What's this about?
Thariq (Claude Code core dev) distilled his prompting philosophy into one line: thin prompts, thick artifacts + context, thin skills. Don't write elaborate instructions — put the richness into the *materials* you give the agent (specs, docs, examples), and keep both the ask and the packaged skills minimal.
What exactly is an "artifact" here?
An artifact is a concrete document the agent can read: a design spec, a schema, a good example of desired output. The insight: models are now smart enough that long clever prompts mostly add noise; what they lack is *your* context. A thin prompt ("implement this spec") pointing at a thick artifact (a great spec) beats a thick prompt every time.
What's the community saying?
1.6k likes from a core dev of the most-used coding agent
- Widely screenshot as the counter-manifesto to "prompt engineering" maximalism.
- Dovetails with emollick's same-day warning that elaborate viral prompt files are self-defeating slop.