garden
Garden
A single-shot autonomous maintenance agent. You invoke it, it finds one thing worth fixing, fixes it well, and opens a PR. Then it stops. Think of it as a gardener walking through the project, pulling one weed per visit.
Inspired by the "doc-gardening" and "garbage collection" concepts from agent-first engineering: steady, small-batch tending beats painful cleanup sprints.
Flow
- Setup -- enter a worktree on a
garden/-prefixed branch - Scan -- survey the project for maintenance issues
- Select -- pick the single most worthwhile issue to fix
- Fix -- make the change, commit it
- Self-review -- dispatch two competing sub-agents to review the fix
- Iterate -- if reviewers find problems, fix and re-review (max 3 rounds)
- Ship -- open a PR against the main branch
Step 1: Setup
Create a worktree and branch:
More from maragudk/fabrik
diary
Write and maintain an implementation diary capturing what changed, why, what worked, what failed (with exact errors and commands), what was tricky, and how to review and validate. Activates proactively during non-trivial implementation work (new features, bug fixes, refactors, research spikes) and at natural session-end moments -- after a PR merges, a feature ships, or a work chunk wraps up -- to capture the narrative while it's still fresh. Does not activate for trivial tasks like one-line fixes, config tweaks, or quick questions.
3decisions
Guide for recording significant architectural and design decisions in docs/decisions.md. Use this skill when clearly significant architectural decisions are made (database choices, frameworks, core design patterns) or when explicitly asked to document a decision. Also suggest proactively at natural session-end moments -- after a PR merges, a feature ships, or a work chunk wraps up -- if a significant decision was made during the session and not yet recorded. Be conservative - only suggest for major decisions, not minor implementation details.
3unsloth
Guide for fine-tuning LLMs, embedding models, vision-language models, and TTS models efficiently with Unsloth. Covers LoRA/QLoRA SFT, reinforcement learning (GRPO, DPO, ORPO, KTO), embedding fine-tuning with sentence-transformers, continued pretraining, and saving/exporting to GGUF, Ollama, or vLLM. Use this skill whenever the user mentions Unsloth, FastLanguageModel, FastSentenceTransformer, FastVisionModel, FastModel, or wants memory-efficient fine-tuning of open LLMs or embedding models on a single GPU, even if they don't explicitly say "Unsloth".
2modal
Guide for running Python code on Modal, the serverless compute platform for AI workloads, batch jobs, scheduled tasks, web endpoints, and sandboxed code execution. Use this skill whenever the user is writing or modifying Modal code (anything importing `modal`, decorating with `@app.function`, `@app.cls`, `@modal.fastapi_endpoint`, etc.), running `modal run`/`modal deploy`/`modal serve`, configuring GPUs/images/volumes/secrets for Modal, or asking how to host inference, fine-tuning, or agent sandboxes on Modal.
2sql
Guide for working with SQL queries, in particular for SQLite. Use this skill when writing SQL queries, analyzing database schemas, designing migrations, or working with SQLite-related code.
2git
Guide for using git with specific preferences -- branch names without `feat/`/`hotfix/` prefixes, backticks around code identifiers in commit messages, asking about GitHub issues to reference before committing. Use this whenever you branch, commit, or write a commit message -- not just when explicitly asked to "commit". These conventions aren't in your default knowledge and you'll get them wrong without consulting this skill.
2