How do I organize a growing library of custom AI tools?
Short answer
Mark every tool with metadata before you add more. Tag each one with version, date, model compatibility, output quality, how often it gets used, and the job it does. When my Claude skill library grew from 55 to 148, three days of marking turned a scattered pile into a system I can actually use.
A few weeks ago I had 55 custom Claude skills and was running out of room to think.
By last Friday I had 148.
I did not build 93 new ones in that time. I found them. They were scattered across notebooks, old projects, half-done integrations, borrowed libraries, and client work I had forgotten about. They existed. I just couldn't see them all at once.
The moment I pulled them together, the problem flipped. Five tools for different audiences is manageable. A library of 148, where you can't tell which one to grab without rebuilding it from memory, is a different problem entirely. A system is a pile of tools with marks on every piece.
What did the audit look like?
I built a rubric with six dimensions and scored each skill against it. Every skill got tagged with version, date, model compatibility, whether it is editable in Cowork mode, output quality, and how often it gets touched. Each one got categorized by job instead of by function. All 148 got sorted so the next person (or next-me) knows what is there and why it exists.
That metadata layer took three days.
Nothing changed about the skills themselves. The work produced no new feature and fixed no bug. But now when I ask "what did I build for brand strategy," I don't grep my brain. I look at the marks and I know exactly which five skills to load, what to watch out for, and what to upgrade next.
The pile was always there. The marks made it a system.
Why doesn't consolidation fix the problem?
Companies with mature AI work tend to scatter their prompts, datasets, scripts, tuned models, and vendor chains across Slack, Drive, Notion, GitHub, and wherever else: a folder here, a laptop there. They know they should consolidate. They assume consolidating means copying everything into one place.
Copying moves the pile. Better names help a little. Neither tells you what a tool does, whether it still works, or what depends on it. You manage 150 tools by marking them, and the marks have to answer the questions you will ask under pressure.
What should every mark answer?
- What does this tool output?
- When did it last change?
- Does it need a model upgrade?
- Who owns it?
- Is it portable, or glued to a vendor?
- What breaks if it breaks?
- Which other tools depend on it?
Answer those once per tool and the library starts answering you back.
Do this today
- Count your custom AI assets in production: prompts, skills, workflows, scripts, tuned models.
- If the count is over 20, block one hour. It will save you five hours next month when you need to know what you actually built.
- Pull every asset into a single list, including the ones sitting in old projects and client folders.
- Mark each one with the questions above, starting with what it outputs and when it last changed.
- Sort the list by job, so the next time you need a tool you look it up instead of rebuilding it from memory.
Meredith's rule
A system is a pile of tools with marks on every piece.
Questions
What metadata should I track for custom AI prompts and skills?
Track version, last-changed date, model compatibility, owner, output, output quality, usage frequency, vendor lock-in, and dependencies. Anyone, including future you, should be able to see what each tool does and whether it still works without opening it.
Is putting all my AI prompts in one folder enough?
Copying everything into one place moves the pile without explaining it. A folder of 150 unmarked tools still forces you to rebuild from memory. Marking each tool with what it does, when it changed, and what depends on it makes the collection usable.
When does an AI tool library need an audit?
Once you pass about 20 custom AI assets in production. Past that point, one hour spent marking every tool can save five hours the next month, when you need to know what you built, which version is current, and what breaks if one fails.
Related service
AI Architecture Review
Want this worked out for your business?
Bring the real problem to a working session and leave with a plan.
Book a session