1. Before any code: the agent writes skill.md
Plan, chosen architecture (and rejected ones, with reasons), tool stack with pinned versions, source paths, accounts, red lines. If it's not in the file, it wasn't decided.
METHODOLOGY
Skill-Centric Workflow
One file is the project's brain. Agents come and go; skill.md stays.
Context windows end. Sessions crash. Models get swapped. Every time that happens, a naked AI project loses its memory — and you pay to re-teach it. The Skill-Centric Workflow fixes this with one move: the project's entire state lives in a single, always-updated file named skill.md.
It is not documentation written after the fact. It is the working memory the agent itself maintains, and the first thing the next agent reads.
2. During the work: the file is kept alive
Every pitfall gets appended with a date. Every done item gets checked only when verified. When reality breaks the plan, the plan in the file changes — not just the code.
3. At handoff: any agent continues seamlessly
Token budget exhausted? Different model tomorrow? The next agent reads skill.md and picks up mid-stride — no archaeology, no re-explaining, no repeated mistakes.
This site is the proof
antist.ai is built and operated this way. Its own skill.md carried the project across sessions, caught a production outage's lessons, and briefed every agent that touched it. The scenario guides on this site each ship a starter skill.md — take one.
Starter templates
- equip-agent.skill.md ↓equip a coding agent →
- faceless-videos.skill.md ↓make faceless YouTube videos →
- solo-website.skill.md ↓ship a website solo →
Original concept & site by Antisubmissivist@gmail.com