My personal collection of Claude Code skills and workflows.
Skill directories live in skills/.
| Skill | Description |
|---|---|
| agent-engineering | Reference for LLM agent development pitfalls and best practices — context engineering, KV-cache / prompt-caching prefix stability, tool (ACI) design, agent loop architecture, multi-agent trade-offs, evals, observability, and security. Distilled from primary sources (Anthropic, OpenAI/Codex, Manus, Cognition, pi, LangGraph, SWE-agent) with load-bearing numbers and citations; a decision-ladder + checklist entry point with six topic references loaded on demand. |
| artifact-design | Design process and taste guidance for building distinctive, production-grade frontend interfaces (Claude Artifacts, web components, pages, apps) — brainstorm a token system, critique it, commit the palette, then build, with render-verified mechanics so output never reads as a generic template. |
| office-hours | Product Office Hours — a structured product thinking partner. Startup mode runs six forcing questions (demand reality, status quo, specificity, narrowest wedge, observation, future-fit); Builder mode does design-thinking brainstorming for side projects, hackathons, learning, and open source. Produces a design doc, never code. Inspired by GStack office-hours. |
| sub-claude | Batch-process CSV/JSONL data by spawning parallel claude -p agents. Each worker is a full Claude agent with tool access (read files, run commands, call APIs) — not just text generation. For 10–1000+ rows of uniform processing where you need fine-grained control over concurrency, retries, cost budget, and structured output. Includes a ready-to-use sub_claude.py. |
Workflow scripts live in workflows/. By default they run on Claude Code's native dynamic workflows (the Workflow tool).
They're also portable via six-ddc/codex-dynamic-workflows — a Claude Code-compatible port that runs the same scripts on other agent CLIs: OpenAI Codex, Gemini CLI, or pi (a general-purpose agent CLI spanning many more model providers) — with a live web run viewer.
| Workflow | Description |
|---|---|
| code-review-extracted | Workflow-backed code review — one finder agent per review angle, an independent verifier for every candidate, then a ranked, capped findings report. Launched by the /code-review skill at high / xhigh / max effort. |
| deep-research-extracted | Deep research harness — fan-out web searches, fetch sources, adversarially verify claims, and synthesize a cited report. |
An orchestrator command paired with model-pinned subagents for running a session with Claude Fable 5 as the lead. Command lives in commands/, agents in agents/.
| File | Description |
|---|---|
| fable | /fable — puts the session in orchestrator mode: the main model only plans, dispatches, adjudicates, and synthesizes. Work is routed by role (deep-reasoner / fast-worker / verifier / Explore) and scaled by model tier via the Agent model override; nontrivial changes get independent verification by verifier, with intensity scaled to risk. |
| handoff | /handoff — distills the current session into a self-contained handoff document (handoff-YYYY-MM-DD-<slug>.md in the working directory): task state, key trade-off decisions at full fidelity, background constraints, and next steps, so a fresh session or agent can pick up the work without re-asking. Sources from the conversation context only — no investigation to fill gaps. |
| deep-reasoner | Pinned to Opus at xhigh effort. Architecture design, complex-bug root-cause analysis, algorithm design, trade-off analysis. |
| fast-worker | Pinned to Sonnet at medium effort. Boilerplate, well-specified changes, running tests/lint/build, batch operations. |
| verifier | Pinned to Sonnet at high effort (pass model: opus for critical changes). Falsification-oriented: re-runs tests and checks each acceptance criterion with fresh context rather than trusting the implementer's self-report. |
Symlink a skill into ~/.claude/skills/:
ln -s /path/to/skills/skills/<skill-name> ~/.claude/skills/<skill-name>Symlink a workflow into ~/.claude/workflows/:
ln -s /path/to/skills/workflows/<workflow-name>.js ~/.claude/workflows/<workflow-name>.jsSymlink the command and agents into ~/.claude/commands/ and ~/.claude/agents/:
ln -s /path/to/skills/commands/fable.md ~/.claude/commands/fable.md
ln -s /path/to/skills/agents/<agent-name>.md ~/.claude/agents/<agent-name>.md