Obra/Superpowers, but Optimized.
Agentic skills framework & software development methodology.
2 followers
Agentic skills framework & software development methodology.
2 followers
Optimized fork of Superpowers ā Retains full original features while adding automatic 3-tier workflow routing, session memory system, integrated safety guards (OWASP-aligned), red-team adversarial testing with auto-fix pipeline, and an estimated 15-30% session overhead reduction. Delivers faster, more reliable, hallucination-resistant coding sessions. - REPOZY/superpowers-optimized
Dear members,
I've been using obra's Superpowers plugin daily for months. It's genuinely great.
But some things kept bothering me:
1. The AI has no memory. Every session it re-explores the same codebase,
re-learns the same constraints, re-proposes the same rejected approaches.
You end up re-explaining context you already explained last week.
2. There are no safety rails. One misread instruction and the agent runs
rm -rf or commits a hardcoded API key. It happens.
3. Sessions get bloated. The longer the context, the worse the reasoning.
Research actually quantifies this ā agents with longer context files use
14ā22% more reasoning tokens and perform worse, not better.
So I built a fork that fixes all of these things and much much more, while keeping the exact workflow everyone already likes.
What I added:
A four-file memory system that lives at your project root ā no database,
no embeddings, no external services. Just markdown files Claude can read
and write like any other file:
session-log.md ā the stop hook writes a breadcrumb at the end of every session automatically.
Skills used, files changed. If you explicitly save state, it records the full decision log:
what was tried, what was rejected, and why.
project-map.md ā generate once with "map this project".
The AI reads it at every session start instead of re-globbing.
Staleness is automatic via git hash comparison.
known-issues.md ā maps recurring errors to their solutions.
Consulted before every debugging session.
Never investigate the same bug twice.
state.md ā for tasks that span sessions. Captures where you were when you stopped.
The idea came directly from Jesse Vincent's episodic memory work for Claude:
https://blog.fsck.com/2025/10/23/episodic-memory
This is the lightweight, dependency-free version of that concept.
For safety I added 8 hooks that run silently in the background:
ā¢ 30+ dangerous bash command patterns blocked before execution
ā¢ 50+ file and content patterns protecting secrets from accidental exposure or commit,
ā¢ A subagent guard that catches when a subagent goes off-track and forces a redo.
For reasoning quality, I implemented Wang et al.'s self-consistency technique (ICLR 2023) at the two moments it matters most: root cause diagnosis and completion verification.
RESULT: Instead of committing to the first hypothesis, the agent generates 3ā5 independent reasoning paths and takes a majority vote.
If they don't agree, it stops and gathers more evidence before touching code.
This one change alone catches a surprising number of confident-but-wrong diagnoses.
The workflow routing is automatic:
ā¢ Tiny tasks (fix typo, rename variable) bypass everything ā zero overhead.
ā¢ Medium tasks get a single verification gate.
ā¢ Only complex tasks engage the full pipeline.
The ratio of overhead to value stays proportional regardless of what you're working on.
The numbers:
- 20 skills covering the full dev lifecycle
- 8 hooks running automatically on every session
- 2 specialized agents (code reviewer + adversarial red team)
- ~15ā30% estimated reduction in session token overhead
- Supports Claude Code, Cursor, Codex, OpenCode, and Gemini CLI
------------------------------------------------
It's free, MIT, and takes 2 commands to install
Claude Code: /plugin marketplace add REPOZY/superpowers-optimized
/plugin install superpowers-optimized@superpowers-optimized
Cursor: /plugin-add superpowers-optimized
For Codex, OpenCode, Gemini CLI ā tell the agent to fetch and follow the INSTALL.md from the repo.
If you're already using obra's original Superpowers, this is a drop-in replacement.
If you're not using any plugin, this is worth trying ā the memory system alone changes how sessions feel after the first few days.
These optimizations were done in the sole intend to ameliorate the plugin. I hope it helps everyone using it as much as it helps me during my coding sessions.
If you appreciate the effort and work, please don't forget to star the repository, it helps a lot!
https://github.com/REPOZY/superpowers-optimized
Shout-out to Jesse, the creator of the original Superpowers plugin.
https://github.com/obra/superpowers
Every dev should try it out, it's free, open source and plug & play.