You commented REPOS, so here is the list. All eight work with Claude Code, and most of them work with Codex and Cursor too — the useful ones are not tied to a single vendor, they are skills, rules and memory that sit next to whatever agent you happen to be running.
Star counts below were read from the GitHub API on 23 September 2026.
The short version: if you install exactly one of these, install superpowers. It changes how the agent works rather than what it knows, and every other repo on this list gets more useful once the agent is planning and testing properly.
| Repo | What it fixes | Stars |
|---|---|---|
| obra/superpowers | Agent skips planning and tests | ~290K |
| affaan-m/ECC | No memory, no safety rails, slow | ~266K |
| mksglu/context-mode | Tool output eats the context window | ~24K |
| max-sixty/worktrunk | Parallel agents fighting over one checkout | ~8K |
| stablyai/orca | No way to run a fleet of agents | ~76K |
| akitaonrails/ai-memory | Switching tools resets everything | ~8K |
| DietrichGebert/ponytail | Over-engineered solutions | ~145K |
| ayghri/i-have-adhd | The answer buried under filler | ~51K |
A skills framework and a development methodology in one. It gets the agent to plan before it writes, write the test first, and review its own work before it tells you it is done. That last part is the one that matters — the default failure mode of every coding agent is declaring victory on code it never ran.
Start here. It is the highest-leverage install on the list.
A harness optimization system: memory, instincts, security rails and research-first development bolted onto Claude Code, Codex, Opencode and Cursor. Where superpowers gives the agent a method, ECC gives it an environment — the things you would otherwise rebuild by hand in every repo.
github.com/mksglu/context-mode
Sandboxes large tool output — logs, test runs, diffs — so it does not spend your context window. This is the least glamorous repo here and the one you notice fastest. A single failing test suite can dump tens of thousands of tokens of stack traces into a session; three of those and your agent has forgotten what you asked for. Trimming the output at the tool boundary is what keeps long sessions useful.
It works through MCP plus hooks, so it applies across agents rather than inside one.
github.com/max-sixty/worktrunk
A CLI for git worktree management, built for parallel agent workflows. Give each agent its own branch and its own folder, and run several at once without them trampling each other's edits or colliding on merge.
Running more than one agent in a single checkout is the mistake everyone makes once. Worktrees are the boring, correct fix, and this wraps the ceremony.
An agent development environment for running a fleet of agents in parallel — desktop, mobile and a remote runtime, using your own subscription and whichever coding agent you prefer. Pairs naturally with worktrunk: worktrunk isolates the filesystem, orca gives you somewhere to watch all of it at once.
github.com/akitaonrails/ai-memory
Long-term memory for agent CLIs, explicitly designed for handoff between vendors. Everything you taught Claude Code about your codebase carries over when you open the same project in Codex or Cursor instead of starting from zero.
github.com/DietrichGebert/ponytail
Rules that steer the agent toward the simplest thing that works. The project describes it as making your agent think like the laziest senior dev in the room, which is exactly right: the best code is the code you never wrote.
Left alone, agents reach for abstraction — a factory where a function would do, a config layer for a value that changes once a year. This is the counterweight.
A skill that stops the agent burying the answer. Answer first, filler cut. Small, and it changes every single response you read for the rest of the day.
These are not eight separate installs so much as four layers. This is the order I would add them in:
ECC overlaps with a few of these and is worth taking as a whole rather than piecemeal. Ponytail and i-have-adhd are rule files — add them last, once the rest is working, and judge them on whether the output reads better.
Not one of these repos makes the model smarter. They fix the harness around it: what the agent can remember, how much of the window it wastes, whether two of them can run without colliding, and whether it checks its own work before claiming it is finished. That is where the 10x is — the model has been good enough for a while; the scaffolding around it is what has been thin.
If you want more in this shape, there is 25 projects to build to get hired in 2026, a Rubik's cube solver built with Claude and Three.js, and the gesture-controlled portfolio this blog lives on.
Follow @avi_vashishta29 for more setups like this.