Skills

My personal skills, grouped by family.

Long-horizon & multi-agent loops

Orchestrators that drive a goal to completion over many turns. long-haul and dual-haul are multi-skill Claude Code plugins — the orchestrator plus its phase skills ship together.

long-haul — solo goal loop (plugin, v1.2.0)

Skill	Description
long-haul	Orchestrator: sharpen a spec → forge a /goal → haul toward it turn after turn in a git worktree, each turn choosing explore (fresh approach) vs exploit (refine incumbent), keeping only measured wins, until the goal holds
haul-spec	Phase 1 — turn the ask into a located, verifiable spec (.longhaul/SPEC.md)
haul-goal	Phase 2 — turn the spec into a /goal-ready completion condition (.longhaul/GOAL.md)
haul-loop	Phase 3 — the per-turn explore/exploit loop; --background detaches an overnight driver
haul-ship	Phase 4 — report goal-vs-achieved, commit the incumbent, open a PR

dual-haul — Claude-vs-Codex race (plugin, v1.0.0)

Skill	Description
dual-haul	Orchestrator: race a different-model Claude against Codex (each in its own worktree) round by round under a /goal, merge the measured winner, then ship a PR. Shared state in .dualhaul/
dual-understand	Phase 1 — improvement-focused brainstorming (.dualhaul/UNDERSTANDING.md)
dual-goal	Phase 2 — turn the understanding into a /goal-ready condition (.dualhaul/GOAL.md)
dual-loop	Phase 3 — the per-turn race: two implementers, judge against the goal, merge the winner
dual-report	Phase 4 — report goal-vs-achieved and who won each round (.dualhaul/SUMMARY.md)
dual-ship	Phase 5 — commit on a feature branch, open a PR, tear down worktrees

Bounded consultation & optimization

Skill	Version	Description
single-consult	1.0.0	Fast in-session sibling of pair-consult: the reviewer is an internal Claude subagent (default fable), not an external Codex peer — so no shared-state/handoff machinery. Same propose · review · respond · re-review · synthesize shape. --model sonnet|opus|fable, --depth, --number
pair-consult	1.2.0	Bounded consultation on one question — A proposes, B reviews, A responds, B re-reviews, A synthesizes and asks the user. Round count and peer effort configurable
pair-optimize	1.1.0	Optimization loop for a DuckDB/SQL query or hot Python path — A measures a baseline + proposes, B challenges, A benchmarks, B audits. Hard rule: no win kept without a measured speedup AND identical output
pair-ratchet	1.0.0	Outer loop over pair-optimize: profile a whole hot path, optimize the dominant bottleneck then the next — each kept win ratchets the baseline — until no session yields a win (loop-until-dry) or a session cap trips

Flags

/pair-consult "<question>"                        # default: 5 rounds
/pair-consult "<question>" --number 7             # n rounds (odd, ≥3)
/pair-consult "<question>" --model high|xhigh     # external peer effort

/single-consult "<question>"                      # internal Claude reviewer, default fable
/single-consult "<question>" --model opus         # reviewer model: sonnet|opus|fable
/single-consult "<question>" --depth xhigh        # reviewer reasoning effort

Unattended cleanup

Skill	Version	Description
lunch-clean-loop	1.0.0	Kick it off, go eat lunch: scans the codebase for readability debt, then loops up to N targets (default 5) through a fresh cleaner → reviewer → tester trio — a cleanup is kept only if an independent reviewer approves it AND a tester demonstrates identical output, else it's reverted. One refactor: commit per kept target on a clean-loop/<date> branch in its own worktree; the user's checkout is never touched. Python cleaners apply the simplify-python ruleset

Usage

/lunch-clean-loop                     # whole repo, N=5
/lunch-clean-loop src/pipeline --n 2  # scoped, 2 iterations

Plan review

Skill	Version	Description	Source
review-markdown-plan	1.0.0	Deprecated (use pair-*). Review a markdown plan by section, annotating only concretely wrong steps	—
get-feedback-markdown-plan	1.0.0	Deprecated (use pair-*). Evaluate the inline feedback review-markdown-plan left, verifying each critique against the codebase	—

Skill & MCP hygiene

Skill	Version	Description	Source
organize-skill-mcp	1.0.0	Scan skill and MCP configs across Claude Code, Claude Desktop, Cursor, Windsurf, Codex, and Gemini CLI — global and per-project — and suggest cleanup	—
find-overlapping-skills	1.0.0	Collapse double-installs, cluster the rest into overlap groups, then resolve each group with the user, disabling non-keepers reversibly	—
inspect-skill-flow	1.0.0	Audit a skill or skill family without editing it: build a typed dependency graph of its flow (steps, artifacts, invocations), render it via draw-graphology, run six weak-link checks (broken handoffs, dead outputs, orphans, dangling pointers, missing criteria, ungated destruction), verify every finding with two refute-only subagents, and write the survivors to a .inspect/ handoff file of ready-to-paste lessons for improve-skills. Reach for it after restructuring a skill family, before trusting a chain with unattended automation, or after installing/symlinking skills into a new home — the moments that leave dangling references and broken artifact contracts behind. User-invoked /inspect-skill-flow	—
improve-skills	1.0.0	Fold one lesson into one existing skill — collect lessons from the session or a direct ask (e.g. one finding from an inspect-skill-flow handoff), propose the exact edit, gate on user confirmation + feedback, then apply to a single skill (user-invoked /improve-skills)	—
agents-md-improver	1.0.0	Audit → score → propose targeted improvements to project-memory files, exactly like the official CLAUDE.md improver but writing into AGENTS.md (with CLAUDE.md a one-line @AGENTS.md reference) so the same memory is shared across Claude Code, Codex, Gemini CLI, …	Copied verbatim from claude-md-improver, one line added to redirect the write target

Writing & code

Skill	Version	Description	Source
humanizer	2.8.1	Remove signs of AI-generated writing (significance inflation, em dashes, rule of three, AI vocabulary, filler, …) so text reads as human-written; based on Wikipedia's "Signs of AI writing"	Source from blader/humanizer, improved here
simplify-python	1.0.0	Make recently written Python read better without changing behavior — idiomatic rewrites, flattened control flow, removed cruft. High-confidence, behavior-preserving rules only	—
slack-comm	1.0.0	Summarize the current Claude session's work into a very concise 3P (Progress/Plans/Problems) Slack status report, review it in chat, then stage it as a native Slack draft (via Slack MCP) for a final look before you send. User-invoked mid-session: /slack-comm	3P/tone format adapted from anthropics/skills internal-comms

Utilities

Skill	Version	Description	Source
draw-graphology	1.1.0	Data-driven network graph visualizations (interactive HTML + static PNG) with graphology + sigma.js — ForceAtlas2 layout, Louvain community colors, degree-sized hubs, and a Playwright screenshot loop so the agent validates its own render. For node-edge datasets too large to place by hand	Methodology ported from the Understand-Anything dashboard
excalidraw-diagram	1.0.0	Create Excalidraw diagram JSON files that make visual arguments — hand-placed conceptual diagrams, vs draw-graphology's computed layouts	Improved from coleam00/excalidraw-diagram-skill
loom-transcript	1.0.0	Fetch a public Loom video's transcript from its share link (oEmbed + share-page scrape, stdlib Python, no auth) and summarize it. User-invoked: /loom-transcript <url>	—

Install

npx skills add ccomkhj/skills

long-haul and dual-haul are multi-skill plugins, so their bundled skills live one level deeper than the flat scan looks. To install them via npx skills, add --full-depth:

npx skills add ccomkhj/skills --full-depth

As Claude Code plugins

long-haul and dual-haul are also published as Claude Code plugins (bundled skills + shared state). Inside Claude Code:

/plugin marketplace add ccomkhj/skills
/plugin install long-haul@ccomkhj-skills
/plugin install dual-haul@ccomkhj-skills

Manual setup

Alternatively, clone the repo and run sync.sh — it symlinks every skill (top-level and plugin-bundled) into both ~/.claude/skills (Claude Code) and ~/.agents/skills (Codex, Gemini CLI, …), one link per skill. Idempotent, and never clobbers a real file:

git clone git@github.com:ccomkhj/skills.git
./skills/sync.sh