Skip to content

Make your AI work first time, for less.

Your AI feature works great when you try it. Real users get something else. The answers are hit or miss, so they hit regenerate, complain in Discord, or leave without saying why.

Our template is a Next.js starter that records what users do with every AI result, so you can see which ones they had to redo and what that cost you. Then Claude Code or Cursor fixes those first, and proves each fix on real past requests before your users see it.

One email when it's ready. No newsletter.

$149 for the first 50 builders, then $249. See what's in the template

Four examples of a user redoing an AI result: a support answer asked again, an invoice total corrected, a headshot regenerated and an email draft rewritten. Each is counted as a redo, with the cost of the result that was thrown away.

What a redo looks like in four kinds of app. Each one is a result you paid for and the user threw away.

The report it writes from your own users

Every AI result counts as worked or redone, from what the user did next. Nobody has to click a rating. Then it shows who the AI fails for, and hands the worst case to your coding agent.

Fig. 1. Example data, 30 days.

Your support bot worked first time for 71% of answers.

Worked first timeof 40,000 answers; higher is better
71%
Wasted AI spendof $480 spent; lower is better
$139/mo
Real cost per answerlist price $0.012
$0.017

Counts as a redo when the customer asked again, corrected the answer, or asked for a human.

Who it fails for

  • Questions about the new Teams plan39% worked
  • Refund and billing questions48% worked
  • Questions not in English52% worked

Your average, 71%. Green worked first time; red needed a redo.

What your coding agent does with it

  1. Read 50 failed answers about the Teams plan.
  2. Found: 38 of them quote the old Pro plan page. The Teams page isn't in the search index.
  3. Tries: Add the Teams plan page to the index. The prompt stays as it is.
  4. Replays 212 past Teams plan questions: 39% → 74% worked first time, same cost per answer.
  5. Opens a pull request. The change goes to 10% of users first and stays only if real users redo less.

replay scored by a judge that was checked against your users' own redos

The numbers are made up. The layout is what the template produces.

Right now you find out from vibes, Discord or the bill

Fig. 2. Builders on Hacker News and DEV, quoted as written. They don't use our template; it doesn't exist yet.

Your coding agent is guessing

Claude Code and Cursor can write the feature, the prompt and the tests. What they can't see is your users: the 40,000 answers they got last month, and which ones they had to redo. So when you ask for better answers, the agent changes the prompt by feel, and you find out a week later whether it helped.

The template gives your agent what it's missing:

results
every AI result, the prompt and model version that made it, and what it cost
next step
what the user did next: kept it, regenerated, edited it, asked again
AGENTS.md
how to read the report, change one thing at a time, and replay real past requests before shipping
rollout
a change goes to a slice of users first, and stays only if people redo less
› make the support bot work first time more often

● Reading the AI results report (last 30 days)
  40,000 answers · 71% worked first time · $139 wasted
  worst: Teams plan questions, 39% (1,204 answers)

● Reading 50 failed Teams plan answers
  38 quote the old Pro plan page.
  docs/pricing/teams.md is not in the search index.

● Change: index docs/pricing/teams.md
  prompt v14 unchanged

● Replaying 212 past Teams plan questions
  worked first time   39% → 74%
  cost per answer     same

● Opened PR #87 "Index the Teams plan page"
  rollout: 10% of users, kept if redos drop
Fig. 3. A Claude Code session with the template's skills. Illustrative; the template is in development.

Your API dashboard shows the list price. Here's the real one.

Every redo is paid for twice: once for the answer the user threw away, and again for the one they kept. The bill can't show this, because it doesn't know which answers were redone.

Fix the answers that fail and you pay for fewer of them, which is where the “for less” comes from. Moving the steps that already work to a cheaper model does the rest.

Work out your own real cost per result

Support bot, October

AI bill
$480.00
answers
40,000
worked first time
28,400 (71%)
spent on answers users redid
$139.20
list price per answer
$0.0120
real cost per answer that worked
$0.0169
Fig. 4. The support bot from Fig. 1, one month, worked through.

A note from the founder

I built Storybirdie, an AI storyboard app for film directors. I had tracing, analytics, an events table and evals. I still couldn't answer the one question that mattered: did my last change make users get what they came for?

When I finally went through the data by hand, it was all there. One director spent 37 tries and 111 credits trying to get a character's clothes right. The app said shot lists were free while it charged 2 credits for each of 1,202 of them. In August the AI cost about $26 against $30 of revenue. No dashboard I had showed any of it.

The template is what I wanted then: a record of what users did with every AI result, and a coding agent that knows how to use it.

Amit Timalsina, who is building the template

Amit Timalsina, building AI Apps That Work

Free tools, while you wait

No sign-up. They run in your browser, and what you paste stays there.

All free tools

One payment, then it's yours

$149 for the first 50 builders, then $249. You get the code. It runs in your app, on your database, with your AI provider keys, and your users' data never comes to us.

It's in development. Join the waitlist and you'll get one email when it's ready. Already built your app? A retrofit kit that adds the same recording and report to an existing app comes next.

The template, in development

$149$249

One payment. Founding price for the first 50 builders.

  • Next.js and TypeScript, with any AI provider
  • Records every AI result and what the user did next
  • A report of who the AI fails for, and what it costs
  • Instructions and skills for Claude Code, Cursor and Codex

One email when it's ready. No newsletter.

Questions people ask

Can't Cursor or Claude Code already do this?

They can write the code. They can't see which AI results your users had to redo, so they can't tell whether a change helped. The template records that, and its instructions teach your agent to use it.

Do I need to set up evals?

No. Most results are scored from what users already do: regenerate, edit, undo, ask again, ask for a human. Where users give no signal, like a lead score, the report waits for the real outcome or asks you to check a small sample by hand.

Can I buy the template today?

Not yet. The template is in development. Join the waitlist and you'll get one email when it's ready, with the founding price of $149 for the first 50 builders.

My app is already built. Is this for me?

The next thing after the template is a retrofit kit that adds the same recording and report to an app you've already built. Join the waitlist and say what you're building, so you hear when it's ready.

Is it only for chatbots?

No. It works for any AI step: support bots, document and invoice readers, lead scoring, email agents, search, translation, and image or video generators. Each kind of step has its own sign that a result needed a redo.

Which AI providers does it work with?

Any you can call from your server: OpenAI, Anthropic, Google, xAI, Mistral, DeepSeek, open models through OpenRouter, and image or video APIs. Switching to a cheaper model without your users noticing is one of the things it's for.

Where does my users' data go?

Into your own database. The template doesn't send your prompts, AI results or user data to us.