# Why Too Many Agent Skills Can Make Your AI Worse Source: https://www.humanstoknow.com/blog/ai-agent-skill-bloat More agent skills can make selection harder even when SKILL.md files use progressive disclosure. Learn where context bloat starts, what research found, and how to evaluate skills. Author: Akshay Apsingi Published: 2026-09-24T14:28:41.987Z Updated: 2026-09-24T14:28:41.987Z ## TL;DR - Progressive disclosure delays loading a skill's full instructions, but an installed catalog can still expose every skill's name and description to the agent. - Overlapping skills create a selection problem. A wrong match can add irrelevant instructions or miss a useful workflow; a 2026 benchmark found selection was the clearer source of performance loss. - Inspect the source, scripts, permissions and real user feedback before relying on a skill; upvotes are leads, not safety certificates. - Search for a skill at the moment of work, try it on a small task, and share honest feedback after using it. ![Illustration of collected knowledge and reusable building blocks.](https://www.humanstoknow.com/api/media/file/sha256-076b3e0fa01316de8421b87cf5911915f9d875e65d8a0b3bbdd47145d7d19152.png) Humans to Know editorial illustration. More skills should make an AI agent more capable. But when you ask Claude to improve a product screen, a crowded skill library creates another job before it changes a pixel: choose among frontend design skills with similar names and confident descriptions. Installing them all hands that decision to the agent every time you work. That trade is easy to miss. The wrong skill may trigger, a useful one may go unnoticed, and every unreviewed file adds another trust decision. Progressive disclosure limits how much of each skill is loaded, but it does not guarantee the agent will pick well. The question is whether a skill improves the task in front of you. ## 1. What does a SKILL.md file actually give an agent? An agent skill is a folder with a `SKILL.md` file containing a name, description and instructions. It can also include references, assets and executable scripts. The [Agent Skills specification](https://agentskills.io/specification) describes a three-stage loading pattern: the agent first sees metadata, loads the full instructions when it chooses the skill, and reads supporting files as needed. That design is useful. A well-written skill can carry a repeatable procedure without pasting its whole manual into every conversation. But the name and description still need to be available for selection. A skill's description is therefore both a promise to the reader and a routing hint for the agent. ## 2. Does progressive disclosure prevent agent skill context bloat? Progressive disclosure keeps the full instructions out of context until needed. It does not necessarily hide the catalog. If your client lists installed skills in the model's context, each additional name and description takes space before the agent has chosen anything. The cost depends on the client, the descriptions and its context budget; there is no reliable universal “tokens per skill” number. The limit is visible in current tooling. [Claude Code documents](https://code.claude.com/docs/en/skills) that it shortens descriptions when its skill listing exceeds its budget. [OpenAI's Codex guidance](https://developers.openai.com/blog/rethinking-skills-and-prompts-for-gpt-6-astra) also warns that too many long or overlapping descriptions make choosing harder. The full `SKILL.md` may wait until later, but a large menu is not free. ## 3. Why can more options slow the work down? Because “which skill?” becomes a task of its own. Imagine a developer choosing among “frontend design,” “UI polish,” “component design” and “design system” skills. The names overlap, while their useful boundaries may be very different. One may critique an existing screen; another may build reusable components; a third may assume a particular styling stack. The human spends time comparing descriptions and repositories. The agent must make the same distinction from limited metadata. A broad description can pull the wrong skill into a task; a narrow but poorly worded description can hide the right one. These are failure modes to test in your own setup, not a claim that every large catalog performs badly. ## 4. Can too many skills make an AI agent worse? Yes, when a larger library makes the agent choose the wrong skill or overlook a useful one. A [2026 research preprint](https://arxiv.org/abs/2605.24050) compared known-helpful skills with libraries of up to 202 across 38 selected SkillsBench task and model combinations using Claude Haiku 4.5 and Sonnet 4.6. At 202 skills, the authors reported a 21 percentage point drop in pass rate. They call the dominant effect *skill shadowing*: wrong or missed skill selection explained more of the drop than extra context, whose measured effect was not statistically distinguishable from zero. This is one benchmark, not a universal skill-count limit. An activated skill contributes instructions to the work. If those instructions assume a different framework, design system or testing convention, the agent has to reconcile them with your repository and your request. Two overlapping skills may even offer conflicting advice. The result can be extra edits, inconsistent choices or a polished answer to the wrong problem. More instructions are useful only when they supply knowledge the task needs. [OpenAI's skill guidance](https://developers.openai.com/plugins/build/skills) recommends a focused workflow with a clear trigger, inputs and output, then testing both correct and incorrect activation. Treat “does this improve my actual result?” as the measure, rather than the size of your installed collection. ## 5. What is the security cost of an unreviewed skill? A skill can contain instructions that try to redirect an agent, plus scripts it may be asked to run. The [open format explicitly permits executable scripts](https://agentskills.io/specification). A bulk install multiplies the files and dependencies you have not examined, and an upstream update can change what you previously reviewed. This is a supply-chain and prompt-injection concern, not proof that community skills are malicious. [OpenAI's prompt-injection explainer](https://openai.com/safety/prompt-injections/) describes how lower-trust content can try to steer an assistant. Before using a skill, inspect its `SKILL.md`, referenced files, scripts, source commit and requested tools. Keep your agent's normal permission controls in place. A popular listing does not grant a file or command extra authority. ## 6. Why isn't a large repository enough for discovery? A repository tells you what its author published. It may not tell you whether someone used the skill on a similar task, what broke, or what changed after feedback. Stars and download counts are useful discovery signals, but they do not answer whether the instructions fit your project today. The missing piece is a shared record of experience beside the skill itself: who made it, where it came from, what users found helpful, and what they would change. That makes discovery a conversation rather than a one-time install command. You still need to read and test the source yourself. ## 7. What can community feedback tell you? An upvote can point you toward a candidate worth opening. A specific comment can do more: “worked for a Tailwind dashboard,” “assumes React,” or “the script needs network access.” Those details help the next person decide whether to investigate. Neither signal proves quality or safety. A new skill may have no votes because few people have seen it; a highly liked one may solve a different problem. On [Humans to Know's skill directory](/skills), use community activity as context alongside the creator, source and instructions. Read comments for concrete use cases, then judge the result against your own task. ## 8. How does Humans to Know help at the moment of work? Humans to Know lets you [browse skills and their creators](/skills) or [connect an agent through MCP](/agent). With the connection in place, an agent can search published skills, open a candidate's full instructions and source details, and read its comments. That lets you ask for help with a new workflow when it appears, instead of pre-installing a large, speculative bundle. An MCP search still uses some tool context; it is not a promise of zero overhead or perfect recommendations. The practical difference is that the agent can retrieve candidates for the current task and explain why one fits before you decide to use it. The website remains available when you want to inspect the listing yourself. ## 9. What would a UI design search with Claude look like? After [connecting Humans to Know](/agent), you could ask: “Find a skill for improving the visual hierarchy and mobile layout of this React settings page. Compare the source and instructions, explain your pick, and don't install or run anything yet.” The agent can use the MCP `search` tool to find candidates, `get_skill` to read promising instructions and provenance, and `list_comments` to see any discussion. Its answer should name a skill, link to it, explain the fit and dependencies, and say what it still cannot verify. You can then try the skill on a small screen and inspect the result. This is an example workflow, not a claim that a listing has been independently tested or that Claude has installed it for you. ## 10. Which UI design skill fits which task? There is no single best “UI design” skill. These current directory listings show why the task matters: | If you need to… | Candidate to inspect | Why it might fit | | --- | --- | --- | | Critique or polish an existing interface | [impeccable](/human/pbakaus/skill/impeccable-38e53fc5db42) | Its listing covers interface review, hierarchy, accessibility and responsive behavior. Check whether its broad scope matches the change you want. | | Design reusable UI component APIs | [web-component-design](/human/wshobson/skill/web-component-design) | Its listing focuses on component patterns and libraries across React, Vue and Svelte. Check its framework assumptions before use. | | Build a dark React dashboard with a defined stack | [frontend-ui-dark-ts](/company/microsoft/skill/frontend-ui-dark-ts) | Its listing names Tailwind, theming and motion. Check the repository and whether those dependencies belong in your project. | These are examples to compare, not endorsements or a ranking. For a more detailed evaluation, use our [agent skill selection checklist](/blog/choose-ai-agent-skills). ## 11. How can MCP feedback improve skill discovery? Give your connected agent Write access and a standing instruction to share feedback after it uses a skill. It can then autonomously upvote a skill that helped and comment on what worked and what still needed fixing, under your account. Those real-use votes and comments create a feedback loop that helps other people choose. ## 12. What should you change in your own workflow? Keep a small set of skills you know you need in the project where you need them. When a new task arrives, search by the outcome, compare a few candidates, read the instructions and source, and test one on a bounded piece of work. Keep the skill only if it helps. Claude Code's [project-level skill scope](https://code.claude.com/docs/en/skills) is one way to avoid making every personal skill available in every repository. Start with [the Humans to Know directory](/skills) or [connect your agent](/agent). Find one skill for the work in front of you, then leave a useful comment if you try it. The next developer should be able to make a better choice because you did.