Mimir
Skills
A skill is a procedure a bot follows step by step for one kind of task: your morning briefing, booking padel, the expense report. A bot sees only each skill's name and when to use it; it reads the steps when a task fits (or when you call the skill by name), so having skills costs little until they're used.
Every skill has the same shape, whoever wrote it: When to use, You'll need, Steps, Check (how to tell it worked), Return (what to tell you) and Stop and ask if.
They live on the Skills tab (⌘3): your bots' skills in this chat, suggestions waiting for you, and the library.
Mimir's skill library #
Skills written for Mimir's tools, ready to add. The ideas come from what people use other assistants for every day; the text is Mimir's own. They're plain text: no scripts, nothing downloaded.
| Skill | For | Needs |
|---|---|---|
| Morning briefing | Today's calendar, open tasks and what you follow, in ten lines | memory (a linked calendar helps) |
| Inbox triage | What in your email needs you, FYI and noise; short replies as drafts; scams flagged | |
| Follow-up tracker | What you're waiting on, and what you promised others | memory, earlier chats |
| Weekly review | Done, slipped, next week's priorities | memory, earlier chats |
| Meeting prep | Before a meeting: the people, last time, what's open | a linked calendar |
| Meeting follow-up | After a transcribed meeting: decided, to do, open questions | a transcribed meeting |
| Research brief | A question answered from current sources, each claim numbered | reading web pages (web search helps) |
| Summarise a link | The gist of a page you send | reading web pages |
| Price watch | Tells you when a product drops below a price or is back in stock | reading web pages, routines |
| Topic watch | A digest of what's new on topics, companies or feeds you follow | reading web pages, routines |
| Write like me | Drafts in your own voice, from how you write | earlier chats |
| Decision memo | Options, trade-offs and a recommendation, after a few questions | nothing |
| Document to actions | Deadlines and obligations from a contract, letter or form | memory |
| Trip planner | A plan by day, a booking checklist, reminders and calendar links | reading web pages |
Add a skill. Nothing from the library is added on its own. Pick who gets it at the top (all bots in this chat, or one bot) and press Add: it becomes an accepted skill like any other, which you can change, share or remove. A card says when something it needs isn't there yet ("needs a linked calendar"); you can add it anyway, and it works once it is.
Schedule it. The ones that are a habit offer schedule it once added: one click makes a routine that runs the skill at a sensible time (morning briefing weekdays at 07:30, weekly review Fridays at 16:00, follow-up tracker Mondays at 09:00). Change the time in Routines.
Using a skill #
A bot reads a skill when your request fits its "when to use". To use one on purpose, type / in the app's chat to
list the bot's skills: /morning-briefing, or /book-padel Thursday 18:00 to use that skill for Thursday. The same
works in Telegram, where the command menu lists each chat's skills (/book_padel), and Discord. Slack keeps / for
its own commands, so there ask by name ("use the book padel skill"). A routine can run a skill too: Or run one of
its skills in the routine form (its prompt becomes /book-padel).
Where skills come from #
From the library, above.
Learned from a task that went well. When a task that took several steps went well (no failed or denied action, no correction after, nothing read from outside), the tidy-up may suggest it as a skill: a name, when to use it, and the steps that run took. A bot uses a skill only after you accept it. Skills belong to one bot in one chat, or to all bots in that chat: pick it under the skill (a shared skill's fixes and new versions stay shared, and two skills the same bot sees can't share a name). A new version waits for you while the old one stays in use.
If you corrected an earlier answer to a similar request, the tidy-up also sees what you said then, and writes it into the skill as a step or a Stop and ask if line ("You left out the parking slips" becomes "read every receipt, even small ones like parking slips"). Only your own words are used, never the reply you corrected. A suggested skill whose "when to use" reads like one the bot already has says so on its card ("Similar to your skill 'Book padel'"): two skills for the same job make a bot pick the wrong one, so keep one (or merge them).
Save that as a skill. Right after a bot did something you'll want again, say "save that as a skill". The bot writes
it (skill_save) from what that conversation actually did, and a Skill to save card appears under its reply, with
Save / Not now (it's on the Skills tab too). The card says which steps it was drawn from, and warns you when the
conversation read outside content. Nothing is saved until you press Save.
Teach a skill. On the Skills tab, Teach a skill: paste a link or the text (how your expense process works, a
checklist, a recipe) and pick a bot. Its AI distils it into a skill: when to use it, the steps or rules, and short
reference sections the bot reads only when it needs them (skill_view with a section). It came from outside, so it's a
proposal you read and accept, never active by itself; a source or skill containing instruction-like text ("ignore
previous…") is flagged on the card. Teaching the same topic again proposes a new version, shown as changes. On a
self-hosted server you can also teach by showing: do the task once on the bots' computer and a bot drafts the skill
from what you did (The bots' computer).
Skill files. Skills use the same SKILL.md format as Hermes, Claude Code and other agents (agentskills.io). Paste one,
or a link to one, into Teach a skill and it's imported as written (no AI involved), as a proposal like any other.
Copy as SKILL.md on a skill gives you the file to share or use elsewhere.
How skills improve #
Skills improve with use. Every time a bot reads a skill, that's counted (the Skills list shows "used 3× · last 2 days ago"). When a run that used a skill then went wrong (an action failed or was denied, the run was stopped, or you corrected the reply), the tidy-up sees the skill next to what happened and may propose a fix: a small patch to the step that caused it, shown as the lines that change, with the reason ("The 'any court' step led to booking an outdoor court, and you corrected it"). It can only patch text that's in the skill, never rewrite it, and a skill stays under 4,000 characters. Apply fix makes it the new version. Runs that had read outside content (a web page, a file, a tool's result) are never used for fixes: a page could steer a run wrong on purpose so its "fix" carries the page's instructions, and a planted step is hard to spot on a card.
Upkeep, once a week. So skills and memory don't pile up:
- a skill nobody has used in 60 days gets a Retire card (Keep it dismisses the card for 90 days). Retired and removed skills stay under retired or removed skills for 180 days, with Bring back;
- two skills that clearly do the same job can get a Merge card (the merged skill replaces both);
- a fact search hasn't brought up in 90 days gets a move to background suggestion (preferences and tasks are left alone; pinned memories never move).
Nothing changes until you accept. This is the opposite of agents that rewrite their own skills unattended: the evidence is what went wrong, and you decide.