catstack

A personal collection of Claude Code skills, consolidated from various project repos so they have one canonical home.

Layout

Each skill lives under skills/<name>/ as a standard SKILL.md package.

Global CLAUDE.md

CLAUDE.md at the repo root is the canonical copy of ~/.claude/CLAUDE.md (personal, cross-project Claude Code instructions — communication rules, evidence rules, session hygiene). install.sh symlinks it into place the same way it symlinks skills; edit here, not in the live ~/.claude/CLAUDE.md, so every machine picks up the change on the next git pull. Claude-only by nature (it's Claude Code's own config file), so install.sh doesn't try to place it for Cursor or Codex.

Hooks

• hooks/diu-stop/ -- a stop-time backstop for the diu skill: checks the final response against diu's brevity rule and pushes back if it looks skipped. Not a single shared file, since Claude Code, Cursor, and Codex CLI each have a genuinely different amount of power at stop time (hard block, soft one-shot nudge, or notify-only with no enforcement at all) -- see hooks/diu-stop/README.md for the per-harness breakdown and install steps.

Skills

• i-have-adhd — the skills/i-have-adhd/ subtree of EdbertChan/i-have-adhd (a fork of ayghri/i-have-adhd), pulled in via git subtree. The upstream repo has extra tooling (hooks, tests, extensions) outside this one skill folder, so a plain subtree add/pull against its root would drag all of that in too — updates instead go through a two-step split:

  # 1. re-clone the fork and extract just the skill folder's history into a branch
  git clone https://github.com/EdbertChan/i-have-adhd /tmp/i-have-adhd-src
  git -C /tmp/i-have-adhd-src subtree split --prefix=skills/i-have-adhd -b extracted
  
  # 2. pull that branch into catstack as a subtree update
  git subtree pull --prefix=skills/i-have-adhd /tmp/i-have-adhd-src extracted --squash

• draft-pr, split-scope — generic PR-drafting and diff-splitting skills, copied from DrafterSkill (packages/skill/skills/), current as of commit d4bb326. The invoker-make-pr and invoker-review-compression skills used in the Invoker repo are project-specific forks of these two.

• principle-* (14 skills) — cherry-picked from the pstack plugin in cursor/plugins (commit 63d938c), by Lauren Tan. Each one is a short, narrowly-triggered engineering rule. Selected after backtesting all 21 pstack principles against real Invoker/smithers/catalyst/etc. transcript history — these 14 either matched a proven-good habit or a real, verified past failure; the other 7 (prove-it-works — redundant with invoker-prove-it/process-guard — plus 6 more still under review) were left out for now:

  • outcome-oriented-execution, foundational-thinking, type-system-discipline — each tied to a real, verified production bug
  • laziness-protocol, fix-root-causes, separate-before-serializing-shared-state, sequence-verifiable-units, build-the-lever, encode-lessons-in-structure, never-block-on-the-human, experience-first — already an established habit in practice
  • subtract-before-you-add, minimize-reader-load, guard-the-context-window — mixed evidence, kept as a guardrail

  This was a manual cherry-pick, not a git subtree — there's no single upstream prefix that maps to "these 14 skills," so there's nothing to subtree pull. To refresh one, re-copy pstack/skills/<name>/SKILL.md from the source repo above at whatever commit is current.

• reflect — mines a conversation transcript for durable learnings and routes them into skill edits, gated on explicit user approval before anything is touched. Adapted from pstack's reflect, rewritten for Claude Code (its own transcript layout, the Agent tool for review fan-out, no dependency on Cursor's create-skill). Claude-only — install.sh skips it for Cursor/Codex. Findings about the user's working style rather than the code route to automate-me instead of becoming an inline skill edit.

• automate-me — mines session history (single-session via reflect's transcript lookup, or cross-session/cross-machine via reflect's corpus_scan.py) for durable working-style preferences and drafts or updates one personal <handle>-mode skill. Adapted from pstack's automate-me, rewritten for Claude Code the same way reflect was: AskUserQuestion in place of Cursor's AskQuestion, direct file writes in place of create-skill, and catstack's own principle-minimize-reader-load/principle-subtract-before-you-add in place of the missing unslop skill. Claude-only, same reason as reflect — install.sh skips it for Cursor/Codex.

• diu — communication-brevity rule (ELI5 under 40 words unless the user asked for depth), extracted verbatim from invoker-diu in the Invoker repo's own skill set. Zero Invoker-specific content, was just filed under the wrong prefix.

• land-stack — SHA-verified PR-stack landing (never resolve the PR to merge by branch name). Generalized from invoker-land-stack: kept the discipline, stripped the Invoker-specific guard script path since that script doesn't exist outside that repo.

• visual-proof — trustworthy before/after visual evidence for a UI-affecting change: never reuse a stale asset, actually open and look at the capture before claiming what it shows, handle genuinely-uncapturable states honestly instead of faking a screenshot. Generalized from invoker-visual-proof: kept the discipline, replaced Playwright/R2-specific capture mechanics with a tool-agnostic before/change/after/compare workflow.

• loop-generator — interview-driven generator for a reusable babysit/watch/retry loop: a loop instruction doc (goal, real target, success/fail invariants, evidence sources, exit conditions) plus a driver script with real safety rules (rebuild the live target set every run, dedupe by identity, never silently widen from inspection into writes). Generalized from invoker-loop-generator: dropped the third artifact (an Invoker-workflow-orchestrator YAML) since that assumed a specific internal system with no analog elsewhere; kept the interview and the two generic artifacts.

  The rest of the invoker-* skills (invoker-remote-ci-verify, invoker-workflow-chain-submit, invoker-invoker-ops, invoker-invoker-setup, invoker-plan-to-invoker) hard-depend on Invoker's own scripts, CLI, database, or YAML schema — correctly stay project-scoped, not catstack candidates. invoker-prove-it is redundant with process-guard. invoker-admin-bypass-sweep is a separate, real gap: it exists only as a local ~/.claude/skills copy with no matching file anywhere in the Invoker repo — it needs to be committed into Invoker/skills/, not here (catstack is for portable skills; that one is Invoker-specific and dangerous by design).