You have connected Twain to Claude, ChatGPT, Cursor, or Claude Code. The tools are listed, the sign-in worked. Now what do you actually say?
This is the playbook. Ten moments in an outbound day, the prompt that handles each, what Twain does behind it, and what it costs. Every prompt below is written the way you would say it to a colleague, because that is all the assistant needs: it picks the tools, asks when something is missing, and reports back.
If Twain is not connected yet, start with the setup guide. It takes two minutes.
Before you start
Three things shape every conversation with Twain through an assistant.
- It is your workspace. The assistant sees the same workspaces, agents, campaigns, automations, contacts, and credit balance as the Twain app, nothing more. Anything it generates is stored on the campaign, with a link, so the app remains the place to edit copy, export, or send.
- Reading is free, writing is priced. Listing, inspecting, and reading contacts back never cost credits. Two actions do: generating outreach for a contact and adding a contact to an automation. The assistant confirms the mode and the price the first time you spend in a session, then reuses that choice for the leads that follow instead of asking every time.
- Generation runs as a job. When you ask for outreach, the contact appears in the campaign immediately and the campaign page shows its progress. Base and many High generations finish within a minute; a longer one keeps running in the background and the assistant collects it when you ask.
1. Pick the right campaign
The moment: you just met someone worth reaching out to, and you are not sure which campaign they belong in.
"I just met the VP of Sales at a 150-person fintech. Which of my campaigns fits, and what would one High generation cost there?"
The assistant lists your campaigns, reads their descriptions and agents, and inspects the closest match. You get the campaign name, what it is for, how many steps its sequence has, and the price of one generation at each mode. No credits spent.
This works better when campaigns have descriptive names. "Founders, seed to Series A" gives the assistant something to reason with; "Campaign 3" does not.
2. Generate for one lead
The moment: you have a LinkedIn URL and a reason to write.
"Generate outreach for linkedin.com/in/sample in the Founders campaign, High."
If this is the first spend of the session the assistant confirms the price. Then it adds the contact to the campaign and starts the generation. The contact shows up in Twain right away, with progress, so a teammate looking at the campaign sees it too.
Within a minute for Base and many High runs, the assistant comes back with the messages, the thesis that explains why this person and why now, the mode that ran, and the credits actually charged. Ultra adds a strategy for the whole sequence. If the run takes longer, the assistant tells you it is still working and waits on the job, or picks the result up when you ask, rather than starting a second one.
Anything you add travels with the request. "They spoke about churn at SaaStr last week" ends up in the research and the copy. A work email or a company domain works when you have no LinkedIn URL.
3. Run a batch
The moment: a Slack thread, a spreadsheet, or a conference list hands you fifteen names at once.
"Here are 15 LinkedIn URLs. Generate all of them in the Founders campaign at Base and give me a table of who got what."
The assistant confirms the total, then starts every generation without waiting on each one, up to twenty in parallel. It follows the batch through the campaign's job list and collects the results as they finish. You get a table: contact, status, credits, link.
When you want the leads in Twain but the copy reviewed there first, say so:
"Add these 40 to the Founders campaign, but don't generate yet."
That import is free. The contacts land in the campaign and you generate and review them in the app when you are ready.
4. Check the cost first
The moment: a big list, a deep mode, and a finite balance.
"How much would 25 Ultra generations in the Founders campaign cost, and do I have enough credits?"
The assistant reads the campaign's price at each mode and your current balance, then does the arithmetic. Ultra is priced per sequence step, so the figure is specific to that campaign's length. A campaign that has not generated anything yet has no sequence to price per step; the assistant will tell you the Ultra cost becomes known after the first generation.
Nothing is spent by asking.
5. Read results back
The moment: a colleague asks what went to a contact, or you want to see where a campaign stands.
"Show me what Twain wrote for the contact at Acme in the Founders campaign."
The assistant finds the contact and returns the messages, the thesis, and the research behind them. The same shape a generation returns, for a contact generated last week, in the app, or over the API.
"Which contacts in Founders haven't generated yet?"
A list of the campaign's contacts, whether each has generated, at which mode, with links. Both reads are free, so this is the cheap way to keep a conversation about a campaign grounded in what is actually there.
6. Regenerate deeper
The moment: one of the Base results turned out to be the account that matters.
"Redo the Acme contact in Ultra."
Same contact, deeper run. The assistant regenerates in place and returns the messages with the thesis and the sequence strategy Ultra adds. Regenerating at the mode a contact already has is free; a deeper mode pays that mode's price. Ultra needs the campaign's agent to have a company knowledge base, and the assistant says so if it is missing rather than guessing.
Twain recognises a contact as already in the campaign by their LinkedIn URL or work email, or, when you gave neither, by the company domain together with the name. So name people when all you have is a domain: two named colleagues at one company are two contacts, while a nameless domain-only contact is one per company.
7. Feed an automation
The moment: the campaign already has an automation that researches, writes, and notifies, and you want a lead to go through it without opening the app.
"Add these three to the Q2 SDR automation."
The assistant checks the automation's configured mode and per-contact price, confirms, and queues each contact. The automation's own steps take it from there. Pricing follows the automation, not the mode you would pick by hand, so the assistant quotes the automation's figure.
8. Set up a campaign from chat
The moment: a new segment, a new product line, and no campaign for it yet.
"Set up a campaign for heads of security at mid-size SaaS companies, using my security agent."
The assistant lists your agents to pick the right one, then asks the handful of questions a campaign brief needs: why these accounts now, what you want them to do, which channels, how many steps. It creates the campaign and returns the link. The campaign's messaging is created by its first generation, so the natural next prompt is the one from moment two.
When there is no agent for the segment either:
"Create an agent for acme.com that sells to RevOps leaders at Series B companies."
The assistant builds the agent from the website and the description, and it is usable in the app immediately. Campaign messaging itself cannot be edited through the assistant. After the first generation, the assistant hands you the campaign link to review it in Twain.
9. Combine with other connectors
The moment: the leads are already in another tool your assistant can see.
In Claude, connectors compose. With Slack, HubSpot, or a spreadsheet connector enabled alongside Twain, the assistant can pull the names from one and hand them to the other in a single prompt:
"Take everyone who replied in this Slack thread and add them to the Founders campaign."
"From my HubSpot list of Q3 webinar attendees, generate Base outreach for the first ten in the Events campaign."
Twain does its part the same way as in moments two and three: the contacts appear in the campaign, the generations run as jobs, and the assistant collects the results. The other connector's permissions and limits apply to its half.
10. Send feedback
The moment: something was off, or something is missing.
"Tell Twain the Ultra thesis for the Acme contact missed that they raised a Series C last month."
The assistant passes the note to the Twain team with the context it has. Bug reports, missing capabilities, and rough edges all go the same way, from the chat you are already in.
Flows that build on each other
The ten prompts above are single moves. Most real work is a short sequence where each prompt uses what the last one returned. Five sequences that come up again and again, with what happens at each step.
From nothing to reviewed copy, in one conversation. A new segment, no agent, no campaign.
- "Create an agent for acme.com that sells to RevOps leaders at Series B companies." The agent is built from the website and the description. Free.
- "Set up a campaign for it: heads of RevOps who just posted a sales-ops job. Goal is a 20-minute call. Email intro plus two follow-ups." The assistant asks the two or three brief questions it still needs and returns the campaign link. Free.
- "What would five contacts cost here at High?" The price per mode, from the campaign. Ultra shows as unknown until the first generation exists. Free.
- "Generate these five at High." The first run also creates the campaign's messaging, so it takes a few minutes longer than the others; the assistant starts all five and collects them.
- "Which thesis is the strongest? Upgrade that one to Ultra." The assistant reads the five theses it already has and pays the Ultra price for one.
- "Send me the campaign link for the team." Review, edits and export happen in Twain.
Add what you know, regenerate for free. The first result is fine, but you know something the research missed.
- "Generate for linkedin.com/in/sample in Founders at High."
- "They announced a Series B last week and are hiring ten SDRs. Regenerate with that." The note is stored on the contact and the run is repeated at the same mode, which costs nothing. The new thesis and messages reflect it.
This is the cheapest way to steer a result: state the fact, keep the mode.
Triage a list. Thirty names from an event, most of them strangers.
- "Add these thirty to the Events campaign and generate at Base." The assistant confirms the total, starts the batch and follows it.
- "Which finished, which failed, and why?" The job list, with each failure's reason. The usual one is that the profile, site or email led nowhere; a failed run charges nothing.
- "For the two that failed, here are their work emails. Try again." A second identifier gives the research something to find; the retry is priced as a first generation.
- "Show me the five with the strongest signals and upgrade them to High." Base bought the triage; High is spent only where it pays.
Weekly review, mostly free. Fifteen minutes to see where campaigns stand before spending anything.
- "List my campaigns with how many contacts each has."
- "In Founders, who has not generated yet, and who was generated at Base only?" The contact list carries both facts per contact.
- "Show me the messages for the last three." Read back exactly as generated.
- "Generate the ungenerated ones at High." The only step that costs anything, taken with the picture in front of you.
Hand off and automate. The result needs to reach a colleague or a system.
- "Get the Acme contact in Founders and post the messages and thesis to the sales channel in Slack." With a Slack connector enabled alongside Twain, one prompt does both halves.
- "From now on, add new contacts like this to the Q2 SDR automation instead." The assistant quotes the automation's per-contact price and queues each one; the automation's own steps take it from there.
When something is off. The assistant relays Twain's own message, so the fix is usually in the sentence.
- Still processing after the wait: ask for the status; the job is running and the assistant knows its id. Do not ask to generate again.
- Could not find enough about this contact: give a LinkedIn URL or a work email; a bare company domain gives the research the least to go on.
- Not enough credits, this generation costs X and Y are available: ask for the cost of a cheaper mode, or top up in the app.
- Ultra needs the agent to have a company knowledge base: add it to the agent in Twain, then ask again.
- Anything else that looks wrong: "Tell Twain that…" sends it to the team with the context attached.
Prompts that work
A few habits make every prompt above land the first time.
- Name the campaign. The assistant resolves names to campaigns; you never need an ID. Descriptive campaign names help.
- Say the mode once. Base, High, or Ultra. The assistant reuses your choice for the rest of the session; saying "in Ultra" on a single request overrides it for that one.
- Give the strongest identifier you have. A LinkedIn URL is best, a work email next. With only a company domain, add the name: that pair is what identifies the contact for a later regeneration.
- Ask for the cost before a batch. It is free to ask, and the price of Ultra depends on the campaign.
- Add the why-now. A line about the trigger, the event, or the connection goes into the research and shows up in the copy.
- Do not repeat a generation that is running. Ask for its status instead; the assistant knows which jobs are in flight.
- Edit in Twain. Every result comes with a link. Message edits, exports, and sending happen in the app.
What costs credits
Read-only actions are free: listing workspaces, agents, campaigns, automations, and contacts, inspecting a campaign or automation, reading a contact back, checking a generation's status, and checking your balance.
Generating outreach for a contact is priced by mode, per contact: 1 credit in Base, 2 credits in High, and 8 credits plus 2 per sequence step in Ultra. Regenerating a contact at the mode it already has is free; a deeper mode pays that mode's full price. A contact counts as the same one when it matches by LinkedIn URL, by work email, or by company domain plus name. The assistant reports the credits actually charged with every result, and that figure is the one to trust.
Adding a contact to an automation costs what the automation is configured to spend per contact, which the assistant reads before confirming.
Creating a workspace, an agent, or a campaign is free. The campaign's first generation, which also creates its messaging, is priced like any other generation.
That is the whole product surface, ten prompts long. Same campaigns, same agents, same credits as the web app, from wherever the conversation is already happening.