agent-toolkit
Minimalistic skills and rules for AI coding agents that assist your daily work in any software engineering project
This toolkit is entirely open source and free to use. Give us a star on GitHub to support us.
Different projects, same repetitive tasks
- Every project is a different world tech stacks, trackers, conventions, AI adoption
- Yet the daily tasks repeat everywhere fetch a ticket, refine it, plan, review, reply
- agent-toolkit lets AI assist exactly those generic skills that can be used in any project
"I kept rebuilding the same AI setup in every project, so I built one that works everywhere."
Core ideas behind every skill
A minimalistic toolkit, not a framework
There are great skill frameworks out there that auto-activate around everything you do. The agent-toolkit deliberately takes the other road:
- a whole methodology
- skills activate before any task
- the agent drives
- pulls you in at its checkpoints
- fuller context, less to remember
- modular, use only the parts you need
- primarily user invoked
- you drive, the agent assists
- human in the loop by design
- lean context, predictable behavior
Refine, Plan, Act
A development workflow suitable for any kind of project
Help on both sides of the code review
Code review is still a key part of most teams' workflow. These skills assist in both directions: when others leave feedback on your PRs, and when you review someone else's code.
- fetch-pr-review downloads all the comments your PR received into a self-contained document that refine-pr-review can consume
- refine-pr-review assists you in triaging the feedback, comment by comment: address, partial or push back
- review-code-assistant reviews someone else's PR locally (works for self-review too), suggesting human-voiced comments and explanations. You decide what to post
Let a sub-agent review the code
A fresh perspective works for AI just like it does for humans: a sub-agent with a clean context, seeing only the changeset and a minimal description, catches surprisingly more regressions and issues than the session that wrote the code.
Your context is often cluttered before you even type
- context-checkup audits everything that auto-loads at startup (governing docs, skills, MCP servers), measures the token cost and proposes lean, reversible trims
- memory-doctor drains the agent's auto-memory block by block, relocating each entry to a home you control or archiving it. You confirm every action
Create and continuously improve the skills and docs your agents rely on
- compact-docs-writer writes docs with maximum token economy
- compact-skill-creator creates or edits skills, combining compact-docs-writer with a set of best practices for effective skills
- self-improve when the agent makes a mistake or doesn't behave the way you want, this skill captures the lesson so it won't repeat the same mistake again
Texts that sound like a real human typed them
We often ask AI to help draft texts that other people will read:
- use-conversational-language tells the agent to write in simple, human language instead of sophisticated AI prose full of — em dashes — and fancy terms
- write-realistic-texts opt-in rule that applies the human voice automatically whenever a text is meant for other humans (code comments, PR replies, chat messages…)
Optional and not installed by default: get them all with
./install-opinionated-rules.sh
or pick only the ones you want. Examples:
- git-read-only-by-default prevents the agent from performing git write operations such as commit, push, reset, etc. unless explicitly instructed
- no-ai-attribution no AI co-author additions on commits and no "Generated with" footers on PRs
- no-nonsense-comments write only code comments that still make sense to a future reader with zero context, prefer no comment over a low-value one, and voice them via use-conversational-language
- and more…