dapperpowers is a distilled brainstorm -> plan -> implement -> verify development flow for Claude Code and Codex. It is derived from the superpowers plugin, but keeps the process lighter and more explicit. The shared skills define the workflow, while each host gets its own thin packaging and agent adapter.
Status: v0.4.3, under active development. Interfaces and skill names may still change.
The skills/ directory is the shared source of truth.
Claude Code loads the shared skills through .claude-plugin/plugin.json and the Claude agent definitions in agents/*.md.
Codex loads the shared skills through .codex-plugin/plugin.json.
The .codex/agents/*.toml files are optional project-scoped adapters for this repo or for projects that copy them in.
Codex plugin installs should still work without those custom agents because the implement skill can fall back to built-in Codex agents with the rendered role briefs.
Full read-only sandbox enforcement for Codex reviewer roles requires the optional .codex/agents/*.toml adapters.
Without those adapters, the fallback reviewers are instructed to stay read-only by prompt.
The agent tier names in plans are portable labels.
Claude resolves implementer, implementer-lite, task-reviewer, branch-reviewer, and escalation-arbiter through agents/*.md.
Codex prefers dapperpowers-implementer, dapperpowers-implementer-lite, dapperpowers-task-reviewer, dapperpowers-branch-reviewer, and dapperpowers-escalation-arbiter when those project agents are available.
When they are not available, Codex should dispatch built-in agents with the same rendered brief and role contract.
The recurring orchestration session should run on the strongest subscription or default coding model you normally use. Reach for stronger or metered models only as bounded oracle passes fed a compiled dossier. The loop holder owns brainstorming, planning, adjudication, and fix re-checks. Oracle passes judge a focused dossier and return, rather than remaining resident in the main loop.
Claude agent frontmatter can pin model and effort for dispatched agents. Codex custom agents inherit the active session model unless their TOML sets a model. The Codex adapter sets reasoning effort for each role and leaves model choice to the active Codex environment.
Claude Code CLI 2.1.x or later for Claude usage. Codex with plugin support for Codex usage.
From GitHub:
claude plugin marketplace add mattmarcin/dapperpowers
claude plugin install dapperpowers@dapperpowers-devFrom a local checkout:
claude plugin marketplace add <path-to-this-repo>
claude plugin install dapperpowers@dapperpowers-devUpdate an installed Claude copy:
claude plugin update dapperpowersFrom a local checkout:
codex plugin marketplace add <path-to-this-repo>
codex plugin add dapperpowers@dapperpowers-devFrom a Git marketplace:
codex plugin marketplace add mattmarcin/dapperpowers
codex plugin add dapperpowers@dapperpowers-devRefresh a Git-backed Codex marketplace:
codex plugin marketplace upgrade dapperpowers-dev
codex plugin add dapperpowers@dapperpowers-devFor a local checkout, pull or edit the repo, bump the synchronized plugin version or apply a Codex cachebuster during local iteration, then reinstall from the local marketplace:
codex plugin add dapperpowers@dapperpowers-devStart a new Codex thread after reinstalling so new plugin metadata, skills, and agent files are picked up.
Restart Claude after claude plugin update because Claude applies plugin updates on restart.
| Skill | What it does |
|---|---|
brainstorm |
Turns an idea into an approved spec through one-question-at-a-time dialogue, 2-3 approaches with a recommendation, and one approval gate. |
plan |
Turns a spec into an executable plan with global constraints, 3-6 task blocks, interfaces, acceptance criteria, tests, verify commands, and agent tiers. |
implement |
Orchestrates execution with a fresh implementer and reviewer per task, fixes via continuation when supported, committed plan checkboxes as the ledger, and one whole-branch review at the end. |
verify |
Requires fresh evidence before claims, including a spec checklist, command output, end-to-end exercise, red-green regression checks, and diff checks for delegated work. |
debug |
Applies systematic debugging when a test fails unexpectedly or a fix does not hold. |
- Few large tasks. A task is the smallest independently-verifiable deliverable and the largest unit an implementer can finish without context degrading.
- Spec-level plans. Plans carry acceptance criteria and interfaces, never implementation code.
- Fixes continue the implementer when possible. If the host cannot continue the same agent, the fix agent gets the original brief, report, diff, and accepted findings.
- Never on main. Implementation runs on a feature branch, and the branch review diffs against the recorded merge-base.
- Bounded oracle passes. The strongest available model is consulted only for focused decisions backed by a compiled dossier.
- Evidence before claims. No completion claim without fresh verification output.
Run the dual-package consistency check:
python scripts/validate_dual_plugin.pyValidate the Claude plugin:
claude plugin validate .Validate the Codex plugin shape:
python <codex-plugin-creator-skill>/scripts/validate_plugin.py .List the Codex marketplace entry:
codex plugin list --marketplace dapperpowers-dev --available --jsonFull design rationale and build history live in PLAN.md.
MIT - see LICENSE.