Skills — Long-term abilities for the AI
Skills give the AI long term abilities A skill is a reusable instruction set you teach Krater once and reuse forever. Think system prompts, but smarter — Krater only loads a skill into the model's ...
Skills give the AI long-term abilities
A skill is a reusable instruction set you teach Krater once and reuse forever. Think system prompts, but smarter — Krater only loads a skill into the model’s context when it’s actually needed, so you can keep a large library without slowing anything down or burning Credits on tokens you didn’t use.
Skills follow the open agentskills.io spec, so anything you build in Krater is portable to other AI tools that support the same standard.
Manage your skills in Settings → Skills.
When to use a skill (vs. a slash command or memory)
| You want… | Use… |
|---|---|
| A persistent workflow, tone, or domain expertise the AI applies on the right requests | Skill |
A one-shot action with a UI affordance (e.g. /image, /research, /summarize) |
Slash command |
| Short, always-on facts about you (your name, role, tone preference) | Personalization tab |
Skills are the heavy lifters. A good skill captures how you want a class of task done, not just facts about you.
The personal facts live in Memory, which is editable in Personalization — the same memory also appears as a pinned Memory — Always on card at the top of the Skills tab. Its contents are injected into every chat, no activation needed.
Three ways to get a skill
1. Start from an example
Krater ships with a curated set of example skills you can copy into your library:
- blog-seo-audit — Audit a blog post or landing page for on-page SEO with prioritized fixes
- competitor-watch — Changelog-style monitoring of a named competitor over a time window
- linkedin-post — Draft LinkedIn posts with a hook checklist and AI-tell avoidance
- meeting-prep — Build a pre-meeting brief with attendee research
- rfp-response — Structured response workflow for RFPs, RFIs, and RFQs
- weekly-digest — Roll up the week’s status with audience-appropriate tone
Click Use on any example to copy it into your library as an editable draft. Already added it once? The button becomes Add another copy so you can keep variants.
2. Create your own
In Settings → Skills → New skill → New skill, give it:
- Name — lowercase letters, numbers, and single hyphens, 1–64 chars (e.g.
support-triage) - Description — up to 1,024 chars. When this skill should fire and what the AI gains from it. Krater shows this to the model in a catalog so it knows when to load the full body.
- Body — Markdown, up to 64 KB. Describe the workflow, rules, or knowledge in detail.
- License and Compatibility — optional metadata for portability.
A good description is one of the most important parts of a skill — it’s what the model sees in the catalog when deciding whether to activate. Aim for “When the user wants X, this skill teaches the model to do Y, Z, and W.”
3. Upload a SKILL.md file
In Settings → Skills → New skill → Upload SKILL.md, import a skill file from disk (.md, .markdown, or .txt, up to 128 KB). The file uses the standard agentskills.io layout: YAML frontmatter with name and description, then the Markdown body. Optional license, compatibility, and allowed-tools fields are preserved.
If the frontmatter name is missing or invalid, Krater falls back to a cleaned-up version of the filename.
The trust gate
Every new skill — whether you created it, uploaded it, or copied it from an example — lands in your library as Needs trust. The AI cannot see or load an untrusted skill, even though it’s marked active.
Click I trust this skill to acknowledge you’ve reviewed the body. Then use the Inject into chat switch to activate or pause it.
This is the agentskills.io client-side trust acknowledgement — it exists because skill bodies can contain instructions you’d be uncomfortable having Krater follow blindly.
How loading works
When you send a chat message:
- Krater injects a compact catalog of your active skills (just names + descriptions) into the system prompt.
- The AI reads the catalog and decides which skills (if any) it needs for your request.
- To load one, the AI calls the built-in
activate_skilltool with the skill name. - The full skill body is returned to the model, which uses it to answer. The body stays in context for the rest of the conversation.
This means you can keep dozens of active skills without bloating every request. Only the catalog is always-on; bodies load on demand. The catalog itself is budgeted at roughly 3,000 tokens — if you have more skills than that fits, the lowest-priority entries are omitted, so keep your descriptions tight.
Editing & deleting
- Edit any skill from the row’s pencil icon. Changes take effect on the next message.
- Delete removes the skill permanently — including any bundled resources. There’s no undo.
- Inject into chat pauses a skill without deleting it — useful when you want to keep a skill around but don’t need it active right now. Paused skills show a Paused badge and never reach the catalog.
Tips for writing great skills
- Lead with the trigger. Start the body with “When the user wants X, do Y” so the model recognizes the moment to use it.
- Be explicit about format. If you want a specific output shape (table, checklist, email draft), describe it down to the section headers.
- Include examples. A worked example beats abstract instructions every time.
- Set guardrails. Tell the AI what not to do — e.g. “Never invent metrics; if you don’t have data, say so.”
- Keep it focused. One skill = one job. Two related-but-distinct workflows should be two skills with two descriptions.
- Keep the description short. It’s on the menu the model scans every message — a crisp one-liner activates better than a paragraph.
FAQ
Are skills shared with my team? No — skills live in your personal library only. Team-shared skills are on the roadmap.
Can a skill call tools or other skills? A skill can instruct the model to use any tool the model already has access to (web search, image generation, file uploads, etc.). It cannot directly call another skill, but it can describe a workflow the model executes step-by-step.
Do skills cost extra Credits? No. activate_skill itself is free — the catalog is a small token addition to your system prompt and the body only loads when the model calls for it. Either way, you’re charged the normal per-message rate based on your model.
What if I have two skills with the same name? You can’t — names are unique per account. Copying an example whose name already exists in your library appends -copy, -copy-2, etc.
Is there a skill marketplace? Not yet. The examples are a curated starting set; broader sharing is on the roadmap.
For questions about skills or to suggest new examples, email hey@krater.ai.