AI workflows for thinking and building
Two independent entry points: think-through helps you choose and apply thinking tools; workflow combines Matt Pocock coding flows with real application proof and relevant pstack/Emil checks. Upstream files stay unchanged.
Install in a repository
Run from the repository where you want the skills:
npx --yes skills@1.7.2 add kolodziejczakM/workflows --skill '*' --agent codex cursor claude-code --yes
Installs all required skills into .agents/skills/, with Claude Code links. No manual clone needed. Private repositories require GitHub access through your configured Git credentials. Keep only your runtimes in the command, then start a new agent session.
Install only think-through globally
Run from any directory:
npx --yes skills@1.7.2 add kolodziejczakM/workflows --global --skill think-through --agent codex cursor claude-code --yes
Installs only think-through for your user account, available across repositories in the selected runtimes. Keep only the runtimes you use, then start a new agent session. Private repositories require the same GitHub access described above.
Think through a problem
$think-through I am choosing between moving closer to work and keeping more space. Help me work through the tradeoffs.
It identifies what blocks your answer, chooses a suitable strategy, and starts applying it. It covers understanding, explanations, estimates/forecasts, and decisions, adding analogy, systems thinking, premortems, or Bayesian updating when useful. Alternatives are discussed when they clarify the choice or you ask. Each completed exercise records a conclusion, key assumption, next step, and when to check the result.
It replies in your language. In the first response, it names the chosen thinking tool and explains why it fits in one sentence, without PDF tool numbers, then gives one or two steps you can take immediately. It does not repeat the introduction on later turns. In the next round, it follows up on the critical assumption if it remains unverified.
For complex or uncertain problems, it shows a compact analysis map: the question, evidence and its basis, blocking unknown, and next step with what it would resolve. The map updates when evidence or the recommendation changes and is skipped for simple answers. It uses a table or diagram without extra quizzes or setup; show-me is optional. This presentation has not been shown to improve human understanding or answer quality.
This is an English adaptation of your thinking-tools framework; the original PDF is archived for source audits and excluded from installed skills. It works in chat without Matt/pstack setup and does not start coding or create a PR. For a repository-local install of this skill alone, replace --skill '*' with --skill think-through in the repository installation command. Use /think-through in Claude Code/Cursor, or /workflows:think-through with the Claude plugin.
Coding workflows
In Codex, use the examples below. In Claude Code or Cursor, replace $workflow with /workflow.
| Task | Invocation | Runs |
|---|---|---|
| Small, clear change | $workflow quick Add JSON output to the export command | implement → verify |
| Discuss before coding | $workflow discuss Add bulk archive | grill-me → implement → verify |
| Larger feature | $workflow feature Add team invitations | grill-with-docs → to-spec → to-tickets → implement-spec → verify → ready PR |
| Build approved tickets | $workflow build Implement the spec at docs/specs/invitations.md | implement-spec → verify → ready PR |
| Check existing work | $workflow verify Check this branch against main and its spec | Verify and report findings; no automatic product fixes |
The first build runs Matt's setup for the client's tracker and domain docs. Planning preserves his decision and approval checkpoints.
Verification reuses or creates the client's verify-<app> skill, exercises acceptance criteria, reviews the final diff, and selects relevant extra checks. Build routes fix accepted findings and reverify. Missing tools or model diversity are reported, never counted as passes.
feature and build automatically push the task branch and create or update a PR, without a separate confirmation. The PR stays draft during implementation and becomes ready after verification passes. Say "local only" or "no PR" to opt out; unavailable Git hosting access is reported as a blocker.
You can keep using Matt's skills individually, then run $workflow verify. In other modes, request a PR or use $pr after proof. Use $retro after the session. Merging and deployment require separate authorization.
Native plugins, optional
Native manifests share the same catalog. Choose one installation method per runtime. Skills CLI installation is smoke-tested; native hosts have not all been exercised.
# Claude Code, after publication
claude plugin marketplace add kolodziejczakM/workflows
claude plugin install workflows@kolodziejczak-workflows
# Codex, after publication
codex plugin marketplace add kolodziejczakM/workflows
codex plugin add workflows@kolodziejczak-workflows
Claude plugins use /workflows:workflow quick ...; select the qualified skill in other plugin hosts. Multi-family review requires host support.
Maintain this bundle
npm ci installs development tools. npm test runs regressions; npm run check checks recorded snapshots and metadata offline. npm run check:upstream verifies against pinned GitHub trees. Update refs in upstreams.json, run npm run sync, and review the diff. npm run pack:plugin builds dist/workflows/.
See maintenance, source research, and third-party notices. The custom glue is MIT; imported skills retain their original licenses and credits.