Background on Claude skill sprawl for AI-Powered Accountants
Claude has a real, functional skills ceiling. I hit it at 196 skills and got a hard stop from the platform. The fix was not deleting capability, it was the same discipline I apply to every financial model: audit, kill the orphans, and roll the rest into a proper architecture. Here is what happened, why it matters, and exactly how to avoid it.
A few weeks ago I opened Customize > Skills to add one more, and the panel made it clear that “one more” wasn’t happening without a clear-out first. So I did what any accountant would do when the tool won’t let you add another line. I counted.
It was 196. Four away from the ceiling.
When I tried to add one more, Claude’s panel stopped me outright with a flat message telling me I couldn’t add more until I cleared some out. That is a real, functional limit, not just things getting sluggish. What I couldn’t find anywhere in Anthropic’s published documentation is the exact number written down. The only 200 they’ll put in a help article is a character limit, not a skill count. That character limit caps every skill’s description field at 200 characters, because that description is the only thing Claude reads to decide whether a skill is relevant before it ever opens the file. So the ceiling is real, I hit it, I have the message to prove it, I just cannot point you to the page where Anthropic states the number. If you go looking for it yourself, you will find the same gap.
Here is how I got there, why it happened, and what fixing it actually looked like.
Key Terms
Skill A SKILL.md file containing instructions Claude loads when it judges the file relevant to your request. Each skill has a description of up to 200 characters that acts as the trigger.
Plugin A bundled folder containing a manifest and one or more skills, installable as a single unit. A plugin replaces dozens of loose skill entries with one install action.
Progressive disclosure The mechanism Claude uses to manage skill loading. Descriptions load into every conversation first. Full instructions only load if the description triggers. Reference files load last, only if the instructions need them.
Does Claude have a skills limit?
Yes. Claude.ai enforces a hard ceiling on the number of enabled skills an account can hold. When you reach it, the Skills panel blocks you from adding more until you remove some. Anthropic does not appear to publish the exact number in its help documentation, but the limit is real and the error message is unambiguous.
I was unaware that I was about to hit it at 196 personal skills. The definition of skill sprawl is when things get too messy you cannot track down what is useful and what needs to be binned.
Some of that count was marketplace boilerplate I had never built or touched, and some was digital-team builds I had shipped for clients sitting in the same environment as my own work. Strip those out and I was still close to the ceiling from my own work alone.
I didn’t set out to build 196 of anything. I built one skill at a time, every time I hit a task I knew I’d do again. A component of the EXL Cloud 3-way model. A monthly client revenue reconciliation. A capstone marking rubric for AIPA students. A worker persona for research, another for compliance, another for orchestration. Each one was a completely rational five-minute investment that saved me an hour next time.
“Rational in isolation” is exactly how a trial balance ends up with four hundred GL codes nobody can explain, or a financial model ends up with nine versions of the same revenue tab because nobody wanted to touch the one that already worked. I’ve spent my career telling finance teams to stop doing this. Stop bolting another tab onto the workbook. Stop hardcoding a one-off assumption instead of building it into the driver. Architect the model once, properly, and let structure carry the weight instead of habit.
Then I did the exact opposite to my own AI stack.
What happens when you hit the Claude skills ceiling through skill sprawl?
Every enabled skill puts its name and description in front of Claude as a candidate on every request. Past a certain point you are not adding capability, you are adding competition.
Skills use progressive disclosure. When a skill is enabled, Claude doesn’t load the whole thing into every conversation. It loads the metadata first, the name and the 200-character description, purely to judge relevance. The full instructions only load when Claude decides the skill applies to your request. Reference files load later still, only if the skill needs them. It is sensible architecture: you can carry a serious library without every message dragging the full weight of all of them.
But the metadata layer isn’t free and isn’t infinite. Descriptions start overlapping. Trigger phrases start colliding. The gap between “Claude picks the right skill” and “Claude picks a plausible skill” gets narrower the more plausible candidates are sitting in the queue. Claude Code, Anthropic’s coding-focused product, publishes an explicit token budget for exactly this reason, a setting that governs how much context window the skill listing is allowed to consume before entries get dropped or truncated. Claude.ai doesn’t expose the same dial, but the underlying pressure is identical: more candidates, less precision per candidate, until something has to give.
The ceiling is not a wall with a number painted on it. It is a slope. You don’t fall off it. You just notice the ride getting rougher, and by the time it is obvious you are already carrying more debt than you meant to.
How to fix Claude skill sprawl?
I ran the same discipline on my skill library that I’d run on a client’s chart of accounts. Same problem, different costume.
Audit before you architect. I have built skills specifically for auditing other people’s skill portfolios: checking for orphaned skills nobody has triggered in months, duplicated logic sitting in three places instead of one, descriptions so vague that Claude is guessing. Turning that same audit on my own library first was uncomfortable and completely necessary. You cannot fix what you haven’t inventoried.
Kill the orphans. Anything untriggered in months, anything superseded by a newer version I never went back to retire, gone. This is the equivalent of the workbook tab everyone is scared to delete in case something references it. Something usually does. Check, then delete.
Consolidate into parent and child, not one flat list. This was the real fix. Instead of fifteen separate top-level skills for every component of a 3-way financial model, one parent orchestrator skill that understands the whole build and calls the right child for the right component. Instead of six standalone worker personas competing for attention on every request, one router that holds context and dispatches. Same total capability, far fewer things Claude has to weigh at the metadata layer.
Version discipline, one canonical source. Where I had built the same capability twice because I forgot I already had a skill for it, I picked one as source of truth and made everything else either point back to it or disappear.
Roll related skills up into a plugin, not just a parent skill. A parent skill is still one more line in the list, just a smarter one. A plugin is a level up: a single installable bundle of skills, commands and tools that installs and shows up as one unit instead of a dozen loose entries competing for attention. I had already done this for two client digital-team builds, packaging a full worker roster as one plugin each client installs in a single action. Then I did the same to my own library in one sitting: 163 skills sorted and repackaged into 15 plugins.
None of this reduced what I can actually do. The library got more capable, because Claude is now choosing between a well-organised shortlist instead of scanning a flat pile for the least-wrong match. The fix for sprawl is never “do less.” It is “structure what you have built so the structure carries the load instead of your memory of what’s in there.”
What is a Claude plugin and how do you build one?
A plugin in Claude’s terms is a folder with a manifest and a skills directory inside it:
plugin-name/
├── .claude-plugin/
│ └── plugin.json
├── skills/
│ └── skill-name/
│ └── SKILL.md
└── README.mdplugin.json needs exactly one field to be valid: a name. Version, description and author are optional but worth including if anyone other than you will ever install it. Every skill I already had, complete with its SKILL.md and any reference files, moved into that structure exactly as it was. Nothing was rewritten or changed. It just got a folder to live in instead of standing alone in a flat list of 196.
Before zipping a single plugin, I ran a validation pass: every manifest must be valid JSON with a name field, every skill folder must contain a SKILL.md, nothing gets packaged broken. Here is where all 163 of my own skills landed:
| Plugin | Skills |
|---|---|
| exl-cloud-model-builder-suite | 30 |
| client-a-ops | 23 |
| aipa-education-content | 20 |
| exl-cloud-reporting-suite | 14 |
| workbook-engineering-toolkit | 11 |
| model-citizn-digiteam | 11 |
| skills-meta-tooling | 9 |
| fractional-cfo-general-ops | 8 |
| pf-training-model-builder | 8 |
| ai-model-eval-framework | 7 |
| legal-contracts-docs | 7 |
| model-citizn-proposals-content | 6 |
| client-b-ops | 4 |
| equity-sbp-ops | 3 |
| credit-portfolio-ops | 2 |
Fifteen installable units from 163 loose entries. The two biggest account for close to a third of everything on their own. That is not a coincidence. The busiest parts of my practice were always going to be the least-architected parts of my skill library, because volume beats intention every time nobody has stopped to check.
The audit surfaced something else: questions I had been quietly avoiding. Two skills reference a client’s employee register under a naming prefix that belongs to a different client entirely. And I have a standalone credit portfolio plugin I am no longer certain doesn’t overlap with a larger client operations plugin two rows above it in the table. A flat list of 196 will let you never notice either of those things. Fifteen named plugins will not let you look away from them.
Claude skills architecture: what to build before the limit finds you
If you are a finance or ops professional building a Claude skill library for yourself or your team, the lesson is not “stop at some magic number.” It is the same thing I tell every accountant who hands me a spreadsheet held together by faith: capability without architecture is fine right up until the moment it isn’t. A skill library that looks impressive in the count is worth exactly as much as a financial model that looks impressive in the tab count. It depends entirely on whether the structure underneath can carry what you’ve built on top of it.
Build the parent before the fifth child. Audit on a schedule, not when the panel stops you. And if you are building this for a team or for a client’s digital workers, package it as a plugin from day one. It is a lot easier to do that up front than to go back and consolidate 196 entries after the fact.
Frequently asked questions
What is the Claude skills limit and how to avoid skill sprawl?
Claude.ai enforces a hard ceiling on enabled skills. When you reach it, the Skills panel blocks new additions until you remove some. Anthropic does not appear to publish the exact number in its help documentation, but the functional limit exists and the error message is explicit.
You should track your skills and maintain them so that you are aware how many you are building and in which domains to avoid conflicts, orphans and a mess.
What is the difference between a Claude skill and a Claude plugin?
A skill is a single SKILL.md file containing instructions Claude loads when it judges the file relevant. A plugin is a bundled folder containing one or more skills plus a manifest file, installable as one unit. A plugin replaces many loose skill entries with a single install action, which reduces the number of competing candidates in Claude’s metadata layer.
What is progressive disclosure in Claude skills?
Progressive disclosure is the mechanism Claude uses to manage skill loading efficiently. Skill descriptions (max 200 characters) load into every conversation so Claude can judge relevance. Full skill instructions only load when Claude decides the skill applies. Reference files load last, only when needed. This keeps context-window usage manageable, but the description layer still accumulates competition as your skill count grows.
How do I know if I have too many Claude skills?
The clearest signal is a hard stop from the Skills panel telling you no more can be added. The earlier signal is degraded skill selection: Claude starts picking plausible skills instead of correct ones, especially when descriptions overlap or trigger phrases collide. If you find yourself re-prompting to force a specific skill, your library has grown past the point where the flat structure serves you.
Can I convert existing Claude skills into a plugin?
Yes. A plugin is a folder structure with a .claude-plugin/plugin.json manifest and a skills/ subdirectory. Each existing skill folder moves in unchanged. The manifest requires only a name field to be valid. Zip the folder and install it as a single plugin, removing the individual skill entries from your flat list.
Does rolling skills into plugins reduce capability?
No. The skills themselves don’t change. The instructions, reference files and trigger phrases all move across intact. What changes is how they’re represented in the metadata layer: one plugin entry instead of many skill entries competing for Claude’s attention on every request. The result is usually more reliable skill selection, not less.
If you want to dive in deeper take a look at our knowledge hub and the Future of Finance Lab
Lance Rubin is CEO and Founder of Model Citizn and EXL Cloud. He is a Chartered Accountant, FMI Certified Trainer, co-author of the CA ANZ Financial Modelling Study Guide, and creator of the AI-Powered Accountant programme on Maven, currently live at Level 1 with Level 2 in design, and author of the companion book “AI-Powered Accounting with Excel and Power BI” (Packt). He has worked at PwC, KPMG, Investec, and NAB, and is based in Melbourne, Australia.