sharpened

Sharpened

Battle-tested skills library for AI coding agents. A fork of obra/superpowers (MIT, © Jesse Vincent) with AI-specific hardening learned from real production sessions.

Why fork

Superpowers gives coding agents a solid methodology: brainstorming → plan → subagent-driven execution → review → done, with TDD and YAGNI as principles. It works.

But it was written for humans doing TDD. In real production use with AI coding agents, we kept hitting failure modes the stock skills don't defend against:

1. Tautological assertions — expect(output).not.toContain(orchestratorId) passes trivially when the template never contains a UUID to begin with. Stock TDD's "watch it fail once" doesn't catch this because the test was always going to be green.
2. Spy at wrong layer — test asserts the idempotent guard in module A fires, but the actual guard is in module B. Removing guard A leaves the test green.
3. Reporting drift — agent claims "109 tests pass" when vitest actually reports 105. Claims "fixed" a field that never made it to schema. Deletes a test to stay green without saying so.
4. Scope creep — task says "extract a helper"; agent also fixes three unrelated bugs and refactors a sibling module. CI stays green, next PR becomes unreviewable.
5. Anti-patterns as a fixed list — stock skills ship with 5 anti-patterns frozen at author's last commit. Real projects need the list to grow every time a new failure mode appears.

Sharpened is the skill set that emerged after ~15 rounds of forcing fixes to these modes. The diffs vs. upstream aren't large; the philosophy is.

Core additions / changes vs. superpowers

Skill	Change
test-driven-development	Reverse-regression is mandatory: every new assertion must red-on-comment-out of the corresponding product-code line, green on restore. Integration tests forbid mock-pty-style chains. Anti-patterns list grows — project-local anti-patterns.md is versioned with the repo.
verification-before-completion	Numbers are verbatim to tool output. Deleted tests must be listed with reason. Completion is per-item verdict (done / partial / skipped + evidence), not a blanket "basically done".
scope-guardrails (new)	PR-X only does PR-X. Accidental fixes to unrelated issues go into a follow-up PR or get reverted. git status must match the task manifest.

Everything else from superpowers is preserved.

Install

Claude Code

# From GitHub (once plugin marketplace supports custom sources)
/plugin marketplace add tt-a1i/sharpened
/plugin install sharpened

# Or local dev
git clone https://github.com/tt-a1i/sharpened.git ~/.claude/plugins/cache/sharpened/1.0.0

Gemini CLI

gemini extensions install https://github.com/tt-a1i/sharpened

OpenCode / Codex

See .opencode/INSTALL.md for bootstrap injection (inherits superpowers' multi-CLI support).

Relationship to upstream

• License: MIT. Upstream copyright © Jesse Vincent; derivative work © Shaokun Tu.
• Attribution: every skill that is a modified version of an upstream skill retains a > Adapted from obra/superpowers note at top.
• Sync policy: we cherry-pick from upstream when skill content is updated; we do not auto-merge, because upstream's open PR backlog on TDD improvements suggests divergent quality bars.
• Issues / PRs back upstream: if a sharpened change is pure bug-fix (not an opinion), we file it upstream too.

Philosophy

Human engineering conventions assume code is read many times, developers have cross-session memory, and code review catches drift. None of that is true for AI coding agents across sessions.

Sharpened's thesis: the rules that matter are the ones machines can verify (grep, wc -l, reverse-regression). Soft guidance ("mocks only if unavoidable") becomes hard rules ("grep mock-node-pty tests/server/ returns zero"). The anti-patterns list grows every time an agent gets caught, and grows only. Nothing gets deleted because a future agent will always try the same trick again.

Sponsorship

Upstream author accepts sponsorship: obra on GitHub Sponsors. Sharpened has none.

License

MIT. See LICENSE.