A minimalist fork of obra/superpowers by Jesse Vincent — rebuilt in July 2026 against verified, current harness and model behavior.
The original Superpowers is a complete software-development methodology for coding agents. This fork strips it to the behavioral discipline that still changes what an agent does, verifies every load-bearing assumption with dated research, and installs as user-level config for Claude Code, OpenAI Codex, and Cursor.
- Leverage the harness and the model. 2026 harnesses do planning, task tracking, subagent dispatch, worktrees, code review, hooks, and persistent memory natively; 2026 models plan and code well. Don't re-implement any of that. Keep only what changes agent behavior at decision points models still get wrong on their own.
- Aim for token efficiency. Usage cost is a first-class constraint. Prefer cheap, always-on guidance over machinery that spends tokens per step. Length is a cost.
- Work with the model's grain. Guidance that contradicts how a model was trained tends to degrade it. Add discipline that fills a real gap, not friction that fights the model.
- Evidence over vibes. Every piece of guidance traces to a dated, sourced claim in
ASSUMPTIONS.md. When a claim falls, the guidance it justifies changes — theself-updateskill re-verifies the registry as models and harnesses move.
AGENTS.md — the always-on operating guide. One essence per discipline, plus a
"→ load skill X" pointer to the fuller procedure. Read natively by Codex and Cursor; reaches
Claude Code via CLAUDE.md. This is the reliable layer — skills auto-trigger probabilistically,
essences are always in context. Final handoffs of distinct bodies of work end with a fixed
receipt — implementation, verification, independent review, and next step/options — so
completion state and the available exits are visible across harnesses without scrolling.
skills/ — ten, loaded on demand. Each exists to counter a failure mode current models
measurably still have:
| Skill | Counters |
|---|---|
| brainstorming | building before the problem is understood and the design agreed |
| research | planning against code the agent only knows by name |
| test-driven-development | implementation-first work, mock-heavy tests |
| systematic-debugging | patching symptoms; not stepping back after repeated failures |
| verification-before-completion | unverified — or gamed — success claims |
| receiving-code-review | sycophantic caving to feedback |
| coding-standards | type/naming/abstraction drift |
| model-selection | wrong model or reasoning effort for the task |
| agent-handover | context rot; dead ends at usage limits |
| self-update | this project itself going stale |
Personal setup never lives in published skill text: your model roster and your stack's coding
rules go in gitignored *.local.md files seeded from checked-in templates at install (see
below).
ASSUMPTIONS.md — the registry of empirical claims behind all of the
above, each with what it justifies, a last-verified date, and evidence in docs/research/.
git clone <this repo> && cd <repo> && ./install.sh| Surface | Claude Code | Codex CLI | Cursor |
|---|---|---|---|
| Always-on guide | ~/.claude/CLAUDE.md (managed block) |
~/.codex/AGENTS.md (managed block) |
Settings → Rules → User Rules (manual paste, prompted by the script) |
| Skills | ~/.claude/skills/ (symlinks) |
~/.agents/skills/ (symlinks) |
reads both trees automatically |
Skills are symlinked, so repo edits are live immediately; the instruction blocks are marker-managed,
so update = git pull && ./install.sh. For a repo that wants checked-in, team-visible
guidance instead: ./install.sh --project <dir>.
One post-install step: install.sh seeds skills/model-selection/roster.local.md from its
template — edit it with your harnesses/subscriptions, model sweet spots, and standing
assignments. It is gitignored (personal, never published) and reaches every harness through
the skill symlinks.
Coding standards are per-stack: published bases in skills/coding-standards/stacks/
(typescript.md, kotlin.md, …) load automatically for matching projects and update with
git pull. For personal additions or overrides, create a gitignored
stacks/<stack>.local.md next to the base — it wins on conflict and survives pulls.
Run the self-update skill quarterly or after major model/harness releases. It re-runs the
research that produced docs/research/ (prompt templates ship inside the skill), diffs the
findings against ASSUMPTIONS.md, and proposes the guidance changes that follow — with the same
keep/drop bar the fork was built on.
Rebuilt 2026-07-02 (the fable-powers rebuild): decision log in
docs/mission-log.md, original feature-by-feature keep/drop reasoning in
MINIMIZATION.md.
Superpowers was originally built by Jesse Vincent and the folks at Prime Radiant. Read the original release announcement. If Superpowers has helped you do things that make money, consider sponsoring the original author's open-source work.
MIT License — see LICENSE. Copyright © 2025 Jesse Vincent; portions © 2025–2026 Thiemo Belmega.