Transcript of Configuration and multi-file skills
Claude
0:02A basic skill works with just a name and0:04description, but here are some other0:06advanced tips that can make your skills0:08really effective in claw code.0:12The agentskills.io open standard has0:15many available fields. We already went0:17over the name, which identifies your0:20skill, uses lowercase letters, numbers,0:22and hyphens only. A maximum of 640:25characters and should match your0:26directory name. a description which is0:28also required which tells Claude when to0:30use the skill. This is a maximum of,0240:33characters and is the most important0:35field. Claude uses this for matching.0:37But we can also add other optional0:38fields. One of them is the allowed tools0:41field which restricts which tools Claude0:43can use when the skill is active. The0:45model field which specifies which cla0:47model to use for the skill. Try to be0:50explicit with your instructions. For0:52example, if someone told me my job was0:54to help with dogs, I wouldn't know what0:57to do. So, we have to assume Claude1:00would think the same way. A good1:02description answers two questions. What1:04does this skill do? And when should1:06Claus use it?1:08Now, if this job description was given1:10to me, I feel a little bit more1:12confident that I could get the job done.1:14If your skill isn't triggering, add more1:16keywords that match how you phrase1:17requests.1:20Sometimes you want a skill that can only1:22read files, not modify them. This could1:26be for security sensitive workflows,1:28read only tasks or more. We have the1:31allowed tools field to make this1:33possible. When this skill is active,1:35Claw can only use those tools without1:37asking permission. No editing, no1:39writing, no bash commands. If you omit1:42allowed tools, the skill doesn't1:43restrict anything. Claude uses its1:45normal permission model.1:49Skills share Claw's context window with1:51your conversation. When Claude wants to1:54use a skill, it will decide to load the1:56contents of that skill into context.1:58However, sometimes you'll need some2:00references, examples, or some utility2:03scripts that are required by the skill.2:06But cramming it all into one 20,000 line2:09text file means you take up a lot of2:10space in the context window. And let's2:13be real here, it's just not a lot of fun2:15to maintain that. This is where2:17progressive disclosure comes in. Put2:19your essential instructions in skill.mmp2:21and detailed reference material in2:23separate files that Claude reads only2:25when needed. The open standard also2:27suggests having a scripts folder for2:29executable code, references for2:31additional documentation and assets for2:34images, templates or other data files2:37that would be relevant for that skill.2:39Then in skill.md link to the supporting2:42files. Here, Claude reads2:44architecture.md only when someone asks2:46about system design. If they're asking2:48where to add a component, let's say, it2:51just never loads. It's like having a2:53table of contents in the context window2:56rather than fitting the whole entire2:57document in there. Keep skill.md under3:00500 lines. If you're exceeding that,3:02then maybe consider should this be split3:05up into different content.3:08Scripts in your skill directory can run3:10without loading their contents into3:12context. The script executes and only3:14the output consumes tokens. Tell claw to3:16run the script, not read it. This is3:19very useful for environment validation,3:22data transformations that need to be3:24consistent, operations that are more3:26reliable as tested code than generated3:28code.3:31Skills support metadata fields. Name and3:33description which are required. Allowed3:36tools restricts available tools and3:38model specifies which claw to use.3:40Descriptions need specific actions and3:42trigger phrases to match for3:44reliability. For larger skills, use3:46progressive disclosure. Keep your3:48skill.md file under 500 lines and link3:50to the supporting files that load only3:52when needed. Scripts [music] can execute3:55without loading their contents, keeping3:57context efficient.
634 words · 104 lines
