Draft a workflow from a description
On the Describe tab, write what should happen — the trigger, the steps, the apps involved — or pick a starter brief, then press Enter or select Generate workflow. The Builder opens straight away on a new private draft, with the workflow assistant beside it, and the draft fills in there as Hive writes it rather than holding you on the form. The draft records where it came from: a built-in template or the model, and which model. Hive checks every draft before it shows it to you: one trigger, every step connected, nothing that loops back on itself, only steps your workspace can run, and every app step given the details it needs (such as who a draft email goes to) from an earlier step when one can supply them. When a check fails, the assistant is told exactly what is wrong and fixes it, up to three times. If one step still cannot be built, Hive keeps the rest of the draft and says what it left out, so you can add it yourself. Hive also checks the draft against what you asked for. A workflow you described as running on a schedule gets a schedule trigger, “send” or “email me” uses your connected mail app’s send step rather than a draft, an approval sits only where a step reaches someone outside your workspace or cannot be undone, and an hourly or app-event workflow does not repeat what an earlier run already did. The assistant fixes these once; anything it leaves is listed beside the draft for you to check. The draft opens with its plan: what Hive took from your description, what it assumed, the steps to check before you deploy and, only where a wrong guess would be costly, up to three questions. Where a workflow sends is yours to type. Neither Describe nor the workflow assistant writes an email address, phone number, link, channel or repository that is not already saved in the workflow, even one your description or a file you attached gives. The step is drafted with that field empty and listed as an open item; type the value onto the canvas yourself. “Email me” still works: your own connected mailbox’s address always counts. Apply refuses a proposal that adds one (“This change names , which is not saved in this workflow. Add it on the canvas yourself.”), and so does Describe (“This draft names , which is not saved in Hive yet. Describe the workflow without it, then add it on the canvas.”). The same goes for a step that would send wherever an AI step’s answer says: Describe leaves that destination empty, and Apply refuses it with “This change sends where an AI step’s answer says, which is not saved in this workflow. Add where it sends on the canvas yourself.” Neither will add or change an Agent step whose answer a step sends to; you can still make that change on the canvas. An AI client can write in these chats too, so only what you type on the canvas decides where a workflow sends. See AI clients and drafts. If Hive cannot draft it at all — no AI model configured, a description it could not turn into a workflow, or drafting took too long — it says why in the assistant panel and on the Describe tab, with Try again. Nothing is created. You can also reword the brief and send it again without leaving the canvas. Your original description and any files you attached stay on the Describe tab too.Send with Enter
In the Describe box, Enter sends your description and Shift+Enter starts a new line, as in the workflow assistant’s chat. Ctrl, Cmd or Alt with Enter sends too. When you type with an input method, such as for Chinese, Japanese or Korean, Enter first confirms the word you are composing. Enter does nothing while there is nothing to send (no description and no attached file), while attached files are still being read, or while a draft is already being written. As you type an app’s name, a list of matching apps opens under the box. Nothing in it is chosen until you press the Down or Up arrow: then Enter or Tab puts the highlighted app’s name in your description. If you have not used the arrows, Enter sends your description exactly as you typed it. Escape closes the list.Connect an app from Describe
Under the box, Your connected apps shows the apps you have connected, with their logos. Select one to add “using” and its name to your description. With nothing connected yet, a few common apps appear greyed under Connect an app. The starter briefs list the templates you can use straight away first. If your description names an app you have not connected, a note under the box says so, with Connect and the app’s name. You can still send: the assistant drafts without that app. Selecting an app under Connect an app, or Connect in the note, connects it without leaving Describe: the app’s sign-in opens, and your description, attached files and chosen model stay as they are. While it connects, the app reads, for example, Connecting Outlook… and the other apps wait. Once you have signed in, the cursor goes back to your description. If the sign-in fails or you close it, nothing is lost and you can try again. An app that needs more details before you sign in, such as a region, opens Integrations in a new tab, already searched for that app. Connect it there, then come back to the Describe tab: your description is still there, and the app is listed once Hive has read your accounts again. If Hive cannot read the list of apps, the area says so with Retry. If a step in the new draft uses an app you have connected with more than one account, the Set up guide opens and asks which account each step should use. You can also ask for a workflow in Ask (“when a new HubSpot deal is created, research the company and post to Slack”). The assistant drafts it on the model you chose for that chat, and Hive shows it as a card with its trigger, steps and apps. Nothing is created until you press Create. See Build from chat.Speak instead of typing
The Describe box and the workflow assistant both have a microphone. Select it and speak: each phrase is added to the box when you pause, after anything already there. While you record, the microphone becomes a recording bar with a level meter and the time so far. Select the tick (Insert text) to stop and keep what you said, or the cross (Discard recording) to put the box back as it was. On Business, dictation uses your own AI credits, at what the transcription costs Hive; on an earlier plan, each Pro or Max seat adds 5 hours a month to the workspace, and after that dictation uses 1 credit a minute.The plan beside a draft
A draft from a description opens with The plan under the assistant’s first message. It lists what the workflow is for, what starts it, what it reads, what it does, who it reaches, what happens in an unusual case and what happens if a step fails — each tagged with where it came from:
Steps to check before you deploy lists every step that Sends a message to someone other
than you (an email, a forward, an emailed invoice, a chat or social post, a comment on an issue or
ticket), Can’t be undone (a refund, a charge, removing someone’s access, or a step Hive’s
catalogue says the app offers no way back from), or Changes data in an app. Hive sets these
labels from the steps themselves, not from what the AI said about them.
Drafts, internal notifications, Hive documents and your own Docs or calendar entries carry no label.
Questions for you appears only when a choice your description left open is costly to get wrong:
who is emailed, what is deleted, money, or whether something is sent outside your workspace. There
are at most three questions, all at once, each with a short reason and two to four answers, the
recommended one first. A cheap choice — a channel, a time, a folder — is never a question: Hive
takes a default and tags it Default.
The draft does not wait for your answers. It is already built on the recommended answer to each
question, so you can deploy it as it is. To choose differently, select an answer under each
question, then Use these answers: the answers go into the message box, and when you send them
the assistant proposes the change for you to apply. An answer that names an email address, phone
number, link or
#channel not already saved in the workflow is refused at Apply: type it onto
the canvas yourself. A workflow drafted from a built-in template,
and any workflow drafted before the plan existed, has no plan beyond its step labels. Neither does
a draft Hive kept differently from how the AI planned it — with steps left out, a loop capped, or
cut at the two-minute limit — because what the AI planned may name steps that are no longer there.
While the assistant works
Drafting a workflow usually takes under a minute, and Hive stops a draft that runs past two minutes. Answering a message can take a little longer. Hive takes the request at once and the assistant works on it in the background. While it drafts a new workflow, the Describe tab and the assistant show what it is doing: planning the steps, checking them, or fixing what the check found. In the workflow assistant, your message leaves the message box and appears in the conversation as soon as you send it; if it cannot be sent, the assistant says why and your message goes back into the box. The Describe tab or the assistant shows that it is working, and the draft or the reply appears on its own when it is ready. You can keep the page open or come back to it — a reply is saved in the conversation either way. If the assistant cannot finish, it says why, and nothing is created or changed. A request that has not finished after 30 minutes is reported as not finished; send it again.Which model drafts
Planning a whole workflow needs a capable model, so the assistant drafts on your workspace’s strongest allowed model: the Powerful model an admin chose for Auto, or else the workspace’s default model, or else Hive’s default — as long as it is not an Efficient model and you are allowed to use it. If your workspace allows none of these, it drafts on Hive’s Efficient model: your workspace’s Background model if an admin set one, or else Hive’s default Efficient model. Where your workspace lets you choose, the Describe composer and the workflow assistant have a model picker, and leaving it on your default drafts as above; otherwise they name the model instead. Drafting uses your own AI credits (on an earlier plan, the workspace’s shared pool, billed to you), at the rates of the model that drafts. If your workspace allows no model the assistant can use, it says so and drafts nothing. On workspaces where the assistant is not available on this runtime, it uses the deployment’s AI model, which the composer names.Open the workflow assistant
In the builder, select the Hive button at the edge of the canvas (“Ask Hive to change this workflow”), or Describe it to Hive under the next step in the header of an empty draft. The panel is called Workflow assistant. On a new, empty workflow it reads Create a new workflow; on an existing one, Change this workflow. It shows the model the assistant uses, or No AI model configured. Only members who can edit the workflow can ask it to change it. What it offers before you type depends on the workflow:- an empty workflow offers starter briefs from the built-in templates, and Import a workflow;
- a workflow with steps offers suggested changes instead: Add an approval before sending, Run it weekly instead and Explain what this does.
Change a workflow
Describe what should be different — or just say what is wrong. You do not have to work out the edit yourself: tell the assistant the problem (“it never emails anyone”), the goal (“I want to hear about these before they go stale”) or what a run did, and it works out what you need and answers with the smallest change that gets it, saying what it changed and why. Ask a plain question and you get a plain answer, with nothing proposed. Nothing is applied for you. When the assistant proposes a change, the reply shows a Proposed change: a summary (up to 500 characters) and a preview of the steps it adds, removes or edits. When the change leaves some of your steps as they are, and connects the steps it keeps in the same order as before, the preview ends with “Nothing else changed.” A change that also reconnects or reorders steps it does not edit, such as moving an approval after the send it guards, does not end with that line: check the canvas before you apply it.- Apply proposed change saves your pending edits, then applies the proposal. If the workflow changed after the proposal was made, Hive reports the conflict instead of overwriting the newer version. A note under the button warns: “Applying this change clears Undo. You can’t undo the change or any earlier edit.” After you apply, Undo on the canvas starts afresh: to go back, ask the assistant for another change, or edit the steps yourself.
- Dismiss leaves the workflow as it is.
How the assistant designs a workflow
The assistant builds for the result you are after, not for the sentence you typed: “email me new leads” is asked for because no lead should go cold, and that decides what the workflow watches and what its last step does. Before it picks a step it settles four things — what starts the workflow, what it reads, where the result lands and who acts on it — and its description says what it settled. Where your request leaves one of them open it builds the reading your words best support and names the other one, rather than inventing a recipient or a destination. What that gives you in a draft:- The fewest steps that do the job. It does not add a step you did not ask for, and does not use two steps where one does the same work.
- Nothing sent on an empty run. Before the first step that sends, writes or asks a person, it adds a rule Condition that passes only when the read found something, so a quiet morning ends quietly instead of sending an empty message.
- A scheduled run that does not repeat yesterday’s work. Nothing in Hive remembers which records an earlier run acted on, so before a step that writes outside Hive it bounds the read to what is new — a date window matching the schedule, or a status the workflow itself changes. Where nothing can bound it, it uses a draft or an approval instead of sending directly.
- An approval in front of risk, by default. It puts an Approval before the first step anyone outside your workspace sees — an email, a message, a post, a write into a connected app — before a step that moves money or removes access, and before a Loop whose steps do any of those, where it guards the whole loop from outside. You do not have to ask for one, and you can say you do not want one. Where a step can hand a person a draft to send instead of sending by itself, it prefers the draft: you still decide, and the run never waits. In a workflow started by a public form, whose values are typed by an anonymous visitor, it guards a write into a connected app the same way.
- A result in counts. A workflow that works through a collection, or reports to a person, ends by saying how many were handled and how many need you — not by listing the steps that ran.
- Every guess, said out loud. Your request will not state everything the workflow needs — which channel, what time of day, which filter, which of your connected apps stood in for one you named but have not connected. The assistant chooses something sensible rather than stopping to ask, and then names each choice it made in the note it writes beside the draft and tags it in the plan, so you can correct it in seconds instead of discovering it on the first run. Only a costly choice becomes one of up to three questions, and even then the draft is built on the recommended answer.
- Steps named for what they do. “Email the client the overdue invoice”, not “Gmail 2” or “Step 3”. A step left with a generic name is listed beside the draft for you to rename.
- Short messages with one next step. A message a step writes to a person says why it was sent and ends with one clear thing to do.
Built from your connected apps
The assistant builds the workflow from the apps you have connected. It knows the steps each one offers in Add step, the reads that fetch records and the actions that change something, and the fields each step needs. It uses a Read from an app step to fetch records on a fixed rule, and an agent reading the app when the step needs judgement, and keeps a workflow within 50 reads a run. The apps you list in your preferences come first, then the apps your description names, then the rest. To read from or act in an app, it uses a step from one of those apps before it reaches for a general AI or notification step. It never uses an app you have not connected. If your description needs one, it drafts the closest workflow it can from your connected apps and Hive’s own steps, and its summary names the app to connect. A preferred tool you have not connected is mentioned there too. When it drafts a Connected app event trigger, it filters on the event types Hive knows that app sends — the same ones the trigger offers under Suggested: — rather than a type it made up, which would pass every check and then never fire. Where Hive knows no event types for an app, it has none to work from, so check the trigger before you deploy. For a workflow that starts when an email arrives in Outlook, it usesmessage.created
only, and reads the email’s subject, sender and body from the event (see
Connected app event).
When a step needs a value your description does not give, such as which Slack channel or which
contact, the assistant leaves that field empty rather than guessing, and its summary says what to
fill in. Where it does have to settle something you never stated — a schedule, a filter, a
threshold, a channel, or an app standing in for one you named but have not connected — its reply
in the assistant names it, rather than letting it pass as something you asked for.
To tell a person something, it uses a step that delivers: a message to a channel your connected
chat tool posts to, a direct message, or a draft addressed to them for a person to send. It does
not use an Internal notification for that — see
Internal notification stays inside Hive.
A message, email or draft it writes over several lines keeps its line breaks; it never shows a
literal \n.
What the assistant checks
Every draft and proposal is checked the same way a deploy is: the steps must be valid, allowed for your role and runnable in your workspace. When a draft fails that check the assistant gets one more attempt; if that fails too, nothing is created or changed and the assistant says why. The assistant checks its own draft before it answers. It can run the same check on a draft as often as it needs, and a draft that fails is sent back with every reason, up to five at a time. A step that names an action its app does not have comes back with the actions that app does offer. The check also covers what a deploy would need from your accounts: every step must use an app you have connected, and so must a Connected app event trigger the assistant adds to a workflow you own. A trigger on a workflow someone else owns listens to their apps, so the assistant leaves it as it is. It fixes the draft within a few attempts, or nothing is created or changed. When you apply a proposed change, each step that stays on the same app keeps the account you chose for it. The assistant never picks the account for a Send message, direct message or email step: it keeps the one the step already uses, and a new step uses the account chosen in its settings. When the assistant drafts a loop that could exceed the 500-step run budget, it lowers the loop’s Max iterations to fit and says so, for example “Capped the loop at 166 items so a run stays within Hive’s step limit.” A workflow you build by hand is refused instead. The assistant keeps your Agent library steps, and the connections into and out of them, when it proposes a change. It can move or remove one when you ask, but it never adds one or changes which agent it runs or what it is given. Add those yourself from Add step.Conversations
Each workflow keeps its Original conversation. New conversation starts another, created when you send its first message. Chat history lists earlier ones (up to 50 per workflow). The assistant reads the recent turns of the conversation you are writing in, so you can refine a draft across several messages without restating what you already said, and it will not ask again for something you have answered. It reads that conversation only — another conversation on the same workflow is separate, which is what New conversation is for when you want a clean start. Where a new message contradicts an earlier one, the new message wins. Only the most recent turns travel with each message, so a very long conversation eventually drops its oldest part. It is also told how the workflow last finished, and whether that was a production run or a test. If you say something is not working and the last run failed, the assistant reads that failure — which step, and why — and proposes the fix rather than asking you what happened. It is told the shape of the failure, not the records a run read or wrote: the step that failed, how Hive classified it, and the message. Where the failure came from a connected app, that message is the app’s own words, and an app often quotes back the value it rejected — so it can repeat something that reached your workflow from outside. It does not open files in your connected apps, and it does not read what a past run found. After it drafts a workflow or proposes a change, Hive checks the result itself: required fields still empty, a read that may run too many times, and a loop it had to cap. It says what is still missing in the reply, and it does not fill those fields in or apply the change until you accept it.Attach files
Select Add files to send files with a message:
Reading attachments needs a configured AI model. Hive reads the text of the files you attach and
sends that text to the assistant with your message. Attaching a file does not let the assistant
open files in Drive or other connected apps.
Preferences
Chat preferences (the sliders icon in the assistant’s header) opens Workflow Chat Preferences: list your Preferred Tools per category and add Custom Instructions (up to 2,000 characters) about behaviour, style and tone. Only your own messages use your preferences. Until you save preferences, a new conversation also shows Customize how workflows are built for you with Set up above the message box. A preferred tool counts when it names an app you have connected: the assistant then builds with that app first. A preferred tool you have not connected is never used. The assistant mentions it in its summary instead.Without an AI model
If no AI model is configured for your workspace, the assistant can only start a draft from a built-in template whose starter brief you send unchanged (case and spacing do not matter). The assistant says so before you type: “Typing your own request needs an AI model — ask an admin. Starters above still work.” Send stays unavailable until the box holds an unchanged starter. Starters appear only on an empty workflow: the suggested changes on a workflow that already has steps need a model, like anything you type. The reply reads “I prepared the … template as a proposed change.”Related
Build a workflow
Other ways to start, and editing the canvas.
Templates
The built-in starting points.
Models
How your workspace chooses AI models.
Test and deploy
Check the draft before it goes live.