - Placement: methodologies/<name> | specializations/shared | specializations/<domain>
  - Inputs/Outputs: ...
  - Complexity: simple | moderate | complex
  - Notes: ...

## Plugin Ideas
<Ideas for babysitter marketplace plugins -- installable packages with install.md
that an AI agent executes to set up capabilities in a user's project>
- **Plugin name**: What it installs and configures
  - What install.md would do: <what the AI agent does during install -- detect stack, interview user, copy processes, set up hooks/configs>
  - Processes it would copy: <which process library entries>
  - Configs/hooks it would create: <ESLint rules, git hooks, CI/CD templates, etc.>
  - Source evidence: <what in the repo inspires this plugin idea>
  - Marketplace placement: <plugins/a5c/marketplace/blueprints/[category]/[plugin-name]/>

## Plugin Marketplace Mapping

<Check existing marketplace plugins before proposing new ones. Map plugin ideas against current plugins/a5c/marketplace/blueprints/ structure>

| Plugin Idea | Marketplace Status | Action | Existing Plugin | Target Placement |
|-------------|-------------------|--------|-----------------|------------------|
| Security Toolkit | UPGRADE | Enhance existing with new scanning processes | plugins/a5c/marketplace/blueprints/basic-security/ | plugins/a5c/marketplace/blueprints/security-toolkit/ |
| Testing Suite | NEW | Comprehensive testing framework | - | plugins/a5c/marketplace/blueprints/testing-suite/ |

**Example existing plugins** (from plugins/a5c/marketplace/blueprints/):
- `basic-security`, `agentsh`, `container-security` - Security tools and sandboxing
- `claude-mem` - Memory and context management
- `dev-browser` - Browser automation and tools integration
- `ctx` - Developer experience enhancements  
- `github-actions-cicd-*` - CI/CD integration templates
- `argocd-gitops`, `devcontainer` - DevOps and infrastructure
- `api-contract`, `changelog-enforcer` - Quality assurance tools
- `autorelease`, `changesets` - Workflow automation
- `contribution-graph`, `community-health` - Project health and metrics

**Plugin naming pattern**: `[descriptive-name]` - no category prefixes, direct plugin names

## Harness Integration Ideas
<For harness-framework repos: ideas for new harness adapters and TUI improvements>
- **Harness Adapter**: New harness integration (like plugins/babysitter-codex for Codex)
  - Adapter implementation: <what would go in packages/babysitter-sdk/src/harness/adapters/>
  - Plugin structure: <what would go in plugins/babysitter-[harness]/>
  - CLI integration: <command patterns, flag mapping, capability detection>
- **Harness Assimilation**: Plugin FOR the target harness that integrates babysitter (NOT a babysitter marketplace plugin)
  - **Capability Assessment**: Verify the harness supports babysitter's orchestration requirements:
    | Capability | Status | Details |
    |------------|---------|---------|
    | **Custom Tools/MCP** | ✅/⚠️/❌ | Can the harness execute custom tools, MCP servers, or bash commands? |
    | **Stop Hooks** | ✅/⚠️/❌ | Does it have stop-hooks or end-turn hooks to interrupt agent conversation for feedback? |
    | **Plugin System** | ✅/⚠️/❌ | Plugin/extension system with manifests and optionally marketplace? |
  - **Integration Viability**: EXCELLENT/GOOD/PARTIAL/POOR based on capabilities (stop hooks are CRITICAL)
  - Target harness plugin: <plugin that goes into the other harness to bring babysitter capabilities>
  - Babysitter integration: <how the other harness would invoke babysitter processes>
  - Capability bridge: <what babysitter features would be accessible from the target harness>
  - Major limitations: <any critical missing capabilities that would prevent full integration>
- **TUI/Orchestration Improvement**: Enhancement to our internal agent harness
  - Current limitation: <what our harness lacks that this repo provides>
  - Integration approach: <how to incorporate the improvement>
  - Implementation scope: <where in our codebase this would go>

## Implicit Procedural Knowledge
<Procedures that are described narratively in SKILL.md files but should be
codified as deterministic JS processes for the babysitter process library>
- **Procedure name**: What it accomplishes
  - Source: SKILL.md section or description text
  - Placement: methodologies/<name> | specializations/shared | specializations/<domain>
  - Why codify: <what makes this better as a process than a skill>
  - Sketch: <brief outline of phases/tasks>

Phase 4 -- Library Mapping and Re-extraction Analysis

CRITICAL: Check existing process library before creating new processes. Many high-value repositories have already been assimilated into the babysitter pr