A daily goal coach on Claude, for under a penny a call
One Goal Sherpa is a goal-setting app with an AI guide that talks like a person, remembers who you are and asks for one goal a day. Claude runs every conversation. Here's how it's wired, what it costs, and what the numbers say after eight months live.
2,041 of 2,534 goals, March to November 2026. Claude runs every conversation behind them for under a penny a call.
The short version
We built One Goal Sherpa and we run it ourselves, live, with real ad spend and real strangers. People use it the way it was designed: eight in ten goals they set get completed, and most of those come with a written reflection.
Behind that, Claude has handled more than 105,000 calls at under a penny each, with code checking every reply before she sees it.
The problem
The app is for women in their 40s to 60s who've already tried the planners and the habit trackers. The most-tapped feeling button in the whole app is “Overwhelmed.” So the product can't add work. It has to feel like a friend who gets it, ask for one goal, and get out of the way in under five minutes.
That puts hard requirements on the AI. It has to sound human every time. It has to answer fast on an iPhone. It can never say something strange to a person in her first week. And because this is a consumer app, each conversation has to cost close to nothing.
What we built
- 01
One goal a day
A short daily loop: check in with how you feel, pick one goal, do it, reflect. The database itself only allows one active goal per person, so a bug can't break the promise.
- 02
The Sherpa
An AI guide that greets her, helps her choose today's goal and celebrates when it's done. Its voice comes from a written voice guide that every prompt and every line of app copy has to pass.
- 03
A memory the size of an index card
A short profile of who she is and what she's working toward, rewritten by Claude every night. The Sherpa reads it before every conversation, so she never has to repeat herself.
- 04
Reasons to come back
A morning quote picked for her, a reply when she writes a reflection, and text nudges written about her actual goal.
- 05
A lane with no AI at all
Quick Set lets her type a goal and go. Zero AI calls, on purpose. Some days she doesn't want a conversation, and the cheapest AI call is the one you don't make.
How Claude is wired in
Everything runs on Anthropic's Claude API through the official SDK. No agent framework and nothing in between. Eight decisions did most of the work.
- Messages API
One front door
The app's AI jobs run through one routing file. It lists each job, which Claude model does it, how long the answer can be and whether it streams. Fifteen jobs run in production today, from the daily conversation to the nightly memory rewrite. Changing a job's model is a one-line edit.
- Streaming
She watches it type
Conversations stream word by word, so the reply starts right away and she isn't staring at a spinner. We also cut our own prep time before the daily greeting starts from 356 to 162 milliseconds by running the database lookups side by side.
- Prompt caching
Pay full price for the instructions once
The Sherpa's voice rules are long and identical for everyone. We send them as their own block, marked for reuse, with each person's details after it. Claude then reads the shared part at a tenth of the price. A second marker moves forward each turn so a long chat isn't re-read from the top. On the daily conversation, the share of each prompt served from cache went from 17% in June to 55% in October.
- Tool use
Answers that fit a form
When the app needs data and not prose, Claude has to answer through a fixed form (a tool call with a schema) that the API checks before we see it. That's how a journal chat updates her Map and how the home screen gets its daily choices.
- Tool use
A quote it can't make up
For the morning quote, Claude gets a shortlist from a bank a human already vetted. It returns only the ID of its pick plus one personal line. It never writes the quote itself, so a fake or altered quote can't reach her.
- Model routing
The right model for each job
Claude Sonnet 4.6 writes everything she reads. Claude Haiku 4.5, the smaller and cheaper model, does the background sorting she never sees. Model names are pinned in code to names that roll forward as Anthropic updates them.
- Server checks
Guardrails in code
The prompt tells Claude what not to do. Code makes sure. Our server holds the first few words of every reply to catch leaked instructions, drops a closing question when there's no box to answer it in, and checks every text message for length and format. If a check fails, she gets a line a human wrote.
- Usage tracking
Every call is metered
Each call logs the job, the model and the tokens in and out (tokens are the units AI is billed in). That's how we know what each job costs, and it's what pointed us to the caching setup that cut the daily conversation's input bill by roughly 40%.
Results so far
Live in beta since March 2026. Product numbers cover real users from March 11 to November 10, 2026, with our own team and test accounts removed. AI numbers cover every call since we started metering on June 2.
The product
871 of them finished onboarding. That's 79%.
Out of 2,534 goals set. Eight in ten get done.
1,321 written reflections. In our August review, people who reflected on day one came back more: 71% versus 49% within a week.
The AI
About two-thirds of a cent, measured across 26,077 calls in October 2026 at Anthropic's list prices. One call is one reply or one background job.
October 2026: 26,077 calls on behalf of 643 people, about $176 at list prices.
Up from 17% in June. From August through October that made the daily conversation's input bill roughly 40% lower than sending it uncached.
The install growth (230 to 915 in 30 days) has its own case study.
What made it work
- 01
Measure the AI bill by job
Because every call is metered, we can see what each job costs. That's what led us to split the prompt into a shared block and a personal block, and it took the cached share of the daily conversation from 17% to 55%.
- 02
If the app depends on it, make it a form
Claude writes freely where she reads. Where the app reads, Claude fills in a form the API validates. That keeps her Map and her daily choices updating reliably.
- 03
Write instructions from the reader's side
Claude mirrors the wording of its instructions. So every field description says who will read the answer and in what voice. That keeps the Sherpa talking to her, never about her.
- 04
Give every AI moment a human-written backup
Every place Claude speaks has a line a person wrote standing behind it. If a reply doesn't pass a check, she still gets a good moment.
How we build it
The code is written with Claude Code, Anthropic's coding agent, working from written plans and decision records kept in the project. A person reviews each change and tests it on a real iPhone before it counts as done.
Numbers get the same treatment as code. One AI model pulls a figure and a second, different model recomputes it from scratch. If they disagree, nobody acts on it.
Built with
- Claude Sonnet 4.6
- Claude Haiku 4.5
- Claude API
- Claude Code
- Next.js
- Supabase
- Vercel
- PostHog
- Twilio
How we checked these numbers
Usage and cost figures were pulled from the app's production analytics on November 11, 2026, then recomputed independently by a second AI model before publishing. Costs are estimates from our own token logs at Anthropic's public list prices, not invoice totals. The speed figure comes from a before-and-after test in August.
Where does the work get stuck for you?
Thirty minutes. We'll tell you whether there is a worthwhile first win.