Anthropic Just Revealed 10 NEW Rules for Claude Skills

Summary: 10 New Rules for Agent Skills

1. Progressive disclosure, one level deep. Keep SKILL.md under 500 lines and treat it as a contents page that links straight to supporting files (pricing, clients, scripts). Claude opens a file only at the step that needs it. Avoid chains ( SKILL.md → file B → file C): with nested references, Claude may only preview the deeper file, reading roughly its first 100 lines.
2. Content lists on long files. Any file over 100 lines gets a short table of contents at the top, so even a partial read tells Claude what's further down and where to jump.
3. Set the degrees of freedom per step. High : plain instructions, for open tasks like brainstorming. Medium : a template with some settings, for shaped output like a weekly report. Low : an exact script, for risky work like invoices or tax documents. One skill can mix all three: write the email freely, generate the invoice by script.
4. Test on the models you'll actually use. One question per tier. Haiku: is there enough guidance? Sonnet: is it clear and efficient? Opus and Fable: does it avoid over-explaining? Old skills may need less instruction now, and a prescriptive skill may run fine on a cheaper model. Testing everywhere is expensive, so prioritize the skills you rely on most and the ones where mistakes cost the most.
5. Write only what Claude can't work out. Anthropic calls the context window a public good. Skip explaining what an invoice is; include your prices, terms, and internal rules. Write the description in third person ("Creates client invoices and sends payment reminders", not "I can help you with…"), since that's the text Claude reads when choosing a skill.
6. Workflows and checklists. For multi-step jobs, give an explicit checklist Claude copies into its reply and ticks off as it goes. 5.5 models have been seen skipping steps. Add go-back lines ("if any total doesn't match the job list, go back to step 2") so a ticked box can't wave through a wrong result.
7. Feedback loops. Run, check, fix, and repeat until the output passes. This isn't only for code. Anthropic's example is a style guide: draft, review against a short checklist (consistent terminology, examples in the standard format, required sections present), note each failure against the rule it breaks, revise, and finish only when everything passes.
8. Three patterns for defined output. Templates , either strict or a sensible default the model can adapt. Examples , as input → output pairs, which show style and detail better than description does. Conditional workflows , a fork that tells Claude which steps to follow under which conditions.
9. Build to be shared. Don't assume a teammate has your tools installed. Name the exact packages the skill uses, and include install steps for when they're missing.
10. Move must-never-break rules into hooks. A rule in capital letters is followed most of the time. A hook is code that runs at a set moment (before a command, before sending) whether or not Claude follows the skill. A hook can be declared inside the skill itself: Anthropic's example is a "secure operations" skill whose hook runs a security check before every command, and stays active for the rest of the session once the skill is used. Pick the few rules where one mistake would really cost you, and check whether each can become a hook.

Transcript

{
  "name": "transcript",
  "attributes": {}
}