Skill Inventory & Patterns

Personal inventory of existing skills and cross-cutting patterns.

Local Skills (research-lab/.claude/skills/)

SkillCategorySophisticationKey Pattern
proposalWorkflowHighLifecycle state machine (IDEATE->SYNTHESIZE)
researchWorkflowVery HighMulti-AI investigation + prompt enrichment
learnDocumentHighProgressive complexity (concepts->experiments)
design-arenaWorkflowExtremeMulti-model tmux orchestration + Caddy proxy
issueWorkflowHighDual-mode (investigation vs. discovery)
pr-reviewWorkflowExtreme3 parallel review subagents + tech-debt lifecycle
exploreDocumentHighProfile-aware code-grounded analysis
swarmWorkflowExtremeTwo-pass hybrid (tmux agents + Task subagents)
setup-repoWorkflowMediumSSH-first bootstrapping
profileWorkflowHighGitHub-aware temporal staleness detection

Global Skills (~/.claude/skills/)

SkillCategorySophisticationKey Pattern
agiWorkflowVery HighAutonomous tmux agents with session persistence
ai-reviewWorkflowHighConfidence x severity matrix, language checklists
breakdown-planWorkflowHighPhase decomposition with entrypoint prompt
committeeWorkflowHighMulti-model deliberation with synthesis
implement-planWorkflowHighSequential subagent execution per phase
initial-planWorkflowHighProposal-to-plan with auto-review pass
orchestrate-planWorkflowExtremeEnd-to-end autonomous pipeline (5 chained skills)
review-breakdownWorkflowVery HighDual-lens parallel review + direct fix application
redate-commitsWorkflowVery HighGitHub activity analysis + proportional gap distribution
tmuxWorkflowMediumFoundation patterns for agent spawning

Cross-Cutting Patterns

Auto-bootstrap: Check if infrastructure exists, create if not (setup-repo, issue, explore, swarm)

Code-grounded analysis: Reference actual file paths and line numbers, not summaries (issue, explore, pr-review)

Multi-AI orchestration: Run multiple models in parallel via tmux (research, design-arena, swarm, agi, committee)

Sentinel files: Monitor completion via marker files instead of parsing output (design-arena, swarm)

Pipe-prompt delivery: cat /tmp/prompt.md | claude -p to avoid shell escaping (swarm)

Archival strategy: Current round always has clean names; previous rounds renamed to -prev-{N} (agi, committee)

Dual-lens review: Two complementary review perspectives in parallel, then consolidation (pr-review, review-breakdown)

Progressive disclosure in practice: All 20 skills keep SKILL.md self-contained. None use scripts/ or references/ directories. All logic is embedded in the markdown.

Notable Techniques

  • tmux send-keys pattern: Always send text and Enter as separate commands with 0.3s sleep between
  • unset CLAUDECODE: Required before spawning Claude in tmux (prevents nested-instance error)
  • Pane ID capture: Use plain variables, not arrays (bash/zsh indexing differs)
  • Destructive command audit: orchestrate-plan scans for dangerous commands before running
  • Temp file tracking: Log temp files, add to .gitignore, clean up at end