The call that changed everything
I got a call from Nora on a rainy Tuesday. She writes a weekly newsletter, interviews guests for her show, and spends two days each week turning hour-long conversations into short, readable summaries. "I can’t afford an editor," she said, "but I can’t afford to keep doing this either."
We sat on a Zoom and I watched her swivel between her inbox, a transcription app, and a half-finished Google Doc. It was obvious: the problem wasn’t AI itself. It was the tiny, exact pain of one person—Nora—who needed an hour of audio converted into a 5-minute, publishable summary with a few pull-quotes, fast.
Why one person matters more than a market
Big markets are seductive. You imagine millions. But the first sale rarely comes from “millions.” It comes from one stubborn person who will test your patience and teach you things you couldn't learn from a spreadsheet. If you can help that one person, you unlock a path to help many like them.
From problem to micro-product
We sketched a plan over coffee: find the exact output Nora wanted, build a simple workflow that saved her time, and charge her something she’d happily pay. No bells, no long-term hiring. Just a tight loop: talk -> build -> ship -> improve.
- Define the exact outcome: "Turn a 60-minute interview into a 400–600 word summary, 3 pull-quotes, and 1 shareable tweet in under 30 minutes."
- List the constraints: cost under $200/month, no in-house editors, reusable templates, and privacy for interviewees.
- Pick the tech that gets you there fastest: transcription + LLM summarization + a few human edits.
How to do this with the smallest possible budget
Here’s what we actually built in the first month, line by line, with numbers because nothing beats real constraints.
- Discovery (free): 5 short interviews with other newsletter writers and podcasters. Ask: what is the one thing you would pay for today? (Find them in Twitter DMs, Substack comments, a Slack community.)
- Landing page (free–$20): One-page preview with a call-to-action and a transactions link. We used Carrd + Stripe—cheap and fast.
- Prototype ($30–$150): Transcribe with Whisper or a cheap transcription API. Summarize with an LLM prompt (few hundred tokens). Use a Zapier or Make automation to move the transcript to the model and then to a Google Doc for a quick human pass.
- Presell and validate (free–$0): Tweet about the workflow, email your early contacts, and offer a one-time discounted pilot to 3 people. If 3 of 10 say yes, you’ve got a product.
- Iterate ($50–$200/month): Improve prompts, add templates, automate formatting, and measure time saved and willingness to pay.
Where the money actually goes
- Transcription API: $10–$50/mo depending on volume
- LLM usage: $20–$100/mo for early prototypes (optimize prompt length)
- Landing page + payment: $5–$20/mo
- Automation tools: $0–$50/mo (start free and upgrade as needed)
The little story that taught us prioritization
Our first month looked rough. The automation failed twice, the first summary came back robotic, and Nora suggested we stop using the phrase "AI-powered" because it scared some guests. We adjusted: human in the loop, friendlier voice, clearer opt-in language. Then Mara, a teacher from Spain, signed up after seeing our thread and paid for a pilot. She loved it. She told three other teachers. One week later we had three paying customers and a list of eight people willing to try.
What changed? Two things: we focused on output quality (not flashy features), and we talked to customers like they were collaborators, not user personas. That’s the secret for tight budgets—you get free product research when you listen.
Practical playbook you can use tonight
If you want to follow this path, here's a compact, actionable checklist—no jargon, no fluff.
- Pick one tiny audience. Example: "indie newsletter writers who record interviews."
- Define one measurable outcome they want. Example: "Publish-ready summary in under 30 minutes."
- Talk to 5 people in that audience. Record answers. Look for identical complaints.
- Build the simplest workflow that delivers that outcome. Use existing tools (Whisper, ChatGPT/GPT API, Google Docs, Carrd, Stripe).
- Offer a paid pilot to 3 people at a price that covers your costs + a little margin.
- Ship, collect feedback, and improve. Keep the human in the loop until the automation reaches the required quality.
Prompt engineering as your MVP
One weird trick that saves money: treat your prompt like a feature. That 100–200 tokens you send to the model is a deliverable. Test multiple prompts with real transcripts. Keep the ones that need minimal human edits. Put the prompt in a template and make it your secret sauce.
“Don’t build an AI product. Build a repeatable human + AI workflow that solves one clear problem.”
Distribution without ad spend
When you’re tight on cash, paid ads are for later. For the first customers, use these low-cost channels where niche people gather:
- Niche Twitter threads and DMs
- Substack/Medium posts and comment outreach
- Relevant Slack, Discord, and Facebook groups (help before you pitch)
- Referrals from your first three paying customers—ask for introductions
- Small bundles or partnerships (a podcast host mentions you in exchange for a discounted pilot)
Metrics that actually matter
Forget vanity metrics. Track three numbers every week.
- Time saved per customer (hours/week)
- Conversion from pilot offer to paid customer (goal: 30%+)
- Retention: how many customers come back for month two
If time saved is high and conversion is good, you have a product. If not, you either picked the wrong problem or your output quality needs work.
When to scale and when to stay small
Scale when the product consistently delivers the promised outcome with low marginal cost. Stay small if you prefer a high-touch, high-margin consultancy model—both are valid. The advantage of starting tiny is that you can choose later. You won’t be tied to a big infrastructure or large team because you didn’t build them unnecessarily early.
Last bit of advice, from one friend to another
Start with a person, not a pitch deck. Help them so well they tell two friends. Keep costs microscopic by using existing models and automations, and keep improving the prompt until the AI feels like an assistant—not a gimmick. You don’t need to be the next OpenAI to make something valuable. You just need to be the person who noticed one tiny, expensive pain and had the audacity to fix it for one person first.
If you want, try this tonight: pick one person you know who has a repetitive, time-sucking task. Send them a short message offering to automate one bit of it for a small fee. Watch what they say. That response is your startup’s earliest data.
