Give it more to work with

A product is a folder for one slice of your business — Checkout, Subscriptions, Onboarding. It collects the initiatives and metrics for that slice. It also carries a workflow, and that workflow decides the columns on your Act board. The second job is the one people miss, so most of this page is about it.

What a product owns

Create a product

1. 1
   Open Products and click New product

   The button sits at the top right of the catalog, and there is a + on the Catalog rail too. Both appear only if you hold the elevated products.edit permission.

2. 2
   Name the slice, not the software

   Checkout reads better than checkout-service. The name shows on every tile and in the workflow picker later.

3. 3
   Set the owner and health

   Owner pre-fills to you, and the server defaults it to you if you clear it. Health is a label you choose — Vernais does not work it out from your data.

4. 4
   Click Add product

   You land on the product's detail page. Status is set to active for you. The product now covers no tools at all.

Map the tools that feed it

1. 1
   Open the tool, not the product

   Integrations, then click a tool, then the Products tab. Its heading reads PRODUCTS FED.

2. 2
   Add the product

   Click Add product. Pick one you already have, or create a new one from inside that dialog.

3. 3
   Enable or pause it

   With the toggle on, the row reads Stripe feeds this product — your tool's name in place of Stripe. Off, it reads Disabled — Stripe paused. The trash icon removes the product from this tool and sends it back to the picker.

Shape the Act board with a workflow

Act works out a board by walking a chain. It takes the task, finds its initiative, finds that initiative's product, and reads that product's workflows. The states become your columns. The arrows become the only moves the board will accept. None of this is visible from the Act tab itself, which is why a board that looks wrong is usually a workflow nobody set.

1. 1
   Open Settings, then Workspace, then Workflows

   Not the product's own Settings tab. The editor lives in the main Settings hub. This is the step most people hunt for and miss.

2. 2
   Pick the product

   The pane leads with a product picker, because workflows live on a product. Each product opens with a starter set: Engineering, Design, Content, Bug triage, and a branching Custom workflow example.

3. 3
   Edit the flow

   Rename it, add or remove states, and draw the arrows on the canvas. Every state carries a status band — To Do, Active or Done — and the system reads that band, never the name you typed.

4. 4
   Close the canvas

   There is no Save button. Every edit persists as you make it, and a Saving… then Saved marker appears in the editor header. Done closes the canvas.

The next Act board that loads for that product's initiatives uses the new states. Work already in flight keeps its column: a task sitting in In Progress stays in an active column instead of being dumped back to the start.

What a product will not do

Four things people expect from a product and do not get. Each one has a real screen that hints otherwise.

Related resources

A sync is a delivery. It drops raw records into your workspace and stops there. Those records are not yet connected to each other, so the AI cannot follow a thread from a Sentry error to the release that caused it. Corvex is the run that builds those connections.

Until Corvex finishes, a new workspace answers nothing.

One constraint shapes most of what follows: when Vernais answers from YOUR CONNECTED DATA, every claim has to trace back to a record it can point at. It will not invent a cause. A question it cannot ground in your data returns nothing rather than a guess. Questions about the outside world, or about general knowledge, it answers like any good assistant — and tells you which world the answer came from.

Run it

1. 1
   Open Integrations

   Left nav → Integrations. The section is called Connect & Sync.

2. 2
   Find the runner

   In the left rail, under Pipeline, click Run Corvex. It is not in the Data section — that catches most people out.

3. 3
   Check the raw count

   The screen shows Raw records at the top. If it reads 0, sync a tool first. The run refuses to start on nothing and tells you there is nothing to process.

4. 4
   Click Run Corvex

   Stages stream in one at a time, each reporting its real numbers. The screen refreshes about every second and a half.

5. 5
   Wait for Pipeline complete

   It takes a few minutes. The run happens on the server, so you can navigate away. A second click while it runs does not start a second run.

6. 6
   Check it worked

   Go to Data. The Nodes table now has rows and the graph draws. Then ask the AI something you already know the answer to.

What the run does to your records

Each stage hands its output to the next. Nothing here is a black box — you can open any result in the Data section and read the source record behind it.

When to run it

• After you connect a tool — sync it first, then run. The Brain works on records, not on credentials.
• In a new workspace — every workspace has its own private graph. Nothing carries over from another one, so a fresh workspace starts genuinely empty.
• After a purge — a purge deletes that tool's records and the graph built from them, while keeping the connection itself. Re-syncing brings the records back. The Brain is what rebuilds the graph over them.

Data can also flow in on its own. Once one full run has completed for the workspace, a sync that brings in new or changed records queues an incremental run by itself. This is on by default. A sync that changed nothing does nothing. See Stream data in live for how that works.

What you see while it runs

The screen is a live receipt, not a spinner. Each stage reports the numbers it actually produced, and four counts sit above them:

• Raw records — what your tools delivered.
• Signals (clean) — what survived de-duplication. These are the records the AI reasons from.
• Duplicates — the copies that were folded away.
• Dedup index — the fingerprints used to spot those copies.

A header line traces the same path in one sentence: raw → dedup index → clean, followed by the duplicate count. The run ends on Pipeline complete or Pipeline failed.

What survives a re-run

The split is simpler than it looks. Everything your tools sent stays exactly as it arrived. Everything the AI worked out gets thrown away and worked out again from those same records.

When the graph stays empty

Related resources

A sync pulls. A webhook pushes. Instead of Vernais asking your tool for new records on a timer, your tool sends each record the moment it happens.

You turn this on per tool, on that tool's own detail page. Vernais mints a private URL and a signing secret. Any sender that can POST JSON can use them — the tool itself, or a small service you own.

Before you start

• Connect the tool first. The Webhooks tab shows Connect first until credentials are saved. A webhook hangs off a connection.
• You need the integrations permission. It is an elevated permission the workspace owner grants per person. Without it the whole Integrations section is hidden, so there is nothing to find.
• Have a sender in mind. Vernais receives. It does not configure the far end for you. Someone has to point a tool or a script at the URL.

Turn on the webhook

1. 1
   Open the Webhooks tab

   Go to Integrations, open the tool, then the Webhooks tab. Before it is on, you see one button: Enable webhook.

2. 2
   Enable it

   Click Enable webhook. Vernais generates an unguessable token and a signing secret. The page flips to the live view: four counters across the top, an ENDPOINT card below.

3. 3
   Copy the URL

   The Webhook URL field holds your address. It reads like https://your-host/api/webhooks/ingest/stripe/<token>. The token in the path is the only thing that identifies your connection, so treat the whole URL as a secret.

4. 4
   Copy the signing secret

   Click the eye icon to reveal Signing secret, then the copy icon. Paste the URL and the secret into your sender's settings.

5. 5
   Check the target stream

   The Target stream counter names where events land — webhook by default. Deliveries show up under that stream on the tool's Events tab.

Sign every delivery

Your sender computes an HMAC of the exact request body, keyed with the secret, and sends it as a header. Vernais recomputes it and compares. Two header names work: X-Atmos-Signature and X-Hub-Signature-256. The value format is identical.

POST /api/webhooks/ingest/stripe/<your-token>
Content-Type: application/json
X-Atmos-Signature: sha256=<hex hmac of the raw body, keyed with your secret>

{"id": "evt_1a2b3c", "type": "charge.failed", "amount": 4900}

Send one record as an object, or several as a JSON array. Vernais keys each record on its id, or failing that its uuid or key, or a hash of the record itself. Resending the same key updates that record instead of duplicating it.

Rotating the secret

Rotate secret issues a fresh secret. It also issues a fresh token, which changes the URL.

Send a test event

Click Send test event. Vernais writes a small synthetic record through the same ingest step a live delivery uses, and logs it. A toast reports Test delivered · 1 event ingested, and a row appears in the history below.

Read the delivery history

RECENT DELIVERIES lists the last 40, newest first: arrival time, status, and the event and byte counts. A rejection carries its reason on the same row.

Make it answerable

A delivery lands as a raw record, and the Events tab shows its status as raw. Raw records are stored, not understood. Corvex is what turns them into the signals, entities and topics the AI reasons over.

Once a full build has finished for the workspace, Vernais folds new records in on its own. A background worker checks every couple of minutes and drains whatever is waiting. Before that first full build, deliveries stack up raw and no question can reach them.

Limits

A webhook fills the gap between syncs. It is not a way to backfill history — a live sync is itself capped, so for anything older, run a sync. See Limits and caps for the rest.

Related resources

Vernais reads your connected tools. It does not read the hallway. When a support lead explains the real cause in a channel, or you write the context up in a page, that knowledge sits outside the graph. Tagging pulls it in. Each tag becomes one signal on the initiative you picked, scored and filed like any other piece of evidence.

Tag a message in live chat

1. 1
   Open the picker

   Type / at the very start of the composer and pick Capture as a signal. Or click the Capture as signal button in the composer toolbar. A search box opens listing your workspace's initiatives.

2. 2
   Pick the initiatives

   Click one or more. Each appears as a removable chip above the input. Use the + button to add another. If the list is empty you get No initiatives yet — create one in the Initiatives section.

3. 3
   Type the message and send

   The message posts to the channel or DM exactly as normal. It also files one signal per tagged initiative. The sent message carries a small target chip for each one, with the initiative code and name.

Send a page into an initiative

A page carries more than a chat line. Use it when you have written the background up properly.

1. 1
   Open the editor menu

   In the page, type / and pick Send page to initiative.

2. 2
   Pick the initiatives

   A picker opens listing the initiatives that are not archived. Choose one or more.

3. 3
   The whole page becomes one signal

   Title and body are captured together as a single signal on each initiative you picked. Not one signal per paragraph — one per initiative. The editor confirms with a count, and tells you if a destination failed.

Find it in the initiative

This is the step most people miss. A tagged signal does not land on the initiative's front page — it lands under two toggles in Identify, both of which default away from it.

1. 1
   Open the Identify tab

   Go to Initiatives, open the initiative, pick Identify and scroll to Signal clusters.

2. 2
   Switch to Post-initiative creation

   The toggle opens on Pre-initiative creation, which shows only what the investigation clustered before this initiative existed. Everything tagged afterwards lives under Post-initiative creation.

3. 3
   Pick the right kind

   Solutions & tasks is selected by default and holds signals that propose a fix. A message that describes what is happening is a statement — click Statements to see it, along with anything judged noise.

4. 4
   Read the two cards

   Related signal clusters and Non-related signal clusters. Your tag is in one of them. Click through to the records table for the full list.

Each signal carries who tagged it. Tagging the same message to three initiatives creates three separate signals — one per initiative, scored separately against each problem.

Related and unrelated

Tagging is not a promise that something counts. Vernais scores every captured signal against that initiative's problem and files it into one of the two buckets, with a 0–10 relation score and a written reason. A tangential message honestly lands in the non-related card rather than padding the case. Related signals are ranked, strongest first.

It judges against the whole problem area, not the narrow root cause. A message about the same funnel counts even when it blames a different cause or proposes a different fix. A message about a different product area does not, however actionable it sounds.

Both buckets are rebuilt from your signals every time you open the initiative. There is no refresh button and no backfill step. A newly tagged message shows up on its own, and the same text always lands in the same bucket.

What a signal can change

Most tags land quietly in a bucket and nothing else moves. Some do more. A signal has to clear two bars first: relevant, and fresh — meaning it adds genuinely new information rather than restating what the initiative already says. Clear both and it can nudge the initiative's confidence, refine the problem statement, or add an insight.

The nudge is bounded. Confidence can move at most 25 points in either direction, and the result is clamped between 0 and 100. One chat message cannot swing an initiative from nothing to certain, and most signals do not move the number at all. The Activity log records what happened: Chat signal added with related — initiative updated., related., or filed as unrelated.

Where tagging silently does nothing

Three failures give you no error at all. They are worth knowing by name.

Related resources

One constraint shapes most of what follows: when Vernais answers from YOUR CONNECTED DATA, every claim has to trace back to a record it can point at. It will not invent a cause. A question it cannot ground in your data returns nothing rather than a guess. Questions about the outside world, or about general knowledge, it answers like any good assistant — and tells you which world the answer came from. A file you attach becomes one of those records.

Attach a file to an initiative

1. 1
   Open the initiative

   Go to Initiatives and open the one the file is about. Click the Files tab.

2. 2
   Drop the file

   Drag it onto the dashed zone, or click the zone to pick one. You can pick several at once.

3. 3
   Wait for the read

   A progress line shows while the server reads the file. A big scan takes longer than a Word doc.

4. 4
   Read the verdict

   The file appears as a card with a badge: ↑ Updated initiative, Fresh evidence, Relevant, Not relevant or Filed. That badge is the AI's judgement of your file against this initiative's problem.

5. 5
   Open the card

   Click it to read the file as a page. If the file moved the initiative, the top of that page says so and lists what changed.

What Vernais does with it

Four things happen to every file, in order. A fifth happens only when the file earns it.

A file is a full signal, not a passive attachment. It joins the clusters, the Correlations graph on the Validate tab and the hypotheses, like any other piece of evidence. See Tag evidence for how a signal is scored.

What it can read

Each file card names the engine that read it, so you can tell an OCR read from a native one. An OCR read of a blurry screenshot is the one worth double-checking.

Four ways evidence arrives

A file is one of four ways evidence reaches an initiative. All four become signals, and all four get scored the same way.

• A file — upload it on the Files tab. This page.
• A page you wrote — in Pages, use Tag to initiative. The whole page goes, not a selection. See Write things down.
• A live-chat message — tag an initiative on the message. See Talk to your team.
• A meeting transcript — it attaches on its own if the meeting was linked to the initiative. See Run a meeting.

Chat attachments are a different thing

A chat attachment goes on one message. Click + in the composer, then Attach file. Vernais reads it with the same engines, hands the text to the model for that turn, and answers your question about it.

Who can see it, and deleting

A file you upload is visible to your whole workspace. Meeting transcripts are the exception: a transcript is restricted to the people who were in the meeting, so a teammate who missed it will not see it in Files at all. Two people can therefore see honestly different file counts on the same initiative.

Related resources

Link the meeting before you run it

A meeting here is a calendar event with video switched on. Add a guest and the event turns into a meeting on its own. There is no separate meetings list.

1. 1
   Open the composer

   Calendar → Create, or click an empty slot on the Day or Week grid. The form opens in the left panel. The Create dropdown also has a Video meeting preset if you want it on from the start.

2. 2
   Add your guests

   Search workspace members. Adding one flips the Video meeting switch on for you.

3. 3
   Link an initiative

   Use the Link to initiative… field. Its owner, assignee and reporter show up as one-tap guests, so you can invite the people who own the problem.

4. 4
   Save

   The event now carries the initiative id. That id is what routes the transcript later.

Answer the invitation

RSVP is not manners here. It is the gate on your own join button.

Open the event and use Going? → Yes, No or Maybe. Only people on the guest list can respond. The organizer is told either way, so a decline is never silent.

Ask for a different time

Guests cannot edit an event. Only the organizer can. Your one lever is Propose a new time.

Pick the slot on the calendar, where you can see everyone's schedule while you choose, then Send proposal. The organizer reviews it on their own grid and picks Accept or Decline. Accepting moves the meeting for everyone and notifies them. Declining tells only you. If you propose twice, the newer one replaces your pending one, and the button reads Propose a different time.

Join the call

Click the event, then Join meeting — or Join now once someone has started it. The room opens inside the app, not in an external tab.

The first person in flips the meeting to live, and everyone else gets a notification to join. Inside you get the participant grid, a docked chat panel, emoji reactions, raise hand, and live captions you can toggle with Captions on.

What keeps the room private

Put a member on the guest list and the door closes: the organizer and the invited members get in, and anyone else is refused. A meeting carrying no member guests is the exception — it stays open to the whole workspace. If you want the room private, invite the people you want in it.

When you join, Vernais mints you a pass for that one room. The pass expires. Its life is tied to that meeting's scheduled end, plus about an hour of slack, and never runs past three hours.

That short life is doing real work. The video service accepts any unexpired pass and has no way to cancel one it already handed out. So the pass is built to die on its own. Ending for all is the other half: it disconnects every person in the room at once.

Record the transcript

Hit Record transcript in the control bar. A red Transcript recording pill appears for every person in the room. There is no silent capture.

Here is the mechanism, because it decides who ends up in your transcript. Each person's own browser turns that person's own speech into text and sends it up. Nothing listens to the room as a whole. That is why the transcript carries real speaker names instead of "Speaker 1".

End it, and the transcript files itself

The organizer clicks End everyone. That ends the meeting and files the transcript. If instead everyone drifts out, the last person to leave files the transcript too — but the meeting stays open, so anyone invited can still rejoin it.

If anything was captured, Vernais assembles the transcript and files it. If nothing was said, nothing is filed. It routes once and only once, so an end plus a stray leave cannot produce two copies. A sweep runs in the background about once a minute and picks up any ended meeting whose transcript never landed.

Where the transcript lands

Every captured meeting writes a page: Pages → Transcript → the date → <meeting title> — <date>. That page is restricted to the invited people, plus anyone who actually spoke in the room.

The page carries a Summary, then Decisions, then Action items as real checkboxes, then the speaker-labelled transcript under a Transcript heading. Each of those sections only appears if the call produced one. If you linked an initiative, a live view of it sits at the top. The model that writes the summary is told to stay faithful to the transcript and to return empty lists when there is nothing to report.

The handoff into an initiative

This is the step the rest of Vernais depends on. If the meeting was linked, the transcript is also filed on that initiative as a signal.

Vernais reads it, judges whether it is actually relevant to that initiative, and files it into the correlations, hypotheses and Decide. It only edits the initiative when it judges the transcript both relevant and new — a call that repeats what is already there is filed as evidence and changes nothing.

Everyone in the meeting gets an inbox note either way. When the meeting was linked and the call produced action items, that note counts how many were routed into the initiative. Otherwise it tells you the transcript is ready.

From there the transcript behaves like any other evidence you hand over. See Tag evidence for the same handoff from a page or a chat message, and Work out the cause for what the initiative does with it.

Related resources

A scrape is a net, not a question. A web search opens two pages, answers you, and keeps nothing. A scrape crawls many sites, pulls out what it finds, and saves every item. You can ask about the collection later.

One constraint shapes most of what follows: when Vernais answers from YOUR CONNECTED DATA, every claim has to trace back to a record it can point at. It will not invent a cause. A question it cannot ground in your data returns nothing rather than a guess. Questions about the outside world, or about general knowledge, it answers like any good assistant — and tells you which world the answer came from.

Start it from the chat composer

The toggle matters more than it looks. With it on, the server takes over and walks you through a fixed flow, one step per turn. Without it, your message falls through to the AI's own improvised version, which asks a third question that no longer does anything. Click + in the composer and pick AI Web scraping, or type /scrape.

1. 1
   Turn on the toggle

   Click + in the chat composer and pick AI Web scraping. A chip appears in the chat bar showing the lane is active.

2. 2
   Name your subject

   Type the topic and send. Name a concrete thing — a brand, a product, and what you want from it.

3. 3
   Answer two cards

   A popup asks how many pages per site, and how broad the search should be. Tap one option on each. Nothing has been scraped yet.

4. 4
   Choose paid or public

   It searches for sources, then shows what it found. A card lists any paid sites by name and asks you to include or skip them. Still nothing has been scraped.

5. 5
   Watch it run

   Once you tap, the scrape starts. A chip appears at the bottom right and a live panel opens on its own.

The two settings

These two cards decide everything about the run. One controls depth inside each site. The other controls how many sites get found in the first place.

Paid and API-gated sources

Some sites block scrapers on purpose. The only way in is their official API, which usually costs money. Vernais separates those out by name and asks you before touching them. Skip is the effective default: only an explicit Include paid sources opts in, so anything else starts a public-only run.

• Blocked without an API — X/Twitter, LinkedIn, Instagram, Facebook, TikTok, YouTube, the App Store, the Play Store, Glassdoor, SurveyMonkey.
• Public and scrapable — Reddit, Trustpilot, G2, Capterra, forums, blogs, news, and the company's own site.
• Why skip works fine — the sources that carry real review text are almost all on the public side.

It never starts in the same turn it searched

This is a deliberate gate, enforced by the server and not by the AI's manners. A start that fires in the same turn as the source search is refused outright. So a reply that says "starting the scrape now" has not started one. The scrape begins on the turn after you tap a choice — never before.

How long it takes

Plan on tens of seconds per page. That is not a setting — it is what the work costs. Each page waits about 8 seconds for a Cloudflare challenge to clear, then a few more for the page to settle. A slow load gets up to 30 seconds before it gives up. Only then does one AI call read the text out of the page.

Watch it, and stop it

A chip appears at the bottom right the moment a scrape starts, and the panel opens on its own. It refreshes about every 1.5 seconds while the run is live. You get the status, the page count, the item count, the scraper's own log lines as they happen, and the items themselves. The panel has a Stop button, and you can also say stop the scrape in chat.

Ask about what it collected

Ask in the same chat. Vernais searches the scraped items by keyword across each item's title, text, and source URL, and returns the best matches with their site. This corpus is kept apart from your connected tools on purpose — a question about your Jira or Stripe data goes to your data, never to the scrape.

To browse without chatting, open the Data section and pick the Scrapes tab in the left rail. The badge shows how many items you have. Filter by target site, search the text, and click through to the original page.

How scraped pages reach the knowledge graph

This happens on its own — you do not run anything. Each scraped item becomes a signal in your workspace's graph, tagged with the integration name web_scrape. It flows through the same pipeline your synced tools use, so it joins the graph beside your Jira tickets and Stripe charges. The panel shows the handoff: Mapping to KG… then Mapped to KG.

Develop your intuition

These are starting points, not rules. When a scrape disappoints, the fix is usually in the request, not the settings. Ask yourself:

• Did I name a subject, or a superlative? "best shirts" finds Best Buy. A brand plus a product plus an intent finds reviews.
• Did I need this saved at all, or did I want an answer this minute?
• Did I turn the toggle on, or did I type the word "scrape" and hope?
• Was the run actually short, or did I check the panel two minutes in?
• Are the sources I wanted on the paid list — and did I skip them without noticing?
• Am I asking the scraped corpus a question my connected tools already answer better?

Related resources

Your words never reach the search engine alone. Vernais hands them to a model first. That model writes several differently-angled searches around your subject, and those angles do most of the work. A vague subject produces vague angles. Vague angles bring back weak sources.

Turn on the scraping toggle first

The toggle forces the scrape flow for that turn. You then get a fixed set of taps: how many pages per site, then how broad the search should be. Nothing crawls until you tap. Without the toggle your message can take an older path that asks an extra question.

Why an adjective drags in junk

A search backend matches words, not meaning. Type best running shoes and best is a word it can match. It finds a store called Best Buy. It finds a dictionary page for the word best. None of that is about shoes.

Vernais defends against this, but only partly. Your raw words are searched too — they are one query among several. The merge takes the top result from every search, then every second result, and interleaves them, so no single search fills the list. Dictionary and definition sites are blocked outright. A model then drops pages that only share a word with your query.

That machinery cleans up after a weak request. It cannot rescue one. The angles are written from your words. If your words name no subject, every angle inherits the hole.

Name a subject the web writes about

Build the request in two parts, and a third when it helps.

• The subject — the brand, product or topic under its real name.
• What you want — reviews, complaints, comparisons, ratings. This tells the model which angles to write, because it adapts to the subject's genre.
• A place word — reddit or trustpilot becomes part of the searches. It is a nudge toward that kind of page, not a filter.

Four requests, rewritten

One site cannot take the whole list

The official site often ranks first on every angle you send. Left alone it would fill the whole list, and you would scrape a company's own marketing instead of what its customers said. The cap pushes it aside. Reddit, Trustpilot, G2, forums and news then have room.

You lose no depth to this. The crawler paginates inside each site it starts on. Six front doors into one site is plenty, and your pages-per-site setting decides how far it walks from each.

What you cannot control

Develop your intuition

These are starting points, not rules. The fastest way to learn what a strong request looks like is to read the sources a weak one found. Then ask yourself:

• Did I name a thing, or describe a feeling about it?
• Would a stranger know what to write about from my words alone?
• Do people publish opinions about this subject, or only my team?
• Did the sources come back from six different sites, or six pages of one site?
• Did I want the web at all, or a number from my own connected tools?

Related resources