Skills
Procedures your agents learn by themselves, which you can teach, improve or undo.
A skill is a procedure with steps that your agents know how to do. When a job goes well, the agent can keep it as a skill; from their next turn, every agent in the business uses it.
See your skills
- In the sidebar, press «Skills».
- You see each agent's skills and those of the «Whole business», with each skill: who learned it («learned on its own by …»), how often it held up or failed, and its pitfalls.
- Open one to read or edit it.
- Every change keeps the text it replaced: under «Earlier versions» you bring any of them back.
Lock a skill
Press the padlock on a skill's row, here or in «Usage» (type /usage in a chat). While it is «Locked», the app refuses every change that is not your own click in its window: the assistant does not rewrite or delete it, what it learns by itself leaves it alone, an import does not replace it, and nobody else turns it off, shares it or moves it. You can still edit it yourself. The same padlock unlocks it.
«Usage» also shows what each skill takes in «This conversation» and in «Last 7 days», and how many times it changed by itself, so you know which one to lock or turn off.
When an agent learns
You see a notice like «Learned: …» or «Improved: …», with «View» and «Undo».
- «Undo» on a new skill sends it to the Skills trash, where you can restore it.
- «Undo» on an improvement puts back the previous version.
- If the skill holds private data, only the agent that learned it uses it.
- If a task goes wrong, the agent can learn from the failure: «Learned from a failure: …».
- If a card on the Task Board stops twice for the same reason, the head of its department writes the lesson as a skill of that department; see Task Board.
Who chooses a skill
By default the assistant can choose a skill by itself when it fits the task. For a skill you want to start only yourself (a deploy, a release), use its icon in the skills list: «only I start the skill; the assistant does not choose it». From then on you start it by typing / and its name, and your own start lets the assistant read it for that chat. «let the assistant choose the skill too» gives it back.
Skills that stop working
A skill the app learned that keeps failing for two weeks is set aside by itself, so it stops getting in the way; it goes to the skills trash like an «Undo». Skills you wrote or imported are never set aside.
New skills at the places you imported from
If the tool you imported your skills from has new or changed ones, the app tells you («Skills of … are new or changed. See them in Settings › Import.»). They are ticked in «Import»; press «Import» and each one is checked first. You can also import the skills a web address lists, when its list is a skills list.
When an agent meets a program it does not know
Instead of guessing a command's options from memory, an agent can read the program's own help first. Where your permission level asks, you get a card like «Sales wants to run «git» on your Mac to read its own help, and keep it as a skill.»
- It reads only what the program says about itself: its version, how to use it, its subcommands and options, and its manual page. It runs in a closed room with no internet, it can only write in a scratch folder, and it installs nothing.
- If the program is not there, it says so and suggests names that look alike.
- What it read is kept as a skill for that program. After the first time the agent uses the program well, the skill also keeps what worked and the mistakes it fixed, so the next time it goes straight to it. Passwords and secrets are never kept in it.
- Every time the program is used, the skill checks its version. If the program was updated, the skill says it may be out of date and the help is read again the next time.
- What the program says about itself is only information: the agent never takes it as an order.
The skills the Assistant brings
The «Assistant» (see Nobreiniers) arrives with skills of its own, which you see under its name in «Skills» like any other:
investigador-web: how to research the web well, searching, reading the sources themselves and confirming what matters in a second place. See Web research.informe-de-consultoria: a report that ends in a recommendation, such as a market study, a strategy or a diagnosis. It fixes the decision, the reader and the scope first; then builds the framework of questions and hypotheses; only then collects data, draws the charts and writes the report, answer first, as a document or a deck, with the data workbook it came from.investigar-repositorio-github: how a code project is built, how healthy it is and what is worth taking from it. It reads the code without running it, pins everything to one commit, and the report gives a verdict, the architecture, the risks and what it did not examine, with a file and line behind every claim.revision-critica-de-articulo: whether a paper's claims follow from its evidence. It reads the whole paper, not the abstract, separates the flaws that undo the conclusion from the ones that only limit it, recomputes before it objects, and can write a referee report.revision-sistematica-de-literatura: the state of the research on a question. It fixes the question and the rules before searching, writes the search down so someone else could repeat it, counts what it left out, and every reference it cites is checked against its real data, in APA, IEEE or BibTeX.herramientas-a-medida: how it makes you a small tool that stays in the app and keeps its own data, such as a price calculator, an order pad or an inventory. See Tools made for you.margen-por-producto: how much each product leaves you. Drop your sales file (with the costs, or tell it the costs) and ask «how much margin does each product leave me?»; you get a spreadsheet on your Mac with your data and a summary by product (income, cost, margin and margin %) made with formulas, so if you change a cost the margin moves. It is ranked, with a chart whose title says the conclusion. If a product has no cost, it asks you, once and together with whether your prices include tax and what a return costs; it never invents one. Costs shared by everything (rent, salaries) are shown on their own line and are not spread over the products without telling you. If you want, and only if you say yes, it repeats the check every month on the new file and tells you only which products lose money and what price would fix it.leer-contratos: what a contract says. Drop a contract (a PDF, a scan or a document) and you get who signs, what it is for, how long it lasts and how it renews, the amounts and how they can change, the penalties, who pays what, how to end it and the notice needed, which courts decide, and what is unusual or missing. Every date and every amount comes with its page (its clause, in a document without pages) and its own words, and was checked against the document; what the contract does not say is written «no lo dice», never guessed. It is not legal advice: it ends with what to ask a lawyer, in plain words. If you want, and only if you say yes, it sets a reminder for each important date, such as the last day to give notice.el-tiempo: the weather where you are, or in any place you name, without your saying where. It answers with a card: now, the next 24 hours' high, low and rain, and each day of the week, in your language and your country's units («what will the weather be this week in Santo Domingo?»). The forecast comes from a public weather service and the place from a public map, and only the place's name leaves your Mac. Most places have no chance-of-rain percentage: the card then shows the rain in millimetres or inches and never makes up a percentage. Today's row covers only the hours still to come. If the card cannot be made, it finds a current weather page, reads it and answers in one line: the place, the day, what the page says and a link to the page. It never adds a forecast the source did not state. The answer names the city it used: if that city is only a first guess from your connection, it asks you to confirm it, and if it has none it asks once. If you say yes, it can also look every morning, for a city you gave, and tell you only when rain or a warning is forecast.buscar-lugares: real places near a point, such as «bakeries within 3 km». For «near me» it uses the area this Mac is in, rounded to about 1 km, if macOS allows it (macOS asks you once, the first time you ask in the app's window on this Mac); otherwise, or when you write from your phone or another device, or a workflow asks, the place in «Settings» › «Profile», and it tells you which one it used. The area goes into that chat like any answer; it is never saved to memory or to a place. You can name places once («this is home», «save the office as …») and say «near home» afterwards: they stay on this Mac, only in your own conversations, and you can ask for the list or forget one any time. The answer is a card with a map link for each place and says which source it used. It looks at a map or directory page in its own browser, checks each place on its own page and gives address, phone and hours; hours the page did not show are marked «hours not confirmed», and «open now» is said only when the page said it. It does not fill your «Clientes» list on its own: it offers once to save them and saves only the ones you say yes to; if you already asked to keep them («save them», «find me leads»), it saves them at once. Each saved place carries the page it came from and the day. Places along your route for the day come in a later version.vigilar-precio: watches the price of a product or a competitor's on a page you give it. It keeps a workbook with one dated row per change (date, old price, new price, source) and tells you only when the price changes or cannot be read. A price that could be read two ways is never guessed, and the currency stays as the page shows it. Ask «should I buy now or wait?» and it answers from that history only when there is enough of it, and says how much it has.
In the research and review skills and in these two, what is reported comes from what was read, recomputed or counted: nothing is filled in from memory, and what could not be found is written as not available. Type / and a skill's name to start one yourself, or just ask for the work and the Assistant picks the skill that fits. The weather, places and price skills work the same way: they report only what a page showed, with its link.
For big decisions: decidir-con-numeros
Every Nobreinier brings the skill decidir-con-numeros, which you see under its name in «Skills». It is for a decision that, if wrong, costs months: a price, a hire, a large purchase, a new product or service, closing a line or changing supplier. It is not used for a small errand; if you ask, or if the decision is that big, the Nobreinier follows it:
- It states the decision in five lines: what is decided, the goal (with a number and a date if it has one), how it is done today, the constraints and who decides. If a wrong guess would be costly, it asks you once, all together.
- It lists what today's way takes for granted and marks each assumption: «fact» (checked, with its source), «estimate» (a number it worked out, with how), «habit» («it has always been done this way») or «opinion» (someone's, named). Habits and opinions are where a better plan usually hides.
- It breaks the problem into its parts (money, time, people, materials, tools, the customers' attention) and finds the real minimum of each, with source, date and unit. If a part has no hard minimum, it says so instead of inventing one.
- It puts what each part costs today next to its minimum and ranks the gaps: the biggest are the opportunity.
- It designs two options from the parts, not from how it is done today: a bold one (the biggest gain if it works) and a safe one (less gain, hardly any risk).
- It attacks each option: what breaks first, what a skeptic would say and the cheapest test before committing (a call, a quote, a week's trial).
- It checks every number against its source and labels it «PASS», «FAIL» or «BLOCKED» (the source does not open or data is missing). It never invents a cost or a figure; if something is missing, it stops and says what.
- It hands you one document with the verdict first: the recommendation and this week's first step, the gap table and the two options, the assumptions with their sources and dates, and a 30-day plan with who does what and how success is measured.
Sending, publishing, spending money or using private data still goes through your approval, as always: this skill only prepares the decision.
Teach on purpose
- Type
/aprendeand the name of an app or a site (/aprende QuickBooks, say) to have it learn to use it whole: every screen, button, box and function. Each function is kept as a recipe, so next time it does it in one step. In your real account it fills everything and stops before saving, sending or paying, unless you ask it to do it for real./learn-skilldoes the same. - When it checks what it learned, it works only in the windows it opened for that, never in your own: it types, clicks and closes nothing there. If you start using the keyboard or the mouse, it stops at once and closes nothing.
- Type
/learn how to …in the composer to have it keep a procedure. - Type
/refineand a skill's name to improve it. - Every skill also works as a
/command.
Save a process of yours
After your assistant does something you will want again, press «+» in the message box and choose «Save as process» («Keeps what you just had done under a name, to repeat it whenever you want»). The box fills with /process. Type a name after it, for example /process month-end close, and send. /proceso works too, and both are in the / menu.
- The work is learned as a skill with your name, marked as your process, and the answer says where to run it again.
- From then on the «+» menu shows «My processes», only once there is one, with «Saved: 2». Each row runs that process again in this chat.
- With no name, it asks you for one. If a skill already has that name, it asks for another, or for you to improve that one with /refine and its name.
- If a skill you already have covers this work, it keeps that skill and lists it as your process.
What you teach helps everyone
What your agents learn and what you teach them (skills, recipes and how-to) leaves your Mac as soon as it is written, without waiting for the next day. The rest, such as the shape of the sites they used and where tasks got stuck, leaves once a day. Nothing about you or your company ever leaves: before anything taught goes up, the app takes out by itself the names of people, accounts, pages and groups, and holds back anything that still looks personal. With that, every update carries everything that was learned, and every copy of the app gets smarter: the new or improved skills and sites that other users shared arrive with the update and your agents use them as the app's own, with nothing for you to do. Yours and your business's always come first: if one of your skills has the same name, yours wins.
It is on; «Share what the app learns and is taught (without your data)», under «Settings» › «Privacy & diagnostics», turns it off and shows what is sent. See Privacy and storage.
