Skills & the Autonomous Skill System

A skill is a package that teaches your agent how to perform a specific task. It bundles task instructions, executable code, reference documentation, and assets in one folder so they can be reused and shared.

Since May 2026, you rarely need to think about skills at all. Your agent autonomously finds the right skill for the job, surfaces a 1-click install card in chat when something better is available, self-repairs defects in skills it has forked, and periodically proposes library clean-ups when things get cluttered. You step in only at the moments that matter.

How your agent uses a skill

When your agent puts a skill to work, three things happen in order:

Search Explore: As the agent starts a task, it automatically scans the Explore tab for skills that could help.
1-click install: When a good match is found, an install card appears in chat. One click adds the skill to this agent.
Run immediately: The installed skill is available right away for the next task — no separate command needed from you.

Note: in scheduled runs, installing new skills from Explore is not allowed. Only skills already added to the agent can be called directly.

At a glance

Question	Answer
Do I have to install skills myself?	No. Your agent finds what it needs and shows a 1-click install card in chat. One button and you’re done.
Who decides which skill to use?	The agent analyses its own task notes, repeated patterns, and failing skill signals, then auto-suggests up to 3 candidates.
What if a skill breaks?	If the agent has its own copy of a skill, it fixes it directly; otherwise it files an improvement proposal to the skill’s author.
What happens when my library gets messy?	The system periodically checks and bundles any clean-up candidates into a notification (at most once a week per user). Nothing is ever automatically deleted or changed.
What about external services like Notion or GitHub?	Connect once and the skill will receive your credentials securely on every run.
Will skills affect my computer?	No. Skills run in an isolated cloud environment — nothing touches your local machine.

Seven moments you’ll meet skills

1. A skill suggestion card in chat

When the agent believes a skill would help with what it’s doing, it shows a card inline in the conversation. The card explains why and offers an Install button — one click installs and activates it immediately.

Suggestions are based on three types of signals:

Notes left after completing recent tasks (what worked, what fell short)
Repeated-pattern detection (when the agent keeps reaching for the same tool)
Failing skill signals (when a skill has errored repeatedly)

These signals are combined into a recommendation score, and only skills scoring 75 or above (out of 100) appear in the card. If you never see a card, your agent is handling things well enough on its own.

2. Bringing in a skill from a public GitHub repository

Beyond what your agent suggests, you can also pull a skill straight from a public GitHub repository.

From Explore: In the Skills tab → Explore, tap Import from GitHub at the top and enter the repository (owner/repo or a full URL). It’s imported right away. The result is registered as a private skill only you can see, which you can then install on any agent you choose.
In chat: You can also just ask your agent — for example, “install the skill from this GitHub repo” — with a link. In this case the agent first shows you what it found in the repository, and the install only goes through once you approve it.

The repository needs a skill description file (SKILL.md), and there are limits on file count and total size. If a repository contains more than one skill, you choose which one to bring in.

3. Connecting an external service once

Skills that work with Notion, GitHub, Slack, or similar services need a one-time login (OAuth connection). You can start that from the Connections section in your agent’s settings, or directly from a skill card in Explore. Once connected, the skill receives your credentials automatically on every run.

For the full walkthrough, see Connecting External Services.

4. A library clean-up card

Your skill library is scanned periodically (at most once a week per user). When clean-up candidates are found, you receive a notification and can review each item from the banner at the top of your My Library tab.

Four clean-up categories:

Category	What it means
Failing (FAILING)	Has accumulated 4 or more errors across all runs
Unused (UNUSED)	Installed more than 30 days ago and never run
Deprecated origin (DEPRECATED_ORIGIN)	The original author has marked the skill as no longer supported
Duplicate function (DUPLICATE_FUNCTION)	Two installed skills that largely overlap in what they do

Clean-up suggestions only ever ask “remove this?” Skills with a new version available, and versions revoked for security, are handled in the Updates received inbox described next — not here. When the same skill appeared in both places, clearing the inbox left what looked like unfinished clean-up behind.

Even when a card appears, nothing is automatically deleted. For each item you decide whether to keep or remove it, and since removal cannot be undone, you get one more confirmation before it is applied.

When the author of a skill you installed ships a new version, it shows up in My Library → Updates received. Each update carries a risk grade so you can tell at a glance whether it’s safe to apply as-is or worth a closer look first.

Update (green, informational): A light change. Apply it whenever you like.
Review changes (amber): The update changes executable code. Review it before applying.
Re-consent needed (red): Permissions or behavior change significantly. Opening the card shows exactly what changed and why; you must review it and tap Update anyway for it to take effect. Until you do, your installed version stays exactly as it is — nothing risky slips in unnoticed.

When several updates pile up, you can apply them all at once from the button at the top of Updates received instead of reviewing each one individually. Only the green and amber grades are eligible for bulk apply — updates that need re-consent (red) or a revoked version each require their own individual review, so they’re left out. You can hand-pick specific updates, or leave nothing selected to target everything that’s eligible.

Each update row also shows a short summary of what changed (for example, “1 new permission · 2 scripts added”). Updates that need re-consent show the re-consent notice instead of a summary — a plain list of facts would make the risk look tamer than it is.

Regardless of grade, tapping View changes on an update card shows exactly what’s different, file by file. The file list stays collapsed until you open it, and changed lines show up as additions (green) or removals (red). You only load the files you actually want to look at, so this stays light even when an update touches several files — and if a single file’s change is very large, only part of it is shown.

If a version is ever confirmed malicious and revoked, anyone who installed it sees a top-priority warning ahead of everything else. If a safe version exists, a button lets you move to it immediately; if not, only a remove button is offered.

6. Reviewing improvement proposals on your skills

When you publish a skill and someone else’s agent encounters a defect, the feedback is automatically filed as an improvement proposal. Multiple reports of the same underlying issue are deduplicated into one proposal.

On the My Library → Improvement Proposals page you can review each proposal and act on it:

Accept: Mark the proposal as reviewed. Then open the skill editor to apply the changes yourself and publish a new version.
Reject: Close the proposal with an optional note explaining why.

Defects in a skill you’ve installed but haven’t customized yet are routed to the original author as an improvement proposal, just like the case above. Once the agent has already self-patched that skill into your own customized copy, though, any further defects are fixed by the agent on the spot, with only a version snapshot kept — nothing is sent to the original author.

For full details on making skills and managing proposals, see Creating Skills.

7. The skill detail page — growth timeline, lineage view, and your agents using it

Skill cards in Explore show a health chip at a glance — a score out of 100 that combines how often the skill has run over the last 7 days, how well those runs succeeded, and how recently it’s still being used. It shows as a green (70+), yellow (40–69), or red (below 40) dot. A high success rate alone doesn’t guarantee a high score: a skill that’s rarely run or hasn’t been used in a while can still score low. Brand-new skills without enough run data show no dot. Skills derived from another skill also show the original author and how many contributors were involved right on the card.

Skill cards also show a small bar graph of recent weekly usage (in Explore and My Library alike, and consistent across web, iOS, and Android). The most recent week is highlighted in a darker color, and the graph is left off skills that don’t have enough usage data yet.

Opening a skill’s detail page shows different things depending on who’s looking:

If you authored the skill: you see a growth timeline showing how its versions built up over time. Node colors distinguish changes the agent made on its own from edits you made by hand, and the current version carries a “live on the market” tag if it’s published. You can roll back to any earlier version node directly from here.
If you installed the skill: you see a lineage view that follows the origin. Any origin version you haven’t pulled in yet shows up as a waiting card, and applying it follows the same risk grade and re-consent rules as the updates inbox above. Once you start customizing it, your copy moves into your My Library list — opening it from there shows a screen for merging the origin’s new versions into your copy, including the list of files you’ve customized. If applying an origin update conflicts with something you changed, and you have an agent using this skill, that agent takes over reconciling it in chat and reports back the result.
Either way: further down the page you can see the list of your own agents that actually have this skill equipped (only agents you own are shown — other users’ agents never appear).
When several agents customize the same skill differently: Say agent A starts adapting this skill its own way, and agent B independently starts adapting the same skill in a different direction. Your My Library list doesn’t show a separate card for every copy — instead, one card carries a small button showing how many copies exist (for example, ×2). Tapping the card or the button expands the list, showing which agent’s copy each one is (like “A’s copy”) along with a short change note describing what’s different from the origin. That note isn’t a canned phrase — it’s whatever the agent itself wrote down in the moment, explaining what was wrong and what it changed. Tap any copy in the expanded list to open its own detail page. This expand behavior works the same way on web, iOS, and Android.

What the agent handles on its own

The following all happen without you having to do anything.

Auto-suggest: After finishing a task, the agent reviews its own notes, repeated patterns, and failed skill signals to look for skills that might help. From candidates that are active, not yet installed, and above a health threshold, the top scorers are ranked and — only if they clear the 75-point bar — presented as a suggestion card.

Self-patch: When the agent finds a concrete defect in a skill it has installed, it applies an exact fix. The patch must match exactly one place in the file; after applying, the skill’s structure is re-validated. If validation fails, the change is automatically rolled back. A version snapshot is saved on each successful patch; the skill’s owner can roll back to any previous version from the skill’s version history view.

Auto-create: After finishing a task, if the agent recognizes a genuinely new, reusable procedure from its own experience, it creates a brand-new private skill and equips it on itself immediately — with no approval step (the same can happen during idle “dream” cycles too). The new skill starts out visible only to you, and you’ll get a notification once it’s created. This auto-create path is cost-gated and cooldown-limited per agent, so it doesn’t happen often.

3-way defect routing: When the agent reports a skill problem, it is automatically classified and routed. Official system skills go to the platform admin queue. Skills the user has forked queue up for the agent’s self-patch. All other skills accumulate evidence in a proposal for the author. Reports of the same issue within 7 days are merged into one proposal, and each user-and-skill combination is capped at 5 reports per day.

Auto clean-up & publish suggestions: Skills your agent creates on its own that go unused for a long stretch are automatically archived — tidied away, not deleted. You can restore any archived skill from its skill page at any time, and you’ll get a notification once it happens. On the flip side, if a skill your agent created keeps performing well, you’ll get a notification suggesting you publish it — publishing is always your choice.

Nine skill notifications

There are nine notifications related to skills.

Notification	Meaning
Improvement proposal received (SKILL_PROPOSAL_RECEIVED)	A new proposal has arrived for a skill you created
Library clean-up available (SKILL_LIBRARY_CLEANUP_AVAILABLE)	Clean-up candidates were found in your library
Skill deprecated (SKILL_DEPRECATED)	A skill you installed has been marked as no longer supported
Weekly health report (SKILL_HEALTH_REPORT_WEEKLY)	A weekly summary of success rates and usage across your skills
Auto-archived	A skill your agent created on its own was automatically archived after long inactivity
Publish suggested	A suggestion to publish a skill that’s been performing consistently well
Version revoked (SKILL_VERSION_REVOKED)	A version of a skill you installed was confirmed malicious and revoked — the same warning described in “Updates received” above (sent immediately)
Auto-created (SKILL_AUTO_CREATED)	Your agent created and equipped a new private skill from its own task experience
Auto-patched (SKILL_AUTO_PATCHED)	Your agent fixed a defect in an installed skill on its own

Improvement proposal, library clean-up, deprecation, and the weekly health report can each be toggled individually in Notification Settings. Auto-archived, publish-suggestion, version-revoked, auto-created, and auto-patched notifications can’t be turned off individually.

Ready to go further?

If you’d like to package a workflow your agent nailed into a reusable skill — or publish one — see Creating Skills.

Advanced

The material below is not required for everyday use. Refer to it only if you need to understand internal skill behaviour or debug an issue.

When an agent packages a workflow into a skill, a draft is first saved behind the scenes and a review modal appears in the interface. You confirm or edit the details, then click to create the skill (private or published). If you cancel, the draft is deleted immediately.

This DRAFT review modal applies when the agent explicitly decides in the moment to package the current workflow into a skill. Separately, as described in “What the agent handles on its own → Auto-create” above, when the agent reflects after finishing a task and judges a procedure worth reusing, it creates a new private skill directly — no draft step, no review modal — and only tells you about it afterward via a notification.

Installing a skill — content preview first

The first time an agent tries to install a skill, the full skill description and bundled file list are shown to you before anything is installed. The agent then proceeds to the actual install request, which still requires your approval before it takes effect.

Lazy loading of skill content

A skill’s full instructions are not always present in the agent’s context. The first time the agent runs a skill in a session, it fetches the complete README body and receives it in the tool result. Subsequent calls within the same conversation segment return only a short receipt — the body is already in the conversation history. When the history is filed (a new segment begins), the next execution fetches the full body again. The agent’s system-level instructions always hold only the frontmatter summary (a short description block); the full body travels exclusively through the execution result channel.

Limits on the change preview

The change preview on update and merge screens has a cap on how much of a single file it will show. Past that cap, only part of the file is shown, along with a note that it was cut off. When several files change at once, the overall response also has a separate cap, so files beyond that limit still appear in the list but without a content preview. The change content itself isn’t fetched just from opening the panel — it’s fetched only when you tap View changes (lazy loading). Expanding or collapsing an individual file after that only reveals or hides content that’s already been loaded.

Bulk apply processes at most 50 items at a time. If you selected more than that, the rest stay in the inbox so you can apply them on a later run.

Who writes the change note

The change note shown in the copy list, the updates inbox, and version history is not a system-generated summary — it’s a short explanation the agent is required to leave every time it patches a skill. Because this value can end up in a listing visible to other users on the market, it passes a secret scan before it’s shown. If the scan flags it, only the note is dropped; the underlying fix is still saved normally.

Installation slots: reference vs. your own copy

When you install a skill for the first time, it is registered as a reference pointing at the original. If the agent later self-patches that skill, a personal copy is created and the agent switches to that copy’s slot automatically. This transition requires no action from you.

A skill pulled from a GitHub repository is different: it starts out as your own private skill from the moment it’s imported, not a reference to an original (unlike installing a marketplace skill). If the repository owner changes the code later, that change is not pulled in automatically — import the same URL again if you need the latest version.

UpServe skills are compatible with the AgentSkills standard, so you can exchange them as zip files with Claude Code, Cursor, Goose, and other compliant clients. For the full spec — directory layout, frontmatter format, and secret injection — see Creating Skills → Advanced.