Reusable agent skills and workflow helpers for Claude and Codex.
Agents: see INDEX.md for a flat map of everything.
# 1. Prerequisites
gh auth login # GitHub CLI, authenticated
# 2. Clone and install
git clone https://github.com/fbserg/skill-issue
cd skill-issue
# Follow docs/install.md, or run the symlink commands in the Install section below.
# 3. Use
# In Claude Code or Codex, invoke:
/epic-plan "add dark mode to the settings page"
# The agent scopes and drafts a GitHub epic with child issues.
# Execute children via:
/resolve-issue <child-number>| Skill | Description |
|---|---|
adversarial-review |
Read-only red-team pass over a plan or diff before risky work lands. |
blitz |
Fast execution of ad-hoc lanes in parallel worktrees. |
epic-plan |
Scope broad work into a tracker issue and right-sized child issues; materializes on GitHub with no approval gate. |
issue |
Front door for GitHub issue work; scopes ideas or routes issue numbers to resolve-issue. |
resolve-issue |
One issue to review-ready PR in an isolated worktree; never merges. |
zero |
Shared destructive cleanup workflow with proof-before-delete classification. |
| Skill | Description |
|---|---|
adversary |
Cross-model red-team: sends a plan or diff to Codex (GPT) for an adversarial pass before committing. |
blitz |
Lightweight executor for ad-hoc lanes: parallel worktrees, no pipeline ceremony. Owns all batches, filed or not. |
deep-research |
Opus-planned multi-source research with disconfirmation lens, GRADE evidence tiers, and a saturation loop. |
epic-plan |
Planner: contract plus child issues that execute via /resolve-issue and /blitz; re-enters from GitHub state. |
resolve-issue |
One issue to review-ready PR in a worktree; never merges. Bounces a true epic to /epic-plan. |
patchcue-resolve-issue |
PatchCue worker workflow for Claude Code; same marked plan/outcome protocol as the Codex resolve-issue skill. |
simplify-sweep |
Batch-clean a pushed commit range via headless Sonnet /simplify per area; orchestrator reviews and commits. |
Claude skills remain available for Claude Code. Codex uses the separate
skills/codex/ tree so tool names and workflow assumptions stay runtime-native.
| Skill | Description |
|---|---|
authenticity-check |
Score human authenticity and flag AI-sounding spans. Not installed by default. |
humanizer |
De-slop AI prose. Not installed by default. |
zero |
Integrate active work and draft PRs, preserve parked work, clean merged branches, push. |
| You have | Use |
|---|---|
| An issue number | /resolve-issue <N> |
| A rough idea (no issue yet) | Ask the agent to file it, then /resolve-issue <N> |
| Multiple issues | /blitz 42 43 44 — parallel lanes |
| Ad-hoc lanes, nothing filed | /blitz — parallel worktrees, no issue required |
| A true epic (multi-session, multiple deliverables) | /epic-plan <topic> → /resolve-issue <child-number> |
Claiming (assign yourself), plan-comment-before-branch, and PR-only delivery are built in — you never merge your own PR.
- Git
- GitHub CLI (
gh) authenticated:gh auth login - Claude Code with local skill directory support
See docs/install.md for full instructions.
skill-issue is the canonical source for these skills. Edit the files in this
checkout, then run python3 scripts/check-install.py to catch any local install
that points at an older copy. Run python3 scripts/check-links.py to verify that
all relative markdown links in the repo resolve on disk.
# Quick install (symlinks — edits to the repo are immediately live)
./scripts/install.shSee INDEX.md for the full tools/scripts table.
The Herdr category holds project placeholders, agent state ages, undo closed tabs, pane sharing, a Codex cache countdown, and a quota display. Each package has its own instructions and tests. Plugins install separately from skills. Publishing these tools does not migrate existing host installations.
Three standalone Claude Code hooks in hooks/claude/. They started as copies of the author's live hooks. The live versions have since moved on, so treat these as starting points, not a mirror.
| Hook | What it does |
|---|---|
pretool-bash.sh |
Block destructive Bash commands (recursive delete of system paths, force-push to main, git reset --hard, dropping tables), Bash writes that escape an isolated worktree, and bare git stash on a dirty tree. |
caffeinate.sh |
Keep a Mac awake while the Claude session lives. No-op elsewhere. |
confetti-gate.sh |
Fire Raycast confetti after a successful main push. macOS + Raycast only. |
See hooks/claude/README.md for settings.json snippets and what was removed.
The authenticity-check and humanizer skills
are vendored from aihxp's Scriveno project
(authenticity-check,
humanizer) and used under the MIT License.
Each vendored skill keeps its original LICENSE (copyright preserved as required);
everything else here is MIT under LICENSE.
Experimental. Read each skill before using it on an important repository.
zero integrates active work, resolves draft PRs, and removes proven merged branches and worktrees. Explicitly parked work stays untouched. Use only at a deliberate cleanup point.
Retired skills and tools are deleted, not archived. Find them in git history.