Amigo Screen

1. 01Understand the problemGetting stuck is a workflow problem, not a skills problem.
2. 02Define the journeyShare, ask, one step, check-in, next step.
3. 03Choose the capabilitiesVoice in and out, on-demand screen reading, guided programmes.
4. 04Design the voice and screen interactionOne short spoken instruction, then it waits for you.
5. 05Structure the appHome, Discover, My Projects, Settings and the live session.
6. 06Privacy and controlYou choose what to share, stop any time, and do every action yourself.

The core loop

1. 1Share your screen
2. 2Voice session opensOne live OpenAI Realtime connection
3. 3Screen checked on requestA single screenshot, only when Amigo needs to look
4. 4One spoken stepThen a check-in before the next

The app, in one map

Blue names are the three destinations in the sidebar. Each row is checked against the app’s source or its live screens.

The idea, in one picture

You

Stuck in software you don’t know. You pick what to share and do every step.

Ask out loud, or type

One short step, then a check-in

Amigo

Listens, looks only when it needs to, and gives one short spoken instruction.

One screenshot, only when it needs to look

Your screen

The tab or window you chose to share, with the floating controller over it. It stays yours to click.

You click, type and scroll. Amigo never does.

You stay in control

Not claimed, because the privacy policy doesn't support it yet: end-to-end encryption, zero data retention, GDPR or CCPA certification.

How a step gets completed

1. 1See the stepInstructions and what “done” looks like
2. 2Do it yourselfIn the real tool; Amigo never clicks for you
3. 3Ask Amigo to checkIt looks at your screen against the “done” condition
4. 4You mark it doneAmigo can verify a step, never complete it

1. 
2. 
3. 

1. 01DiscoveryProblem, audience and the core loop, written as briefs before any build.
2. 02UX strategyVoice first, screen on demand, and the user does every action.
3. 03UI designPublic site and app, art-directed against a written brief.
4. 04AI-assisted developmentBuilt with Claude Code: 346 commits between August and October 2026.
5. 05IntegrationSupabase auth and data, the OpenAI relay, Stripe checkout and webhooks, Shopify OAuth.
6. 06TestingPlans reviewed by Codex; cost fixes verified with token counts from test calls; usage logged before cutting more.
7. 07DeploymentNetlify deploys every push to main, with security headers set.

1. Authenticated web appHome, Discover, My Projects, Settings, live session
2. Voice assistantLive speech with on-demand screen reading
3. Guided programmes7 programmes and 4 short tasks, each in checkable steps
4. Floating controllerPicture-in-Picture session controls
5. Accounts and sign-inGoogle or email, with row-level security
6. Subscriptions and paymentsStripe checkout, billing and webhooks
7. Public marketing siteLanding page, pricing, FAQ, legal pages
8. Production deploymentNetlify from GitHub, live at amigoscreen.com

Three decisions that made it fast and affordable

1. 01

   Challenge

   Slow voice turns

   Decision

   One OpenAI Realtime connection

   Result

   −2 network hops

   removed per turn

2. 02

   Challenge

   A screenshot every turn

   Decision

   On-demand screen capture

   Result

   Fewer vision calls

   and faster turn-taking

3. 03

   Challenge

   High running cost

   Decision

   Fixed prompt caching, gated the mic in silence

   Result

   ~30–60% lower

   roughly, than the modelled cost

Technical breakdown: provider comparison and running costs

All figures are raw AI cost before margin, for a five-hour allowance. The entry plan, Basic, charges £12 for those 5 hours.

Entry pricing

£12/mo

The entry plan: 5 hours at £2.40/hour, about a 50% margin over raw AI cost. Set low on purpose to learn from real usage.

Raw AI cost

Modelled

$22.50

Audited

$9–16

Modelled at $22.50, audited at $9 to 16 after the caching fix.

Biggest cost line

$5.76

$5.76 was mic audio streamed during silence, before the gating fix.

Modelled raw cost per five-hour session

OpenAI Realtime

~$22.50

Claude + speech layer

~$11.50

Claude plus separate speech-to-text and text-to-speech was about half the cost, but three moving parts and a hop of latency each. Gemini was evaluated, not modelled to five hours. Cheapest on paper wasn't cheapest to ship: voice has to feel instant, so I stayed on OpenAI Realtime and cut the cost inside it.

1. 1

   Fix prompt caching

   A recap rewrote the start of the prompt about every 100 seconds, so the cache never held. Made the prefix static and verified with cached-token counts from a test call.

2. 2

   Gate the mic in silence

   Settings that existed but were never wired in, now connected with a pre-roll and a hangover so soft speech isn't clipped. Estimated saving: $3 to $4 a session.

3. 3

   Measure before cutting more

   A cheaper model tier could save about $7 a session but risks weaker screen reading, so it is being tested first. Talk time and screenshot frequency are now logged.

One mismatch worth naming: the product charges in sterling and pays OpenAI in dollars, so the margin moves with the exchange rate as well as with usage.

What's next

• Test the cheaper model tier without losing screen reading
• Bank the silence-gating saving
• Revisit pricing once real usage data is in

No usage numbers, retention or revenue appear on this page, because none have been independently verified yet. What the project does show is one person carrying a product from the first brief to a live, paid service: the interaction design, the build, the infrastructure and the launch, with every decision traceable to the code that shipped.