| name | assimilate-popular-workflows |
| description | This skill should be used when the user asks to "find skills in the wild", "assimilate popular workflows", "discover SKILL.md files in repos", "research external skills", "find workflow patterns", "survey the skill landscape", "what skills exist out there", or wants to investigate public repositories for extractable processes, babysitter plugins, and reusable procedural insights. Searches GitHub for SKILL.md files, classifies repos by archetype, and maintains structured research under docs/reference-repos/. |
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:
| What it is | Where it goes | Examples |
|---|
| Full generic dev methodology (entire workflow paradigm) | methodologies/<name>/ | agile, gsd, tdd, scrum, kanban, waterfall |
| Common cross-domain pattern (reusable across many specializations) | specializations/shared/ | audit-pipeline, expert-advisory, progressive-disclosure |
| Domain-specific process | specializations/<domain>/ | security-compliance, devops-sre-platform, data-science-ml |
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):
| Category | What the plugin installs | Examples |
|---|
| Security & Sandboxing | Lint rules, git hooks, scanning processes, sandboxing policies | basic-security, agentsh |
| Context & Memory | MCP servers for memory, lifecycle hooks for auto-capture | claude-mem, mempalace |
| Knowledge Management | Wiki systems, knowledge graphs, semantic search engines | llm-wiki, graphify, qmd |
| Developer Experience & UX | Status indicators, session landing pages, skill recommenders | ctx, status-line, welcome |
| Tools Integration | Browser automation, external tool integration, MCP tools for new capabilities | dev-browser, prompt-master |
| CI/CD Integration | GitHub Actions workflows, harness-specific pipeline templates | github-actions-cicd-* |
| DevOps & Infrastructure | IaC templates, deployment configs, cloud provider setup | project-deployment |
| Quality Assurance & Testing | Test frameworks, coverage gates, linting configs, pre-commit hooks | testing-suite |
| Workflow Automation | Rate limit handling, auto-retry logic, lifecycle event hooks | rate-limit-handler |
| Theming & Environment | Sound hooks, design systems, conversational personality, themed assets | themes, sound-hooks |
| Harness Integration | Alternative harness adapters, TUI improvements, orchestration frameworks | opencode-adapter, workflow-orchestration |
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:
gh search code "filename:SKILL.md" --json repository,path,url --limit 100
gh search code "description:" "filename:SKILL.md" --json repository,path,url --limit 100
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:
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
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
- Drop any hit from
a5c-ai/babysitter (this repo).
- 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.
- 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.
- Dedupe by
repository.nameWithOwner.
- Group hits by repo -- one repo may contain many SKILL.md files.
- 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:
| Archetype | Description | Action |
|---|
mega-skill-pack | Repo exists to distribute many skills across domains | Deep-dive: catalog all skills, extract patterns |
methodology-repo | Repo represents a specific workflow or methodology | Extract the methodology as a potential babysitter process |
internal-maintenance | Skills exist only for the repo's own CI/dev workflow | Skip -- not transferable |
other-harness | Skill is specific to a non-Claude harness (Codex, Cursor, etc.) or focused on harness invocation/CLI orchestration | Skip -- not transferable to babysitter processes |
claude-plugin | A Claude Code plugin with skills as part of its offering | Investigate plugin structure, extractable integrations |
harness-framework | Alternative AI coding harness/framework (OpenCode, Antigravity, etc.) or Claude Code orchestration/TUI improvements | Extract for harness assimilation (new adapter + plugin) and/or TUI/orchestration improvements |
domain-skill-pack | Skills focused on a specific domain (e.g., data science, DevOps) | Extract domain processes and patterns |
utility-with-skill | A tool/library that ships a SKILL.md for usage guidance | Extract the usage pattern as a potential shared process |
not-a-skill | Repo uses SKILL.md as generic docs, no Claude Code connection | Skip -- no frontmatter, no agent context |
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:
- Custom Tools/MCP: Can execute custom tools, MCP servers, or bash commands
- Stop Hooks: Has stop-hooks or end-turn hooks to interrupt agent conversation for feedback
- 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 process library. Before extracting processes, map them against existing library content to identify:
- Direct matches - processes already implemented that could be enhanced with new insights
- Near matches - similar processes that could be generalized or specialized
- Gaps - novel processes not yet in the library
Library Structure Check
The babysitter process library is located at library/ with these key directories:
library/methodologies/ - Full development methodologies (agile.js, atdd-tdd/, bmad-method/, cc10x/, etc.)library/specializations/ - Domain-specific processes (ai-agents-conversational/, etc.)library/cradle/ - Core babysitter processes (bug-report.js, feature-request.js, etc.)library/contrib/ - User-contributed processes
Mapping Process
For each extractable process identified in Phase 3 research documents:
-
Search for existing implementations:
find library -name "*.js" -type f | grep -i "<process-concept>"
ls library/methodologies/
ls library/specializations/
-
Classify the relationship:
- UPGRADE - existing process that could be enhanced with new patterns/insights from the repo
- VARIANT - similar process that could be generalized or adapted
- NEW - novel process not represented in the library
- OBSOLETE - existing process that could be replaced with superior approach from repo
-
Document the mapping:
Add a "Library Mapping" section to each research.md:
## Library Mapping
| Extractable Process | Library Status | Action | Existing Path | Target Placement |
|-------------------|----------------|--------|---------------|------------------|
| Superpowers Debugging | UPGRADE | Enhance with new TDD integration patterns | methodologies/superpowers/superpowers-workflow.js | methodologies/superpowers/ (enhancement) |
| TDD Workflow | VARIANT | Could generalize atdd-tdd with pure TDD variant | methodologies/atdd-tdd/atdd-tdd.js | methodologies/pure-tdd/ (new variant) |
| Research Pipeline | NEW | Novel 23-stage autonomous research methodology | - | specializations/shared/autonomous-research.js |
| Security Audit | NEW | K8s security scanning process | - | specializations/security-compliance/k8s-security-audit.js |
Library placement rules for Target Placement:
- methodologies/[name]/: Full generic dev methodologies only (agile, tdd, scrum, kanban)
- specializations/shared/: Cross-domain reusable patterns (audit-pipeline, research-methodology)
- specializations/[domain]/: Domain-specific processes:
security-compliance/ - Security, compliance, auditing, scanningdevops-sre-platform/ - Infrastructure, deployment, monitoring, platformdata-science-ml/ - Data processing, ML workflows, analyticsfrontend/ - UI/UX, component architecture, design systemsbackend/ - API design, microservices, database, performancemobile/ - iOS, Android, cross-platform mobile developmentai-agents-conversational/ - Agent development, LLM integration patterns
Re-extraction Strategy
When a repository offers improvements to existing processes:
- Read the existing process to understand current implementation
- Extract the novel insights - what does the repository add that we don't have?
- Plan the enhancement - how to integrate new patterns without breaking existing functionality
- Document the upgrade path - what changes would be made and why
Example upgrade documentation:
### Upgrade Analysis: superpowers-workflow.js ← obra/superpowers debugging enhancements
**Current implementation**: Agent development methodology with TDD, debugging, and planning frameworks
**Repository insights**:
- Binary search debugging strategy
- Systematic error categorization (syntax/logic/integration/environment)
- Rubber duck debugging integration
- Prevention-focused root cause analysis
**Proposed enhancements**:
- Add binary search phase for large codebase debugging
- Implement error taxonomy classification within superpowers workflow
- Enhance debugging strategy selection logic
- Integrate prevention analysis into superpowers methodology
**Backward compatibility**: Existing superpowers methodology preserved, enhanced with new debugging patterns
Phase 5 -- Process Codification
For entries marked as NEW or UPGRADE from the library mapping analysis, proceed with process extraction. Use the process-builder skill patterns from .claude/skills/process-builder/SKILL.md.
For NEW processes:
Process files go in .a5c/processes/assimilated/ as staging candidates. After review, they are promoted into the process library at their designated placement path.
For UPGRADE processes:
- Create enhanced version in
.a5c/processes/assimilated/ with suffix -v2 or -enhanced
- Document the differences from the current version
- Plan migration strategy for existing users
- After review, replace or merge with existing process
.a5c/processes/assimilated/
├── [org]-[repo]-[process-name].cjs # Staged NEW candidate
├── [existing-process]-enhanced.cjs # Staged UPGRADE candidate
└── ...
# After review, promoted to process library:
# methodologies/<name>/ # Full generic dev methodologies only
# specializations/shared/ # Cross-domain reusable patterns
# specializations/<domain>/ # Domain-specific processes
Use .cjs extension because .a5c/package.json sets "type": "module".
Each process must:
- Import
defineTask from @a5c-ai/babysitter-sdk
- Export
async function process(inputs, ctx)
- Include
@references pointing back to the source SKILL.md
- Include
@process assimilated/[name] tag
- Include
@placement tag indicating the target library path (e.g. @placement specializations/security-compliance/k8s-audit)
- Include a
@graph JSDoc block referencing relevant atlas graph node IDs (domains, skillAreas, topics, roles, workflows). Read packages/atlas/graph/domain/ to find valid IDs. At minimum include one domain: node. Example: @graph\n * domains: [domain:software-engineering]\n * topics: [topic:security-scanning]\n * roles: [role:sre]
- Honour the source repo's license in the JSDoc header
Phase 6 -- Maintain indexes and history
Maintain three files in docs/reference-repos/ alongside the per-repo research directories:
README.md -- Master index of tracked repos
The main index of all repos with extractable value. Only repos that have research docs with at least one extractable process or plugin idea belong here.
# Reference Repos
<!-- Generated by .claude/skills/assimilate-popular-workflows. Re-run to refresh. -->
Last refreshed: YYYY-MM-DD
Total repos tracked: N
## By Archetype
### Mega Skill Packs
| Repo | Stars | Skills | Extraction Priority |
|------|-------|--------|---------------------|
| [org/name](org/name/research.md) | N | M | High |
### Methodology Repos
...
### Claude Plugins
...
### Domain Skill Packs
...
### Utilities with Skills
...
backlog.md -- Candidate repos to investigate
Repos discovered during Phase 1 that haven't been investigated yet. Append new candidates here during discovery; remove them once classified and either tracked (moved to README.md) or rejected (moved to processed.md).
# Candidate Backlog
| Repo | Stars | Source | Notes | Added |
|------|-------|--------|-------|-------|
| org/name | N | gh-search / clawhub / topic:X | Brief note on why it's a candidate | YYYY-MM-DD |
processed.md -- History of all evaluated repos
Every repo that has been investigated goes here, regardless of outcome. This prevents re-processing the same repo in future discovery runs. Include the classification result and reason for skipping (if skipped).
# Processed Repos
| Repo | Stars | Archetype | Outcome | Date |
|------|-------|-----------|---------|------|
| org/name | N | mega-skill-pack | Tracked -- 3 processes, 2 plugins | YYYY-MM-DD |
| org/other | M | internal-maintenance | Skipped -- no transferable value | YYYY-MM-DD |
| org/another | K | not-a-skill | Skipped -- generic docs, no agent context | YYYY-MM-DD |
Cleanup rules
- Do NOT keep research directories for skipped repos with no extractable value. If a repo is classified as
internal-maintenance, other-harness, not-a-skill, or otherwise has zero extractable processes and zero plugin ideas, record it in processed.md only. Do not create a directory under docs/reference-repos/.
- Only create
research.md for repos that have at least one extractable process or plugin idea. Each repo gets exactly one file (research.md), not separate index/extractable-value files.
- CRITICAL: Check for duplicates before processing. Before investigating ANY repository:
- Check
processed.md - skip if already evaluated
- Check
README.md - skip if already tracked
- Check existing
docs/reference-repos/[org]/[repo]/ directories
- Remove duplicates from
backlog.md when found
- License must be verified. Every
research.md must include the license field. During enrichment, extract license.spdx_id from the GitHub API. If the license is not MIT, BSD, or Apache-2.0, skip the repo and record it in processed.md with the reason.
Notes
- ALWAYS check existing library first. Before extracting any process, map it against the current process library (
library/methodologies/, library/specializations/) to identify UPGRADE opportunities rather than duplicating effort.
- Prioritize upgrades over new processes. Enhancing existing processes with new insights from high-value repositories often provides more value than creating entirely new processes.
- Never copy SKILL.md content wholesale. Extract the procedural insight, not the prose.
- Respect source licenses. Include attribution in every extracted process file.
- Skills that are purely prompt-engineering (just a system prompt with no procedure) have no extractable process value -- note them as
not-transferable in the inventory.
- Domain-specific skills are extraction targets, not just methodologies. A "kubernetes-specialist" skill may contain a k8s deployment audit process (
specializations/devops-sre-platform/). A "react-expert" may contain a component architecture review (specializations/frontend/). A "debugging-wizard" may contain a systematic debugging process (specializations/shared/). Always read domain skills for multi-step procedural content before dismissing them as "expert personas." The process library has three placement tiers: methodologies/ (full dev paradigms), specializations/shared/ (cross-domain patterns), and specializations/<domain>/ (domain-specific processes). Most extracted value goes into specializations, not methodologies.
- Skip skill-management processes (skill-routing, skill-discovery pipelines, skill-validation, skill-metadata checks). These are babysitter-internal concerns, not transferable domain processes. Their associated plugin ideas (e.g., a skill-registry-browser plugin) may still be valid.
- Skip multi-model coordination processes (multi-model review, heterogeneous AI team orchestration). Babysitter's harness adapter system already handles multi-model dispatch natively. These don't add value as library processes.
- Skip patterns already covered by the SDK: human-in-the-loop review cycles (covered by breakpoints), harness CLI invocation/degradation (covered by harness adapters), effect dispatch coordination (covered by the runtime). Only extract processes that add domain-specific or workflow-specific value beyond what the SDK primitives provide.
- Memory systems are always plugins, never processes. Memory management (tiered storage, decay, reflection, promotion) belongs in the Context & Memory plugin category. Do not place memory-related workflows in the process library -- they are plugin-internal logic installed via
install.md.
- The
internal-maintenance archetype is the most common. Expect 60-70% of hits to be skipped.
- Rate-limit awareness:
gh search code is throttled at 30 req/min. Split searches by language qualifier if hitting caps.
- When a repo appears in
processed.md, skip it unless explicitly asked to re-evaluate. For tracked repos (directory exists under docs/reference-repos/), compare pushedAt dates to decide if re-investigation is needed -- update in-place rather than recreating.
- Re-extraction for process upgrades: When explicitly asked to re-extract from high-value repositories to upgrade existing processes, update the existing
research.md with new insights and add the "Library Mapping" section to identify UPGRADE opportunities.
- For very large skill packs (20+ skills), sample the most-starred or most-recently-updated skills rather than researching all of them in a single pass.
- After completing research, suggest the user run
/babysitter:contrib for any upstream-worthy process candidates.
- See
references/classification-heuristics.md for detailed archetype classification examples and edge cases.
