Anthropic publishes Skill authoring best practices: keep SKILL.md under 500 lines
bibryam · x · 2026-11-01
Anthropic released official best practices for writing Agent Skills in its Claude Platform docs, centered on conciseness: the context window is a shared resource, and only Skill metadata is pre-loaded at startup.
Key points:
- Assume Claude is already smart; add only context the model lacks, making every token justify its cost
- Tune freedom by task fragility: heuristics for open-ended tasks, strict checklist scripts for fragile ones
- Write precise descriptions so Claude can discover when to use each Skill
- Use progressive disclosure: a concise SKILL.md entry point with references and scripts loaded on demand
- Keep SKILL.md under 500 lines and references no more than one directory level deep
- Build evals first and iterate, with at least three evaluations before extensive docs
More from coding & agent
- Stardock's GalCiv IV dev shows how AI agents kill 5:1 asset-pipeline overhead and finally clear years of player requests — draginol · 2026-11-01
- Dev uses AI coding to turn a handheld into an e-book reader and drawing tablet — banteg · 2026-11-01
- ASC CLI adds territory-scoped PPP pricing, deep IAP validation, serial matching — rudrank · 2026-11-01
- asc versions release Now Reconciles After Server Errors, So Shipped Releases No Longer Turn CI Red — rudrank · 2026-11-01
- asc CLI now reconciles release status after server errors to avoid false CI failures — rudrank · 2026-11-01
- asc CLI adds terminal-based streamlined purchasing toggle and StoreKit ownership filtering — rudrank · 2026-11-01