Skills (Beta)
Last reviewed: 2026-11-20
A skill is a set of instructions for work that comes back: preparing a meeting, sorting your tasks, a weekly review. Plainva comes with fourteen of them, and you can write your own. Skills use the open Agent Skills format — a folder with a SKILL.md — so they also work in other AI apps that read the format.
Using a skill
Start a skill with one click: under Skills below the input field of an empty conversation (the featured ones first), under Skills in the AI tab, on the phone in the segment Skills of the screen AI, or from the command palette. The conversation then runs with the skill: its instructions go along, and it uses only the tools and folders the skill names — a skill that names no tools gets those that only read or show.
You can also simply ask. In a conversation with tools the AI knows the names and descriptions of your active skills — as many as fit into 6,000 characters, the others stay startable by hand — and loads one when your question matches — “prepare my next meeting” is enough.
The skills that come with Plainva
| Skill | What it does |
|---|---|
| Daily orientation | What matters today: due tasks, appointments and what you worked on lately. |
| Weekly review | The past seven days and the week ahead, with three suggestions. |
| Project status | Goal, progress, open points and the next step of a project. |
| History analysis | Looks back over a period: what changed, how projects developed, which decisions you wrote down or revised, and what is still open — from what your notes gained day by day. A line you deleted is counted, never shown. |
| Meeting preparation | Prepares a meeting from earlier notes and open points, or writes one up afterwards. |
| Task triage | Sorts your open tasks: what now, what can wait, what to drop. |
| Research | Researches a question on the web and in your notes, with every source named. |
| Mail and calendar | Goes through recent mail and the coming appointments: what needs an answer, what to prepare, which tasks follow. |
| Write and rewrite | Summarises, shortens or rewrites a note — as text you take over. |
| Knowledge upkeep | Finds notes that say the same, are out of date or are connected to nothing. |
| Link cleanup | Checks the links of a note: leading nowhere, missing, one-way. |
| Memory care | Looks through the memory: entries that say the same, contradict each other or are out of date — and drafts what to merge and what to take out. |
| Privacy check | Finds what in a note should stay on this device, and suggests a rule. |
| Reflection | Looks back on the notes of a day or a week with you — kindly, never a diagnosis. |
All of them only read: none changes a note or sends anything. The one that drafts is Memory care: what it suggests for the memory waits until you accept it. Only Research uses the internet, and only Mail and calendar reads your mail — see below. The privacy check and reflection are meant for a model on this device; with a cloud model the send overview says so. Switch any skill off under Skills — the switch holds for this vault on this device. Make your own version copies one into your vault, where you can change it.
On the internet and in your mail
Research is the one skill that comes with Plainva and uses the internet. Starting it is your choice for its conversation, like the globe under the input field: where you switched on The AI may use the internet in this vault, it searches and reads pages — and while your notes are in the conversation, every page and every search still asks first, as described under On the internet in AI Assistant. Where the switch is off, it researches in your notes only and says so. The same holds when the AI loads the skill by itself in a conversation you started without the internet.
Mail and calendar reads mail through the same question as any conversation: Read your mail? the first time. It never reads the text of a message itself; a second reader without tools writes a report about it. Both are described in AI Assistant.
A skill of your own brings the internet into its conversation only when its allowed-tools line names web_search or fetch_url. Review and approve then says Uses the internet where you allowed it for this vault. before you approve it. A skill without an allowed-tools line never brings the internet along, but it uses what its conversation has: in a conversation that already has the internet, it can search and read pages like the conversation itself. A skill whose allowed-tools line names neither has no internet in its run, whatever the conversation has.
A test run never uses the internet, and it never asks: mail it was not allowed to read in this session stays unread.
Proposing changes
A skill of your own proposes changes only when its allowed-tools line names the tools for it: propose_edit and set_property for suggestions on a note’s text and on its properties, create_note, create_entry, create_task and add_journal_entry for drafts, rename_note, move_note and delete_note for plans. Review and approve then names each of them and says May suggest changes, leave drafts and lay out plans — they wait for you to accept, create or confirm. and, below it, Except with “Apply directly” or in a folder with “Change without asking”: there it changes and creates notes at once, with Undo. Deleting, renaming and moving still ask. Nothing in the vault changes before you accept, create or confirm — except where the conversation has Apply directly on or the note lies in a folder with Change without asking: there a skill that names propose_edit, set_property, create_note, create_entry, create_task or add_journal_entry writes at once, with a version before and Undo (see Apply directly in AI Assistant). A skill that names no tools proposes nothing — one you approved earlier gains nothing either — and a test run leaves nothing behind. What the three forms are: Proposing changes in AI Assistant.
Your own skills
The button New… offers a skill, a script and an import. New skill asks for a name, a description — the AI chooses a skill by it — and the instructions. Plainva writes them as .agent/skills/<name>/SKILL.md into your vault, where they travel with it like any note. Edit opens the file like a note.
Import… (in the same menu) takes a skill as a .zip or .skill file. Before anything is written, Plainva checks it: exactly one skill in the format, no path outside its folder, the size limits. It names the licence, scripts it will not run and tools it does not have. Hidden files — names that begin with a dot — are not part of a skill and are left out.
Nothing runs before you approve it
A skill in your vault that is new or changed — through sync, another program that wrote it, or an edit on this or another device — does not run until you approve it on this device. Such a skill waits at the top of Skills under Waiting for your approval, and in Settings → AI in this vault. A skill you create here (New skill), copy here (Make your own version) or import here (Import and approve) is the exception: it is approved on this device at once, because you saw it here. A sentence at the top of the view explains what skills are; the order is waiting for approval, Comes with Plainva, Your own, scripts, from servers, the row Instructions for this vault (with AGENTS.md as its second line), test runs and what was learned. Withdrawing an approval asks first. Review and approve shows what the skill may do, what changed since your last approval, its instructions, files and where it lies. The approval holds for exactly this version; any change lifts it again. Approvals are stored on this device, never in the vault.
The same holds for an AGENTS.md at the top of your vault: once approved, its standing instructions go into every new conversation. Neither a skill nor AGENTS.md can lift your privacy rules, and a skill never gets more than a conversation has — it can only narrow it. A rule you add in the memory, or accept from the AI, is one more line of this file; see Memory.
Skills of your servers
A server you connect under External tools can bring skills of its own: guides for its tools, written by whoever runs it. They stand under Skills in the group From servers, each with the name of its server — never among your own and never in place of one: the AI loads such a skill under a name of its own, such as mcp_tracker/code-review. A server’s skills are offered only in a vault that uses the server.
Review and approve shows the server and where it is reached, the skill’s name, what the server writes about it — as its quote —, every file with its size and the start of its checksum, and what the skill may use. Plainva reads none of its files to show you this; Show instructions reads its SKILL.md now, and each file the AI may read of it, and checks every one first — one look shows at most 20 files and 1 MB together. The approval holds on this device for exactly these files and this description: if the server changes one of them, the skill waits again, and its review lists what changed — and, if you read its instructions before you approved it, compares the new ones with them line by line.
A server’s skill may use Plainva’s tools that only read or show, and the tools of its own server that the vault allows — every call is asked as always. Whatever tools it names, it gets none of Plainva’s tools that change anything and none of the app’s commands, no internet, no script, and never the tools of another server. Once such a skill — its instructions or one of its files — is part of a conversation, the rest of that conversation keeps to these limits, whichever skill is loaded after it. The AI loads it in a conversation you start yourself, when your question matches; a conversation a skill started, a routine and a test run do not offer it, and neither does a device that is fully local. Each file is read from its server only when it is needed and checked against what you approved; a file that differs is not used, and the files beside its SKILL.md reach the AI as the server’s data, never as instructions.
Read again asks the server for its list now. Add by address… takes a skill the server offers without listing it, by the address of its SKILL.md; it then waits for your approval like any other. Remove server under External tools takes the approvals of its skills in this vault along.
Test runs of skills
A skill can bring test scenarios: a message that starts it, and what a good run does. Thirteen of the fourteen skills that come with Plainva have them — all but Memory care; for your own, write them into tests/scenarios.json in the skill’s folder:
{
"version": 1,
"scenarios": [
{
"id": "rates",
"message": "Check the offer against last year's rates.",
"tools": { "required": ["read_note"], "forbidden": ["run_command"] },
"cites": ["Offer"],
"never": ["internal margin"]
}
]
}
tools names the tools a good run uses and those it must leave alone, cites the notes its answer names, never text that must not appear in it. A skill has at most eight scenarios.
Test run with ⟨model⟩ — at the bottom of Skills (Test run…), or in a skill’s menu — runs the scenarios against the model a new conversation would use. Nothing starts by itself: the dialog first says how many scenarios would run and against which model, and you set a Ceiling in US dollars; the run ends between two scenarios once it is reached. A limit of 400,000 tokens ends the run as well, also where a price is known; a model on this device needs no ceiling.
Each scenario is an ordinary run of its skill: it reads your vault like a run by hand, passes the same overview before sending, counts towards your usage and leaves its conversation in the history, where its next run replaces it — unless it is open on screen or you went on with it; then it stays. Afterwards every scenario shows its verdict in words, and the skill’s row says how its last run went. A result is about one model and one version of the skill: once you choose another model or change the skill, the row says so instead of showing a result that no longer counts. Some scenarios that come with Plainva ask about notes of Plainva’s own test vault; in your vault they do not apply, and the dialog counts them apart instead of failing them.
Learning from a conversation
A conversation can leave something behind: a fact worth knowing, a rule, or a skill that did not go far enough. Learn from this conversation — in a conversation’s menu in the list, and under its last answer — asks for that. Nothing reads your conversations in the background.
A dialog first says what would happen: the conversation goes once more to the model that led it, and to no other — what you wrote and what was answered, with the names of the tools that were used. Nothing a tool returned goes along, and no note. Where a skill of your own ran in the conversation, its instructions go along, so that a better version can be suggested. Learn starts the review; it costs one request.
What comes back are drafts, each with the Evidence the review gives for it, and none of it counts before you accept it. An entry for the memory and a rule are decided on their cards, as described in Memory. The draft of a skill has the button Review instead.
A conversation that read a web page, an e-mail or an external tool suggests entries for the memory only: what a stranger wrote does not become a rule or a skill. The same holds where the conversation rests on notes that are kept from the cloud or from the internet, and entries from it carry that rule. A conversation that ran with a model on this device is reviewed on this device.
Accepting a suggestion for a skill
Review shows what would change, line by line, and what the skill may do — which stays as it is: a suggestion changes a skill’s instructions and nothing else. Its tools, folders and limits are never set by a model. The dialog also says whether the current version was tested, what a run costs more or less, and which conversation the suggestion came from. Rework turns the comparison into a field you can type in.
Accept writes the new version and approves it on this device, because you saw it here. On your other devices the skill then waits for their own approval, like any change. A suggestion for a new skill starts with Plainva’s defaults: it reads and shows, and changes nothing. A skill you imported and the skills that come with Plainva are never rewritten by a suggestion.
Watched versions and the way back
A skill that a suggestion has rewritten is watched for three runs of the new version, and its row counts them. If a run runs out of steps or tokens, goes in circles, is cut off or is refused — a run you stopped, or one the provider did not answer, does not count —, the skills view names the skill under Watched versions and offers two things: Back to the version before, or Keep. Nothing goes back by itself.
Earlier versions… in a skill’s menu lists what the vault’s version history keeps of the skill’s file. The dialog holds the version you choose against the skill as it is now — its lines, and what it may do — and warns where the earlier version may do more. Restore this version writes it back and approves it on this device. The versions are kept on this device.
For a skill that changed in any other way — through sync, or an edit on another device — Review and approve says the same under Compared with the approved version: which tools came and went, the folders, the limit.
Under What was learned the skills view opens .agent/logs/learning.md: one line for each skill and each rule that was accepted from a suggestion, with the day and the conversation. The file travels with your vault.
Tidying up
Skills pile up. Under Tidy up, the skills view names what this device noticed by itself. No model is asked for it and nothing is sent, and each line is a question, not a finding:
- Two skills that say almost the same, so that the AI picks one or the other. Compare puts them side by side.
- A skill of your own that has not run for more than 90 days. Switch off takes it out of the catalog; it stays in the vault.
- A skill whose list names a tool Plainva does not have. It runs without that tool.
- A skill that did not pass its test with the model chosen now.
- A skill whose runs keep ending without an answer, and a way you went by hand in three conversations. For both, a review of the last such conversation can suggest something — as described under “Learning from a conversation”: it costs one request and asks first.
Each line offers one step, and none is taken for you. Don’t show again puts a line away on this device; it comes back when the matter itself has changed.
What goes to the provider
The send overview lists the instructions that go along under Instructions: the skill of the conversation, the list of skills the AI may load, and AGENTS.md. Instructions from your vault going to a cloud for the first time bring the overview back. Invisible characters in a skill never reach a model.
Limits of the beta
A skill runs no scripts of its own; for small programs that read your vault, see Scripts. Your own skills are not offered to AI apps connected through the MCP server; only five of the skills that come with Plainva are: Daily orientation, Weekly review, Project status, Meeting preparation and Task triage.