Skip to content

sungjunlee/skills

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

32 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

skills

A small collection of portable agent skills.

These skills are meant to stay compact, composable, and easy to adapt. Each skill should do one useful thing, keep its core workflow in SKILL.md, and move only optional supporting material into references/.

Skills

Planning

brainstorming

Turn a vague idea into one selected direction and a compact Design Handoff, then recommend the next owner. Explicit-only and engine-agnostic.

brainstorming inspects safely discoverable repository context, asks only decision-changing questions, compares viable approaches, and converges on a Design Handoff. It recommends a successor (feature-spec, a tracker-neutral decomposition handoff, implement/relay, stop, or a grill capability) as output — it never invokes one, mutates a tracker, or infers capability availability.

/brainstorming "we need some kind of notifications, not sure where to start"

Source:

  • skills/planning/brainstorming/SKILL.md
  • skills/planning/brainstorming/references/routing.md
  • skills/planning/brainstorming/agents/openai.yaml

feature-spec

Compile a completed Design Handoff, settled proposal or conversation, or clear tracker task into a tracker-neutral, implementation-ready Feature Spec.

It preserves settled decisions and acceptance criteria, records exact repository-contract contradictions for human resolution, and closes with exactly one of an implement | relay recommendation, a decomposition handoff, or a blocked-human-decision Human Decision Handoff. It recommends successors without invoking them or mutating tracker state.

Source:

  • skills/planning/feature-spec/SKILL.md
  • skills/planning/feature-spec/agents/openai.yaml
  • skills/planning/feature-spec/references/spec-template.md
  • skills/planning/feature-spec/references/routing.md

Engineering

implement

Execute a settled feature spec, tracker task, or clear prompt in the current checkout with the lightest safe current-session engine.

implement performs a pre-edit safety gate, chooses inline, serial-worker, or bounded-parallel execution, and keeps diff inspection and authoritative verification with the orchestrator. Work that needs durable recovery, isolated lifecycle, or elevated-risk handling is returned as a relay handoff before mutation; relay is never invoked silently.

Source:

  • skills/engineering/implement/SKILL.md
  • skills/engineering/implement/agents/openai.yaml
  • skills/engineering/implement/references/routing.md
  • skills/engineering/implement/references/worker-contract.md

Productivity

delegate

One-shot delegate a prompt to a chosen agent/model and return its output. No worktree, manifest, or review.

When the full relay-dispatch flow is overkill — you just want a quick answer from a specific model — delegate shells out directly to the chosen CLI.

/delegate opencode-go/deepseek-v4-pro "refactor this function to use streams"
/delegate reasonix/deepseek-v4-pro "write unit tests for the parser"
/delegate cline-pass/glm-5.2 "review this diff"
/delegate opencode/glm-5.2 "summarize this diff"
/delegate opencode "explain this file"  # no model: CLI default

Source:

  • skills/productivity/delegate/SKILL.md
  • skills/productivity/delegate/references/cli-invocations.md
  • skills/productivity/delegate/references/dispatch-guardrails.md
  • skills/productivity/delegate/references/model-catalog.md

Review

gosu-review

Run a real multi-subagent panel review on the current artifact.

Single-agent reviews collapse into one voice. gosu-review keeps each panelist's independent lens visible, then surfaces the orchestrator's tensions, consensus, and per-persona verdicts.

Use it when you want multiple context-specific experts to review code, a plan, a skill, a document, a decision, or repository state.

/gosu-review
/gosu-review skills/review/gosu-review/SKILL.md
/gosu-review "review this product launch plan"

Source:

  • skills/review/gosu-review/SKILL.md
  • skills/review/gosu-review/references/personas.md

Related Larger Projects

These are intentionally not included here because they are heavier workflows rather than small drop-in skills.

  • dev-backlog — keeps GitHub Issues as the source of truth while adding local sprint files as the execution hub for humans, Claude Code, and Codex.
  • dev-relay — runs a repeatable plan -> dispatch -> review loop where an executor works in an isolated worktree and an independent reviewer checks the PR before merge.
  • CraftKit — file-first toolkit for authoring, critiquing, tuning, surveying, and carrying forward prompts and coding-agent skills.

Install

Skills in this repository use the required nested path skills/<category>/<skill-name>/. Ask the skills CLI to scan the full depth; without --full-depth, nested skills may not be discovered.

List all discoverable skills before installation:

npx skills add . --list --full-depth

Install from this repository with the same full-depth scan:

npx skills add . --full-depth

Alternatively, copy or symlink the desired skill directory into your agent's configured skills directory.

For example, to install gosu-review:

ln -s "$PWD/skills/review/gosu-review" ~/.agents/skills/gosu-review

Adjust the destination for your runtime. Codex and Claude Code may use different skills directories depending on local setup.

Repo Layout

docs/
  engine-capability-contract.md
evals/
  schema/
  fixtures/
  cases/
  results/
  reports/
scripts/
  verify-replays.mjs
skills/
  engineering/
    implement/
      SKILL.md
      agents/
        openai.yaml
      references/
        routing.md
        worker-contract.md
  planning/
    brainstorming/
      SKILL.md
      agents/
      references/
    feature-spec/
      SKILL.md
      agents/
        openai.yaml
      references/
        routing.md
        spec-template.md
  productivity/
    delegate/
      SKILL.md
      references/
  review/
    gosu-review/
      SKILL.md
      references/

Conventions

  • Keep each skill small enough to read quickly.
  • Prefer strong workflow rules over long explanations.
  • Use references only for optional seeds, examples, or checklists.
  • Avoid pretending to use tools that are unavailable.

Maintainer Verification

Run the semantic replay verifier and confirm nested skill discovery before landing changes:

npm test
npx skills add . --list --full-depth
git diff --check

The engine capability contract in docs/ is maintainer guidance. Runtime skills must remain self-contained and must not depend on reading it.

About

Small collection of portable agent skills

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors