Plan, test, debug, and ship with more care in Claude Code and Codex.
One plugin with task skills. Optional helpers and style stay in repository-only extras.
Install · What changes · Why megapowers · Skills · Docs
Coding agents are fast, but they can skip steps: edit before they know the cause, say "done" without re-running the tests, or push and deploy without a second look. megapowers gives Claude Code and Codex a skill (a short playbook) for each of those moments. When your request matches one, the agent can load it and follow it.
It is exactly one plugin, with tested installs for Claude Code and Codex. The harness-neutral skills and their supporting files may also be copied into other Agent Skills harnesses. Local agent bindings and hooks stay in the user's setup; copying skills does not install a plugin or the repository-only output style. It uses each tool's own agents, permissions, worktrees, and memory, and ships no daemon, scheduler, model router, formatter, or status line.
| You ask | What the agent does |
|---|---|
| "Why is this test flaky?" | systematic-debugging: finds the root cause before editing anything |
| "Add rate limiting to the API" | test-first-implementation: writes a failing test, sees it fail, then makes it pass |
| "Commit this and open a PR" | verify-and-finish: re-runs your checks and reports what actually passed |
| "Deploy to production" | safe-effects: confirms the exact target with you, then reads the result back |
| "Get a second opinion on this auth change" | independent-review: guides review by another available model, following your reviewer instructions |
Skills guide the agent; they do not force it (how often they load). Your repository's own instructions and tools always come first.
Claude Code
claude plugin marketplace add lawzava/megapowers@release
claude plugin install megapowers@megapowersCodex
codex plugin marketplace add lawzava/megapowers --ref release
codex plugin add megapowers@megapowersThen start a fresh session. Loading the shipped skills needs no Go runtime.
Both plugin installs track the release branch, which only moves to signed
release tags. Pinning, install scopes, and uninstall are in
docs/install.md.
OpenCode uses a copy/path install or HTTP skill catalog, not a plugin install.
The primary route is the release catalog. Add it to the skills array in
opencode.json:
{
"$schema": "https://opencode.ai/config.json",
"skills": [
"https://raw.githubusercontent.com/lawzava/megapowers/release/plugins/megapowers/skills/"
]
}Updates arrive when the release branch moves. Each catalog entry uses the
plugin version so changed releases refresh OpenCode's cached skill files.
Alternatives:
- Point the same array at a checkout's
plugins/megapowers/skillsdirectory, for example"/path/to/megapowers/plugins/megapowers/skills". Absolute paths are used as written; relative paths resolve from the active OpenCode working directory, not the config file. - Copy skills with
npx skills add lawzava/megapowers -a opencode.
Use one route to avoid duplicate skill IDs. These routes load skill text and
supporting files, not the repository-only Claude Code output style. The
OpenCode skills documentation defines the
skills array and HTTP catalog format.
- Shared skills on Claude Code and Codex. One plugin supplies the same task guidance on both.
- Checks before you ship. Verification and side-effect skills guide the agent to check evidence and authority before acting.
- Tested with and without the plugin. Installed-plugin A/B studies compare the same tasks with and without megapowers. Historical totals need revalidation after a grader correction. Results below.
- Independent model review.
independent-reviewguides the agent to choose another available model, following your instructions. Cross-vendor review is preferred but optional. The skill supplies no provider launcher. - Loads on demand. The catalog supplies trigger descriptions. A skill's full text loads only when you need it.
Facts checked on 2026-09-25. Install counts on skills.sh are self-reported by that site.
| megapowers | obra/superpowers | mattpocock/skills | |
|---|---|---|---|
| Works with | Claude Code and Codex, one plugin, shared skills | 16 agent tools, including Claude Code, Codex, Cursor, Gemini CLI, and Copilot CLI | Any skills-compatible client through npx skills add |
| Get it from | This repository's marketplace | The official Claude Code marketplace and, per its README, the official Codex marketplace | skills.sh, 4.2M installs across the collection |
| What you get | Task skills; optional helpers and Claude Code style in repository-only extras | 14 skills, including brainstorming, writing-plans, executing-plans, subagent-driven-development, TDD, and diagnosing-superpowers |
Single-purpose skills: code-review, tdd, triage, diagnosing-bugs, handoff, writing-for-agents, grill-me (1.2M installs) |
| Published measurements | Pack comparison and A/B studies, with SHA-pinned packs | Per-change probe counts in release notes (for example, TDD control 8/10 against treatment 5/10 when a section was cut) and a quorum eval lab | None found in the repository on 2026-09-25 |
Which one fits:
- superpowers if you use other agent tools such as Cursor, Gemini CLI, or Copilot CLI, or want commercial support. It has official marketplace listings on both sides, release notes that quote probe counts per change, and a diagnostics skill since v6.4.1 (2026-09-19).
- mattpocock/skills if you want to adopt small single-purpose skills one
at a time through skills.sh. Its
grill-meis a user-invoked pointer skill, which matches the Claude Code docs' advice for skills that over-trigger. - megapowers if you use Claude Code, Codex, or both, and want one plugin with shared task guidance and published with-and-without measurements.
Each skill has a short description the agent always sees. The full text loads
only when your request matches. The experimental labels come from
skills/catalog.json.
Plan and build
| Skill | Use it when |
|---|---|
design-and-plan |
A feature or interface needs a spec, a tradeoff decision, or a step-by-step plan |
close-off (experimental) |
You invoke it by name at the end of a session: it checks the task is complete, records learnings in the project's own surface, cleans up what the session started, and says it is ready for archiving |
grill-me (experimental) |
You invoke it by name to stress-test a plan or idea through a structured interview before any action |
test-first-implementation |
Adding or changing behavior, fixing a confirmed bug, or refactoring: red, green, refactor, with each run observed |
systematic-debugging |
A bug, flaky test, incident, or regression has an unknown cause |
orchestrating |
Independent tasks can run in parallel on the tool's own subagents, each owning different files |
Ship safely
| Skill | Use it when |
|---|---|
verify-and-finish |
Before saying done, committing, merging, publishing, deploying, or handing off: re-runs the real check |
safe-effects |
Before a deploy, message, charge, migration, destructive query, DNS change, or any write that leaves your machine |
independent-review |
Security, auth, billing, concurrency, or data-integrity work needs an independent opinion from another model |
Research, long runs, and writing
| Skill | Use it when |
|---|---|
evidence-research (experimental) |
A decision needs evidence from outside the repository |
autonomous-run (experimental) |
An approved goal must keep going unattended across context resets or sessions |
humanizing-prose |
You want human-facing text drafted or edited without filler, and without dropping or inventing facts |
writing-agent-instructions (experimental) |
You are writing or revising a skill, AGENTS.md, or CLAUDE.md, running a retrospective from the current conversation, or auditing saved memory |
Repository-only setup and upkeep skills are in extras/; they are not installed with the plugin.
docs/orchestration.md explains which kinds of task go to which skill.
An optional repository-only style makes replies shorter and more direct: the answer first, about 100 words by default and up to 250 unless you ask for depth, and no em dashes. It keeps the built-in coding instructions.
- Claude Code: install it manually from
extras/output-styles/ into
~/.claude/output-styles/, then run/output-style Megapowersor selectMegapowersin/config. The plugin no longer ships it. - Codex: the plugin does not apply a session-wide style. Invoke
humanizing-prosefor writing tasks that need its editing guidance.
In one operator's September 2026 Codex sessions, styled final messages had a median of 70 words, and 0.6 percent contained an em dash, against 36 percent without the style. These numbers are not a token-saving or cost claim.
Each pack ran on the same 8 cases, 3 runs per case, with and without the
pack, in isolated OpenCode sessions. A blind judge from the other vendor
graded each run: Sol graded Opus runs and Opus graded Sol runs. The 8 cases are
the ones where some run failed in an earlier 20-case baseline. Opus is
claude-agent/opus; Sol is openai/gpt-6.1-sol-fast#high. Packs were pinned
to these commits: megapowers b75e795, mattpocock/skills 4ae2317,
obra/superpowers 22f0b46, pstack (cursor/plugins) 6b03002,
addyosmani/agent-skills 301b02b.
| Pack | Opus without → with (of 24) | Sol without → with (of 24) | Time with vs without (Opus / Sol) |
|---|---|---|---|
| megapowers | 18 → 22 | 23 → 24 | −5% / +6% |
| mattpocock/skills | 17 → 22 | 24 → 23 | +6% / +19% |
| superpowers | 19 → 21 | 23 → 21 | +14% / +24% |
| pstack | 18 → 19 | 24 → 22 | −2% / +2% |
| addyosmani/agent-skills | 20 → 18 | 24 → 24 | +5% / +25% |
Not every row comes from a fresh run. The megapowers, mattpocock/skills and superpowers Opus rows are the same 8 cases taken from earlier 20-case runs. Those runs used the same skill content as listed above, and their saved runs were graded again with the current grader. The pstack, addyosmani/agent-skills and superpowers Sol rows ran fresh on the 8 cases.
Opus gained 4 to 5 passes with megapowers and with mattpocock/skills. This small, selected sample does not establish reliable differences between packs. Sol passes almost every case without any pack. On the full 20 cases, megapowers scored 51 → 55 of 60 on Opus and 56 → 57 of 60 on Sol. Method, cases and commands: installed-ab study.
The October 2026 A/B studies reported gains in judgment cases and ties on their coding tasks. Their totals have not been revalidated after an October 12 grader correction, and the original result bundles are unavailable locally. They do not establish current efficacy or a speed or token-cost benefit. Historical results and limitations are retained in evals/RESULTS.md.
The evaluation suite includes a three-turn maintenance case that checks whether later changes preserve earlier behavior. The scorer labels effort levels and reports paired actor durations beside outcomes. The study protocol explains how to run matched effort comparisons. Having these checks is not evidence that the current plugin improves their results.
The study below measured earlier models with graders since found to fail correct answers.
From an installed-plugin A/B study completed 2026-09-05: 1,080 valid trials
over 27 cases, with ten control/treatment pairs per case on each tool. The
control ran without the plugin and the treatment ran with it. Codex ran
gpt-6-astra on CLI 0.153.3, and Claude ran claude-fable-5-1 on Claude
Code 2.1.258, both at high effort.
| Task checks passed | Codex without | Codex with | Claude without | Claude with |
|---|---|---|---|---|
| 18 development cases | 85/180 | 106/180 | 89/180 | 101/180 |
| 6 held-out cases, frozen before the run | 50/60 | 60/60 | 52/60 | 53/60 |
With the plugin installed, the skill a task needed loaded in 77 of 150 Codex
trials and 83 of 150 Claude trials, so skills help but do not always fire.
The study predates the 0.29.0 version stamp. Neither tool passes the
full benchmark, and the results do not show general model superiority. Full
results and method: evals/RESULTS.md.
More results and caveats
| Development checks (18 cases) | Codex control | Codex + megapowers | Claude control | Claude + megapowers |
|---|---|---|---|---|
| Task outcome | 85/180 | 106/180 | 89/180 | 101/180 |
| Workflow | 163/180 | 170/180 | 164/180 | 169/180 |
| Outcome and exact activation profile | 85/180 | 74/180 | 89/180 | 73/180 |
| Required activation profile | n/a | 77/150 | n/a | 83/150 |
"Required activation profile" scores whether the skill a case requires
actually loaded, as shown in the trace. It is scored apart from task outcome.
"Outcome and exact activation profile" also required the exact expected set of
skill reads, so an extra successful skill read could fail it. On this row
treatment scores below control on both tools. The current runner permits
successful reads of other known skills unless explicitly forbidden; these
historical totals have not been regraded under that policy.
All ten Claude pending-review replies passed the task checks but did not load
verify-and-finish. The instruction-authoring case moved from 0/10 to 10/10
on Codex and 0/10 to 9/10 on Claude. Codex's pending-review case moved from
0/10 to 10/10. The separate autonomy/continuity diagnostics did not pass
(Codex 0/30 to 0/30, Claude 1/30 to 0/30).
The 2026-09-02 Codex trigger-recall run reported 0/117 implicit recall despite
a rendered catalog
(trigger-recall README), so on
Codex the skill-trigger checks only report and never fail a run.
scripts/validate.sh and evals/run-all.sh prove repository mechanics, not
agent quality. Protocols are in docs/advanced/evals.md.
The skill catalog loads per session; each skill body loads only when that skill is used. Earlier token estimates included startup text that is no longer shipped and do not describe this version.
Exactly what installs
- 13 skills under
skills/, listed inskills/catalog.json. - An HTTP skill catalog at
skills/index.jsonfor OpenCode. - Guidance and passive metadata only: no executable code, hooks, or MCP entries. Optional helpers and setup skills live in repository-only extras/.
The plugin does not edit global user configuration and supplies no model,
provider, or agent configuration. For non-trivial delegation, orchestrating
can read an optional personal registry at
~/.config/megapowers/agent-capabilities.md; it is advisory and grants no
access.
Check registration with the plugin-list commands in docs/install.md, then ask the agent to load a named skill and summarize it. For optional local diagnostics, see extras/.
| Symptom | Fix |
|---|---|
Old megapowers:Megapowers style setting no longer resolves |
Install the repository-only style manually from extras, then select Megapowers |
| Wrong version after an upgrade | Compare the marketplace head with the approved release tag before registering it |
- Skills are instructions, not enforcement. Model and tool behavior change.
- The repository-only Claude Code style cannot suppress tool-result rendering.
- The plugin supplies no destructive-command guard or command gates. Use harness permissions, OS sandboxing, and least privilege for enforcement. See SECURITY.md.
- Independent review guidance requires checking secrets and disclosure scope before sending material to the selected reviewer.
- The evaluation studies that call real models cost money and need a reviewed credential broker. Their self-tests do not replace a real run.
- Claude Code and Codex are the tested plugin targets; OpenCode uses skill sources. Exact boundaries are in docs/harness-support.md.
- Install, pin, update, uninstall
- Harness support and freshness
- Orchestration and task shapes
- Independent review workflow
- Evaluation and release evidence
- Security policy, Contributing, Changelog
scripts/validate.sh
bash evals/run-all.sh --json results.jsonlThe deterministic suite fails on malformed, incomplete, indeterminate, timed-out, or harness-error results. Changes to behavioral guidance need deterministic regression coverage. See CONTRIBUTING.md.
megapowers is MIT-licensed. It began from Superpowers by Jesse Vincent and retains other source credits in ATTRIBUTION.md. It is not an Anthropic or OpenAI product.