Getting Started with Claude
Practical AI for the work you actually do. Three workflows built live in 60 minutes, the prompts behind each one, and the skills and scheduled tasks that turn a good prompt into something that runs without you.
Three workflows, run end to end.
Rather than tour the tool, we picked three jobs people actually do every week and built each one live, showing the mechanics inside it. Every prompt below is the one we ran in the room. Copy them as they are, or edit the bracketed parts and go.
The prompts, the skills, and the dashboard.
Every prompt we ran in the room, the scheduled task that runs the follow-up unattended, the five skills behind the workflows, and the sample NPS dashboard. Plus the occasional note from Momentum on what we're building next.
Founder prep: research and call prep on a template.
Three principles carry almost anywhere, not just founder prep. Bring any input: recorded transcripts, typed notes, even a photo of handwriting. Combine internal and external: start from what you already know, then have Claude research everything you don't. Output to your template: a format you define once, so every result comes back in the same shape.
We ran it live in the room against a real company, so the research was genuine rather than invented. The input was a set of light discovery-call notes, the kind you already have after a first conversation. The output was a sourced one-pager in a fixed template.
The difference between this and reformatting an intake: the notes said "most teams." The research came back with the actual number, the real pricing, and a production-capacity risk the notes never mentioned.
From call notes to a sourced one-pager.
Start with your notes and your format
Both files go in before you type anything. The template is what stops you reformatting the output by hand every time.
Research everything else, and combine it
"Where our notes and public info disagree or leave gaps, flag it" is the line people skip, and it is where the value is. It surfaces the places your notes and the public record don't match, before the call rather than during it.
See the one-pager it produced
Joshua Pope (CEO) and Rowan Ferrabee (CTO), both UWaterloo engineering. Founded 2019. Hardware-plus-subscription scale-up selling the Trajekt Arc, a pitching robot that recreates any real pitcher for hitters to train against, now close to standard equipment in Major League Baseball.
Implied revenue is arithmetic from public lease pricing (65 units × $15–20K × 12), not a company disclosure. Labelled as an estimate in the output.
- The land-grab is over, and a lockout lands on the renewal cycle. The CBA expires 1 Dec 2026 with a lockout widely expected into early 2027. Discretionary player-development spend freezes exactly when the 2023–24 leases come up for renewal.
- Capacity is a balance-sheet problem, not just a factory problem. Every unit is financed before it earns, revenue arrives across 36 months, and no institutional round is on public record. "We can only build so many a year" may really mean "we can only fund so many a year."
- Growth beyond MLB breaks the price point and the go-to-market at once. College and high school cannot absorb MLB pricing, cheaper incumbents hold that band, and there is no sales or CS function to serve hundreds of small accounts.
- "Of the leases that have hit their three-year term, what percentage renewed, and at what price or unit count?" Adoption is the known story. Renewal is the only evidence the Arc is embedded rather than a novelty tried once.
- "What is tying up your cash, the factory or the fleet?" Decides whether "should we raise" is an equity question at all, or an asset-financing question being framed as a venture round.
- "If there is no baseball from December to March, what happens to bookings?" Forces diversification onto a real deadline and tests whether they have modelled the lockout.
- No financials anywhere public. Revenue, gross margin, runway all unknown.
- HQ contradiction. Notes read Kitchener-Waterloo; reporting puts HQ in Mississauga with manufacturing in Cambridge.
- Headcount contradiction. Notes say ~15. Public sources range 21 to 52.
- "Most teams are using it" is the risk, not the win. At 25 of 30, the new-logo runway is roughly four clubs.
If context runs out, work in passes
Give each question its own run, then combine. A Project helps here too: your instructions live there instead of eating each chat's memory. On a bigger model the single prompt above usually holds without splitting.
Turn it into a Project
Sidebar → Projects → New project. Two things get saved once: the instructions are the "how", and the context is the format.
One line, and you get the whole brief
Same full brief, but you pasted no prompt and no format. The project holds both. As Heather said in the room: I built this once. Now any founder is: open the project, drop in what I have, name the company.
Turn the brief into a deck
Ask for the outline before the slides. It is much faster to fix a wrong running order in a list than in a built deck, and once the outline is right the slides are mechanical.
First, have Claude learn your brand from your own website
This is the step most people skip, and it is why AI-generated decks all look the same. You only do it once. The output is a small file you keep and attach to any project that needs to look like you.
See the design.md it built from momentumproductco.com
- Abel for every heading, h1 through h4. Weight 400 only, never bold. Line height 1.12, letter spacing 0.012em.
- Inter for body copy at 17px, line height 1.65. Weights 400, 500, 600, 700.
- Never substitute a generic sans for Abel in headings, even when body copy is Inter.
- One accent, used sparingly:
#F26C26. Darker#C24F10for links and small text on light backgrounds. Soft#FDEBDFfor fills. - Ground
#FBF9F6, cards#FFFFFF, ink#211E1A, muted#6E675D, rules#E8E2D8. - Dark sections invert to
#211E1Aground with#F3EFE8text and#A79E90muted. - No second accent. No blue.
- Border radius 3px everywhere. Nothing rounder.
- Content max width 1080px, prose max 68ch.
- Eyebrow above most headings: uppercase, 0.74rem, letter spacing 0.16em, weight 600, in accent-dark.
- Alternate light and dark sections down the page rather than boxing everything in cards.
- Sentence case in headings, not title case. Headings often end in a full stop.
- No em dashes. Use commas, colons or full stops instead.
Make the deck in your own design
Same pattern as the founder prep Project: save the design file once as context, and every deck after this starts on brand instead of needing a rebuild. If you use a separate presentation tool, the design file travels there too, and note that any third-party tool is a separate company with its own terms.
An automated playbook for meetings.
Think of it as a lifecycle with two halves. After the call: summarize, pull action items with owners and dates, draft the follow-up in your voice, recommend next steps. Before the next call: recap where you left off, research what has changed, draft the agenda from what is still open.
Keep it in one chat or a Project and the two halves connect. "Before the next call" already knows what happened at the last one.
Close out the call, then set up the next one.
Summarize the call and pull the action items
Upload the transcript or your notes first. In the session we ran it against a sample partnership discovery call with Dana at Ridgeway.
Draft the follow-up in your voice
It writes the follow-up while the call is still fresh, in your voice rather than robot-voice. You can build a voice file from a few samples of your own writing, then refine it over time: whenever the skill produces something that doesn't sound like you, edit the skill rather than just fixing that one draft.
Put the items where you'll actually see them
Connect Gmail and Calendar once in Settings → Connectors and Claude can file these for you directly. Note it drafts the email rather than sending it, which is where you want the human in the loop.
Recap and research, pointed forward
The same internal-plus-external move as the founder prep workflow, aimed at the next conversation instead of the first one.
Draft the agenda
Keep the whole lifecycle in one chat or one Project and the two halves connect. "Before the next call" already knows what happened at the last one.
Then make it run without you.
Everything above still needs you to start it. This is the one thing a chat window cannot do: work that happens while you are not there. It runs in the cloud, so your laptop does not need to stay open and you do not need to be at your desk.
Set up the scheduled task
One sentence is enough to create it. As Heather put it: I didn't add a capability just now. I removed myself from it. That's the whole difference.
The scheduled task running since the day of the workshop
The version Heather actually runs is narrower and more careful than the demo. It calls a saved skill rather than restating the steps, keeps a state file so nothing gets processed twice, and does nothing at all on a day with no founder calls.
Daily Founder Call Follow-up
Three details worth stealing. It calls a skill by name rather than repeating the steps, so the recipe can improve without touching the schedule. The state file means a call is never summarized twice, which is the failure mode that kills most daily automations. And "if there were no founder calls, do nothing" is what stops it filing an empty report every evening and training you to ignore it.
If you're building your first one, start it read-only. Have it message you the summary for a week before you let it write into your task list or your CRM, so you can see what it would have done.
Make it a skill.
Two ideas that stack together. A skill is the recipe: codified know-how for a task done the same way every time, which Claude picks up when it fits. An automation is the doing: multi-step work that runs on a trigger or a schedule, sometimes while you are away.
A scheduled automation can call a skill. The recipe card, and the sous-chef who cooks it on schedule.
The test for whether something should be a skill: same steps every time, a standard output, you run it on a cycle, and more than one person does it. If all four are true, stop rewriting the prompt.
- A prompt you reuse. Save the one that worked in a note. Any plan.
- A Project or a Skill. Instructions saved once. A skill fires on its own when the job fits.
- A trigger. It runs at 5pm whether or not you open Claude.
Most people stop at the first rung. The prompt that worked gets lost in a chat window, and next month the whole thing gets rebuilt from scratch. Every rung after that is just refusing to rebuild it.
The program-feedback eval, built live.
The best skill candidate from the survey: a monthly survey export that always needs the same five things done to it before anyone can read it.
Clean the export and theme the feedback
Upload the raw export exactly as it came out of the survey tool. You don't clean it first. Handing over the mess is the job. The labels in the real file included "rev", "Rev", "rev scale up" and "Rev Scale-Up" as four spellings of one program, and scores written as "9/10", "9" and "ten".
Turn that procedure into a skill
Save it under Customize → Skills. Next month you upload the export and ask for Program Feedback Eval by name. Same steps, standard output, run every month, more than one person does it: that is the test for whether something should be a skill.
Tracking the waves, not just cleaning one file
The cleaning report is what makes the number trustworthy. It shows July's 8.73 came from 27 clean rows out of 28, with 2 duplicates flagged and 5 rows missing fields. Without it you are trusting an average you cannot audit.
The dashboard it produces is published here, in full:
Trend over time, a per-program breakdown, movement across the six waves, and the cleaning report for each one. The survey data behind it is simulated, made up for the workshop so the cleaning and trend steps had something realistic to work on. The dashboard says so at the top.
The skills running underneath.
Every workflow above started as a prompt, got reused, and eventually became a skill. Here is what that looks like once it settles. These are the actual skills behind the sessions demoed in the room.
founder-call-prep
Workflow 1
Recaps the last call, reconstructs what has moved since across email, calendar, CRM and Slack, pulls open commitments, and builds the agenda. One brief per upcoming call.
What it does
You ask to be prepped for founder calls in the next three days, or for a catch-up before a specific call. Also on a schedule.
This one is for a call that hasn't happened yet. Its sibling, founder-call-followup, closes out one that has.
- Recaps where the last call left off
- Reconstructs what has moved since, across Gmail, Calendar, HubSpot and Slack
- Pulls your open commitments from the follow-through tracker
- Builds an agenda from what is still open
- Slack DMs one brief per call
founder-call-followup
Workflow 2 · on a schedule
The full post-call pipeline: summary, owner-tagged action items, the engagement doc, the email back in your voice, CRM tasks, and a recap of everything staged. This is the one running unattended every weekday at 5:30 PM.
What it does
You say you just got off a call, ask to close out today's founder calls, or the scheduled task runs at 5:30 PM.
It stages rather than sends. The email lands in drafts, so nothing goes out without you.
- Summary and owner-tagged action items
- Research on the threads left open
- An internal Slack draft to a colleague
- The engagement doc, and open calendar slots
- The email back to the founder, in your voice
- HubSpot deal and tasks, plus the follow-through tracker
- A Slack DM recapping everything it staged
meeting-action-items
Workflow 2
Pulls meeting notes, extracts prioritized action items, and loads them into task lists. Tracks the last run date, so only new meetings are processed and nothing duplicates.
What it does
You ask what came out of your meetings, to sync meeting notes to tasks, or to be caught up on what you need to do.
It records the last run date and only processes meetings newer than that. The same idea as the state file behind the scheduled task above.
- Pulls notes from my call recordings
- Extracts and prioritizes the action items
- Loads them into Google Tasks
- Routes work items to the Momentum list, personal ones to Things to Do
nps-trend-tracker
Workflow 3
Cleans one or more survey exports and tracks average NPS over time, overall and by program. Maintains the running history rather than treating each export as a one-off.
What it does
You want to track NPS over time, compare this month's survey to last month's, or update the running tracker. It triggers even when you only attach one file and imply more are coming.
This is the Program Feedback Eval from Workflow 3, grown up. The sample dashboard is what it produces.
- Cleans the export and reports what it removed
- Normalizes messy program labels and score formats
- Themes the "most valuable" and "could improve" free text
- Averages NPS overall and by program
- Appends the wave to the running history and rebuilds the rollup
coach-matcher
Workflow 3
Matches founders from an intake export to the right coach, with ranked reasoning, draft intro emails, and a workload report. Says "none with confidence" when the roster has a gap.
What it does
You have a batch of founder intakes and a coach roster and need them connected. It triggers on "who should these founders see" or "triage this intake batch," not just the word match.
It is allowed to return nothing. When no one on the roster genuinely fits, it says "none with confidence" and names the gap instead of returning the least-bad option.
- Checks eligibility, but scores everyone and says who failed and why
- Reads the free text over the dropdown, and flags where they disagreed
- Ranks a shortlist with reasoning, plus two alternates
- Drafts the intro emails
- Reports coach load against capacity
Writes in an actual voice rather than generic AI. This is what makes the drafted follow-up emails sound like a person. Built from real writing samples and refined over time.
Turning one into a scheduled task.
A skill is the recipe. A scheduled task is what makes it run without you. One of the six above is on a schedule today: founder-call-followup runs as Daily Founder Call Follow-up, weekdays at 5:30 PM, in the cloud. It has been running since the afternoon of the workshop. The full instructions are above.
Scheduling your own
The end-of-day meeting follow-up is the easiest first one, because you find out within a day whether it earns its place. Start it read-only and let it prove itself for a week before you give it permission to write anything. And include that last line, because a task that files an empty report every day is a task you will start ignoring inside a fortnight.
What's safe to put in.
A gut-check that applies to any AI tool, not just Claude.
Green, go
Public info, market research, your own drafts, anonymized notes.
Yellow, depends
Internal docs and founder background. Depends on your plan and your settings.
Red, keep out
Revenue numbers, personal information, anything under NDA, unless cleared under the org account.
On a personal plan, keep it to public-facing research. Don't paste in internal revenue a founder shared with you in confidence.
Your plan and your data
On a personal plan you choose whether your chats train Claude. Opted out means not used for training and deleted within 30 days of you deleting them. Opted in means up to five years, de-identified, and only for chats started or resumed after you turned it on. Set it in Settings, then Privacy.
Team and Enterprise work data is not used for training by default, which is the safer footing for internal work. One exception: thumbs-up and thumbs-down feedback can retain that chat for up to five years, and admins can turn feedback off org-wide.
Verified against Anthropic's published policies as of August 2026. Worth re-checking before you rely on it.
Lock it down in five minutes
- Incognito chat. Ghost icon, top right of a new chat. Not saved, not remembered, never trained on. Every plan.
- Training off. Settings, Privacy, Model Improvement. Off means deleted in 30 days rather than kept for five years.
- Memory. Settings, Memory. Pause it, or reset it and everything Claude knows about you is gone. This is the one that creates exposure for personal information.
- Connectors. Settings, Connectors. Disconnect anything a given job doesn't need. Least privilege, same as any other tool.
- Org retention. An owner sets how long chats live. Left alone, the default is indefinite. Team and Enterprise only.
Four of those five apply on any plan, and none of them takes a minute.
Use the lightest model that does the job.
- Haiku, fast and light. Quick summaries, extractions, first drafts.
- Sonnet, the balanced default. Most everyday work: summaries, writing, analysis.
- Opus, deep reasoning. Research, synthesis, the genuinely tricky stuff. The founder-prep workflow above.
- Fable, the heaviest lifts. When you want the most capable model available.
Sonnet is the right engine for most of what's on this page.
What to remember when you sit down to try this.
- Start read-only. Have a new workflow report back before you let it write anything. You get to see what it would have done, and nothing can go sideways while you're still learning to trust it.
- Bring your own context first. Your notes, your transcript, your template. Then have Claude research what you don't have. Combining the two is where the value is.
- Make it say when it disagrees with you. The Trajekt brief caught two contradictions between the call notes and public reporting, before the call rather than during it.
- Ask for sources, and label estimates as estimates. A $12–16M revenue figure derived from public pricing is useful. The same number presented as fact is a problem.
- Let it refuse to answer. Give it permission to say "none with confidence" and it will surface the gaps in your own data. A tool that always returns something just buries them.
- Report the cleaning, not just the average. An NPS number you can audit is worth more than one you have to trust.
- Climb the ladder. Reuse the prompt, then make it a Project, then make it a Skill, then put it on a schedule. Most people stop at step one.
- Don't stop at the prompt that worked. Save it, make it a Project, make it a Skill, then put it on a schedule. Each rung is just refusing to rebuild the same thing next month.
- Give an automation permission to do nothing. "If there were no founder calls, do nothing" is what keeps a daily task worth reading. Pair it with a state file so nothing gets processed twice.
- Check the privacy settings before you need them. Four of the five take under a minute and apply on any plan.
Try one thing tomorrow.
The one to start with
Then save the prompt. Reuse it, make it a Project, make it a Skill. That's the whole ladder, and it starts with one prompt you didn't throw away.
Learning this is a team sport.
New rules ship weekly, so the fastest way to get unstuck is to ask in public. Share what you tried and what came back. Someone else has hit the same wall.
- The Slack community. The Momentum AI Working Group, around 200 people learning this together. Start in #just-getting-started-with-ai or #ai-tips-n-tricks. Ask anything, get a real answer fast. No pitches, no selling.
- The workshop library. momentumproductco.com/ai-workshops holds every session we've run, each with its prompt pack, templates and the questions that came out of the room.
- Just ask. Email heather@momentumproductco.com or find Heather on LinkedIn. Tell her the task you're stuck on and she'll point you at the right thing, or build it with you.