Repeatable procedures your AI agent runs with judgment. Every run writes a trace. The agent gets better over time.
An agent is a folder of markdown files that an AI runtime follows. It has steps, a setup command, and a runs/ folder where every execution writes its trace. The agent reads past runs before it starts a new one, so it makes better decisions each time.
This is not a prompt chain or a static runbook. The steps give the agent room for judgment: they say what to accomplish and when the step is done, not which API call to make. The result is a procedure that works across different systems because the agent adapts it.
Every agent declares two kinds of inputs:
The first run starts from the procedure alone. It operates, observes, and writes what happened into its runs/ folder. The second run reads that trace before it begins. If the first run discovered a selector or a retry pattern, the second run knows about it and applies it.
Playbooks, decision logs, and checklists accumulate inside the agent's folder. The agent consults them before every step. Nothing is retrained, no model is fine-tuned. The context is just files, and the agent reads them.
Every agent follows the same structure. A manifest, numbered steps, a place for runs, and optional commands and functions.
Manifest with metadata, description, parameters, and connection requirements.
Numbered markdown files the agent follows, each with a clear objective and done condition.
Traces from completed runs. The agent reads past runs to make better decisions.
Setup interview and user-invokable actions. The setup command personalizes the agent.
Functions, playbooks, checklists, templates. Add whatever folders the agent needs. The structure is yours.
One command. The agent lands in your instance's modules/ folder, ready for the setup interview.
Every agent is a plain folder of markdown files. Download it from the agent page and drop it into any instance. No lock-in, no runtime dependency.
A browser agent that learns your sites. It performs browser tasks through your real Chrome, accumulates per-site knowledge, and crystallizes repetitive tasks into plain Python scripts; on later runs the AI only dispatches the script and heals failures. Broken knowledge heals itself; blockers like captchas escalate to you.
A diagram agent that turns an initial text into a polished explanatory visual, hand-crafted as SVG. It scopes the single claim the diagram must make, proposes 2-3 ASCII wireframe compositions built from proven archetypes, and expands the chosen one using a written design system so results stay consistent across runs. The user decides composition; the design system decides style. Finished diagrams accumulate in a library the agent consults on later runs.
A marketing post agent that puts the headline first. Each post starts from a chosen angle, then a headline workshop generates candidates from proven persuasion formulas (Cialdini, Kahneman, Caples), scores them, red-teams the best, and lets the user pick or rewrite. The body follows in two layers: a System 1 hook to stop the scroll, then System 2 proof to close the deal. The user controls every content decision. Product knowledge accumulates in a shared products/ folder so repeat runs stop re-asking.