AI Compass / Quick reads
Files first, then code
An AI guesses where nothing is written down. Five short files before the first command save a lot of corrections afterwards.
In depth: Project start with rules and plan files
Four basic rules, often attributed to Karpathy, belong according to AI LABS in the CLAUDE.md of a parent folder, not in a skill: state assumptions, simplest solution, change nothing outside the task, define a check before “done” (1 Sept). For a new project, the same channel recommends a few files before the first build: a requirements document from a round of questions, negative rules for what must never be built, a progress file and a lessons file, and tests derived from the requirements before the code (10 Apr). What must never happen belongs in a hook according to the docs, because the CLAUDE.md is context, not a block (docs checked).
Tool and alternative
Everything works with Claude Code and git; you don’t need the creators’ templates. The rules file lives in the repo multica-ai/andrej-karpathy-skills (MIT according to the README; the old address forrestchang/… redirects there, externally checked). Use the lessons skill task-observer only with many skills, as its own README says. According to the docs, a CLAUDE.md should stay under 200 lines.
Sources
- AI LABS, “These New Claude Skills Are Actually Insane”, 1 Sept 2026 (7:01–8:04)
- AI LABS, “Give Me 13mins, I Will Show You The Right Agent Setup”, 10 Apr 2026 (1:04–9:42)
- multica-ai/andrej-karpathy-skills, rebelytics/one-skill-to-rule-them-all (task-observer) and Claude Code docs Memory, retrieved 17 Oct 2026
Advertising in the source: Supabase (1 Sept), Wayin Video (10 Apr, 6:30–7:35), each plus paid community.
Want to try this with your team? That’s what the AI Lab is for. See the AI Lab