Have the assistant do something

Say in one sentence what you want done, check the steps it proposes and confirm them.

In short

The assistant has one box for both. A question is answered and never changes anything. A request becomes steps with every value filled in. Two examples are "new card Fix the login redirect" and "run the payments suite and if it fails file a bug on Platform". Telling which is which costs nothing: it is read from the words. If it guesses wrong, "Ask AI to answer it" under a plan and "Plan it as steps" under an answer switch it. The steps happen only when you confirm them.

Before you start

• Who: anyone signed in. Each step needs the same level you would need by hand: Run to run what is saved, Edit to change or schedule it. Do never deletes anything, and sign-in, passwords, access tokens and billing stay yours. A step that sends or invites waits for you to confirm it alone.
• Plan: AI is part of every plan, Free included. On Free the platform's key answers questions about how {{brand}} works. For questions about your own work, add your own key (it is free to add) or choose a paid plan (see /help/ai/answers-left). Reading a request without a model costs nothing on every plan.
• First: only the modules your plan, your role and your switches allow are offered. If your organization switched AI off, Do is off too.

Steps

1. Press "Ask AI" in the top bar. Or press Ctrl+K (⌘K on a Mac), type what you want and choose Do with AI - that is always planned as steps.
2. Type what you want in one sentence, such as new card Fix the login redirect or plan Release review tomorrow 3pm. Or click one of the examples.
3. Press Enter. The steps appear with their values filled in.
4. Check each value and where it came from. Press "Change" to edit; "Save changes" checks them again.
5. Press "Confirm" - with several steps it reads "Confirm all 2 steps" - or "Confirm this step" to run one step at a time.

Result: each step shows "Done", and a message offers "Undo" and "Open". What a step made carries a "Created with AI" mark.

Check and change a step

Each step says what it will do in one sentence, then lists what it will use. Beside each value is where it came from: you said, AI suggested, AI, from the record, from this screen, default, or you changed.

• A step marked "Needs your input" opens its fields at once: something is missing, or a name you used fits nothing, or more than one thing. Pick from the list.
• A record is always shown by its real name, and only if you can open it. A name that fits nothing you can open is answered the same way whether it exists or not.
• When a step touches something else, it says so under the values - which requests a suite will send, and to which hosts.
• Do is told which board or card you have open. With a card open, say "split this card into steps" and it reads that card's description to write them.

Confirm

• Steps run in order. If one fails, the run stops there, what was done stays done, and the step shows "Did not run" with the reason. "Try again" runs it once more.
• Confirming uses none of your AI answers.
• A plan waits ten minutes for you and says until when. After that, ask again.
• A plan is yours alone: it works only for you, in the workspace and the session it was made in. If what you may do changes before you confirm, the plan is refused and you ask again.
• Each step counts against your plan's limits exactly as if you had done it by hand.

Undo, and the Created with AI mark

"Undo" is on the message after a run and on each finished step in the panel. It takes back exactly what that step made - the cards, the item, the note - and nothing else.

• Only you can undo your own step, for 24 hours.
• Undo works while nobody else has changed what the step made. A card someone has since edited, moved or commented on is left alone, and Undo says it changed meanwhile.
• A changed or moved card goes back to what it was. A schedule is removed while it never ran and nobody took it over. A queued UI run is cancelled while it is still queued. Text added to a note is taken off while it is still the note's last part and unchanged.
• Running a request, a suite, a connection's checks or a query cannot be undone: it ran. The plan says so before you confirm.

Anything made this way shows "Created with AI", with who confirmed it and when, to everyone who can open it. This covers a card, a planner item, a note, a request, a monitor or a draft vacancy. A card split into steps shows "Changed with AI". The mark names the step that ran, never what you typed, and the same line appears in Activity.

Two steps in one sentence

A sentence can chain two steps, where the second waits on what the first comes to: "run the payments suite and if it fails file a bug on Platform". It is one plan with one confirmation.

• The first step must come to a verdict: a saved request, a test suite or a Cache Tester connection's checks. A queued UI run and a schedule come to none, so nothing can wait on them.
• The step that waits says when it runs, for example "Only if step 1 fails". If the condition is not met, it shows "Skipped" and why, and nothing is counted against your limits.
• The bug is filed on the board you named - or on the first board connected to the project you may file on - with the failing run as its evidence. If an open bug already follows that failure, nothing new is filed and the step says so.
• What you may do is checked again before every step. If something changed while the first step ran, the second does not run, and the plan says why.
• A chain in the product's own phrases costs nothing on every plan. With a paid plan or your own AI key, the model can also write chains in your own words, up to 8 steps.

Sentences that only open or show something

A sentence that changes nothing is carried out at once, with nothing to confirm and no AI answer spent:

A sentence never shares, grants or sends anything at once: such a step always waits for your confirmation.

What it can do

• Boards: create one or more cards with a column, an assignee, a due date and a priority. Split a card into checklist steps. Change a card's title, assignee, dates or priority. Move a card to a column. Bring a card to another board. File a bug from a failed run.
• Planner: plan a task, an event or a reminder, with a time, an estimate and people; move an item to another time - every occurrence of a repeating one moves.
• Notes: write a note in a notebook, turn the note you have open into a summary note, or add to the end of a note.
• API Tester: save a request with its checks, create a monitor for a saved request, run a saved request as saved, run a test suite, or schedule a suite. The time is read in your own time zone.
• UI Tester: run a ready track as saved. A track that needs a secret value nobody gives is run from its own screen instead.
• Cache Tester: run a connection's saved checks.
• Jobs: draft a vacancy from a sentence. It is always saved as a draft: nothing is published, and nobody is told, until you publish it yourself.
• SQL Tester: run one query that only reads, on an environment of yours. It runs in a transaction that is always rolled back and keeps at most 100 rows.
• Designer: draw a first sketch of an architecture or a workflow - up to 12 parts and their connections, on a plain grid for you to arrange.

Only what your plan, your role and the modules switched on allow is offered. The panel lists it under "See the 9 things it can do here" (the number is how many). Each step needs the same level you would need by hand: Run to run what is saved, Edit to change or schedule it. A run against an environment marked Production is refused at the Run level unless the project or site allows it.

Beyond these, Do can take any action a screen offers you that adds or changes something, with the same confirmation (see /help/ai/ask-ai-to-do-it).

Good to know

• Do never deletes anything: for a delete it opens the record, so you press the button yourself. Sign-in, passwords, two-step sign-in, access tokens, single sign-on and billing stay yours.
• A step runs with your own permissions, in the workspace you are in. What you could not do by hand is not offered.
• A request in the product's own phrases is read without AI and uses none of your answers. The examples that work this way are marked "No AI needed". Anything else is read by the model, so write it as you would say it.
• A request can be up to 4,000 characters, and one request becomes at most 8 steps. One person can ask for at most 10 plans a minute.
• A request or a monitor made this way must point at a public address; a private network address is refused, as everywhere in API Tester. A monitor made this way alerts by email only.
• A query runs only when it is a single SELECT statement. Anything that writes, locks, waits or calls out of the database is refused in the preview, whoever wrote it.
• When a word could mean two things - "move" is both a card and a Planner item - the plan asks which you meant.

What it costs

• A request read without AI costs nothing and works on every plan. On the Free plan, a request only a model can read needs your own key - it is free to add - or a paid plan.
• A request the model reads uses one of your AI answers - two or three when it also reads the record you have open. The plan says how many it used and how many are left.
• During a trial, before the platform's AI reads a sentence, the box says how many answers it would use and waits for "Use one answer".
• A request for several steps needs a stronger model. With your own key it always works; on the platform's key it depends on your plan. Without it you get the first step, and the panel says so.

When your organization limits AI

• With AI switched off for the organization, Do is off too - even for requests that need no model.
• A module the organization keeps away from AI offers no steps.
• With personal details masked, they are masked in what the model reads here as well.
• Where the organization chose that only its own key, or only the platform's AI, answers, your own key is not used for Do there either. See /help/ai/what-the-assistant-reads#organization-settings.

If something goes wrong

• "Only the first step is shown. Plans of several steps run on your own AI key (Settings → AI) or on a paid plan." To fix it, add your own key. Or confirm the first step and ask for the next one separately.
• "You can no longer do this here: the module is switched off for you, or your role does not allow it." Your access changed after the plan was made. Ask someone who can, or ask again in another workspace.
• "Nothing can be done with AI in this workspace right now": the modules that take actions are switched off for you, or your role does not allow them.
• A step shows "Did not run": read its reason, fix it (for example a value), and press "Try again".
• The plan expired after ten minutes: ask again.

Related

• /help/ai/asking-the-assistant
• /help/ai/why-did-this-fail
• /help/ai/answers-left