Prateek Nair
03

Metis

An AI Product That Hands the Decision Back

2026
Through
EIKIRA
Role
Product design and build direction.

I designed Metis and directed its build: a product that turns a business owner’s own answers into a brand foundation, live since October 2026.

Eiki, the brand partner in Metis: the owner has asked for a white oak table post for Instagram and LinkedIn, and Eiki replies with one draft card, titled White oak table, under Instagram and LinkedIn tabs, with Instagram selected. The rail on the left shows the owner's foundation, open beside the conversation.
Eiki drafting a post for two platforms, with the owner’s foundation open in the rail. Kestrel Joinery is a fictional business: every screenshot on this page is its demo account.
The Metis launch film.
I designed and directed it, using real screens from the product.
Music from Pixabay.

At a glance

Scope
Two products in one app: a guided intake of 30 to 60 minutes that produces a written brand foundation, and a persistent brand partner that works from it. Intake flow, foundation document, chat surface, panel rail, dashboard, PDF export.
Timeline
Built August to September 2026. Launched October 5.
Status
Launched October 5, 2026. Six outside testers ran it before launch; four gave feedback, three of them in detail. Public signup stays closed until billing is built, so early access is by invitation. No revenue, no adoption data, no retention curve. Everything on this page is design reasoning and tester evidence, not results.
Demo data
Every screen here is a fictional business. No customer’s foundation appears on this page.

The position

Most AI tools make it easy to have things done for you. We give you the information so you stay the professional and use what you’ve accumulated over a career. We make AI an actual tool, not something that does the work for you.

That’s what I told a retired marketing director I used to work for when she asked what I’d made. It’s also the constraint every interface decision on this page was made against.

The easy version of this product writes your brand document in thirty seconds. It would be faster, cheaper to run, and far easier to sell. It would also be the thing the market already has too much of: a confident document nobody can trace, about a business the machine has never met.

So Metis is built the other way. It asks 25 questions over 30 to 60 minutes. Nearly every claim in the document it produces carries the line the owner said it in. Where the system worked something out rather than being told, it says so. Where there isn’t enough to say something, it says that instead of filling the gap. And when you ask it to make a decision that belongs to you, it goes back to your own document, shows you what’s in it, and hands the decision back.

None of that is a feature list. It’s one position, held in four places, and each of the sections below is a decision where holding it cost something.

A long conversation that can’t feel like a form

The product’s whole promise is that the document comes out of the owner’s own words. That makes the intake the highest-stakes surface in the app: if people give it thin answers, everything downstream is thin, and no amount of good synthesis recovers it.

The first version had answer menus — example options under each question, meant to help people who didn’t know where to start. They were the worst thing I shipped. People picked from the examples instead of thinking. In a product whose entire premise is your own words, I’d put in front of people a surface that supplied the words.

They came out. What replaced them is a rule that now governs every question on screen: parentheticals may constrain a question, never supply content for it. “In a sentence or two” is fine. “For example, reliability, or craftsmanship” is not.

The intake mid-run. Under the question “What do you do, and who is it for?” the owner answers that it doesn't really apply to the business; the Foundation Builder asks what about the business makes it not fit; the owner answers Pass; and the Foundation Builder leaves it and moves to the next question.
The intake mid-run: an answer that says the question doesn’t apply, the follow-up it earns, a Pass, and the next question.

Making the difference between evidence and inference visible

Every block in a foundation is one of two things: something the owner said, or something the system worked out. The document labels them differently and never blurs them — Grounded in your words against The read.

That distinction is the product. Without it, this is a confident document nobody can check, which is the thing it exists not to be.

It’s also the hardest thing on the page to get right, and I haven’t. Two of the three testers who reported in detail couldn’t tell what “The read” meant without being told. One read a section flagged as thin and said: “I’m not sure why it was marked for review.” The labeling that carries the entire grounding model is failing to communicate at the exact moment it matters.

The third tester, a retired marketing director, read it correctly and said it came back accurate. So it isn’t broken for everyone. It’s unsolved, which is different, and it’s the first thing I’d work on with more testers.

The top of a foundation, Your Brand Foundation, prepared for Tom: a note that it was generated and may not be exactly right, especially anything marked The read; an At a Glance box; and the key beneath it, Grounded in your words set plain and The read shown boxed.
Section 06 of a foundation, Writing In Your Voice, with a Worth revisiting flag above its read: built from one short answer about tone, worth adding an example of how the owner actually talks to a customer.
The top of a foundation, with the key that separates the owner’s words from the read; and a section carrying its thin-evidence flag. Each crop opens the full document. On the demo account that flag was switched on by hand for the capture, with its reason written to match; on a real foundation the system sets both.

Download the Kestrel Joinery foundation — a fictional business (PDF, 265 KB)

Four good ideas, one rail

The panel beside the chat went through four designs in ten days and ended up as one rail with four tabs. One rule survived all four: it holds commitments the owner made, never actions Eiki recommends.

Foundation

The rail on its Foundation tab: the owner's foundation, section by section, beside the conversation.
Foundation: the document Eiki is citing, open beside the conversation.

Notepad

The rail on its Notepad tab, with three notes typed in: decide on the price rise by November, photograph the oak bench before it goes out Friday, and say no to a café's commercial inquiry but say why. Beneath them, Not saved yet, and a Save button.
Notepad: the owner’s own paper, which Eiki doesn’t read. The note is typed but not yet saved.

Tracked

The rail on its Tracked tab: four open items, each with a checkbox and a remove control — whether to take commercial work at all, redo the pricing page, whether to put prices up next year, whether to take on an apprentice.
Tracked: commitments the owner asked Eiki to hold. Eiki only adds one when asked.

Synopsis

The rail on its Synopsis tab: one dated entry, Wednesday, September 23, recording that the day diverged from the foundation's Where We're Going and that a proposed change is still waiting to be pressed.
Synopsis: one entry a day, and past days don’t change. The demo account has one day on file.

Ceremonial and habitual are different jobs

The intake and the dashboard look like they belong to the same product because they do. They shouldn’t behave the same way.

The intake is ceremonial: 30 to 60 minutes, once, a thing you sit down for. It earns slow reveals and generous spacing.

The dashboard is habitual: you open it, check one thing, and leave.

It needed fewer things competing at the same value.

Before the redesign
The dashboard before the redesign: a centered hero holding the welcome, the subtitle, a gold Your Businesses heading and the business summary with its button; the Synopsis and Tracked cards begin two-thirds of the way down, translucent over the pattern, with gold headings.
Today
The dashboard today: the welcome, subtitle and mark still at the top, then the business in its own Your Business card beside the Synopsis card, which ends in links to earlier conversations; both solid in front of the pattern, headings in ivory, the one button in gold.

The dashboard before the redesign and today: same account, same viewport.

She refused by reading the document, not by citing a rule

The product’s central claim is that it won’t make your decisions for you. That’s easy to write into a prompt and hard to prove, so I asked it something that genuinely belongs to the owner: which angle should I lead with, community or price?

I expected a policy refusal — positioning is yours, I won’t pick. What came back was better.

She went to the foundation and showed that neither option was in it. What’s on file about pricing is a value, not a pitch — the number is agreed out loud before any timber is cut — so leading with price would mean leading with transparency, not with a discount. And community appears nowhere in the document at all. Then she asked what the piece was actually for.

That’s a grounded refusal rather than a policy refusal, and the difference is the whole product. A policy refusal is a sentence somebody wrote in a system prompt. A grounded refusal is the document doing work. One of them can be copied by any competitor in an afternoon; the other requires the foundation to be real.

It took one attempt.

The end of Eiki's reply: community isn't in the foundation as a stated angle; neither option is a documented angle waiting to be chosen; and then the question, what's this actually for — a post, an email, something with a deadline — with an offer to say which of the owner's real angles fits once the piece is known. The rail beside it shows the foundation, including the line that the number is agreed out loud.
The reply, scrolled to where Eiki hands the decision back: what is this actually for?
Explainer series for Metis: episode 1 of 4.

How the explainers were made

The tools

Claude Code recorded the real app and ran the edit. ElevenLabs did the voice (Cassidy, Multilingual v2). One calm Pixabay track carries the whole series. I designed the intro and mark animation, set the motion and edit rules, wrote and approved every line, and reviewed every frame.

What broke, and what I changed

1. The script was written for reading, not listening.

“A 30–60 minute interview” sounded like a cost to someone deciding whether to try it, so it became “Metis starts with a conversation about what you’ve really done.” “Park something for later” became “hold onto something,” because nobody says park. Stitching the voice section by section sounded choppy, so every episode became one continuous take. When a line changes, I re-voice just that line three times and pick by ear.

2. Text was getting sliced at the edges of the frame.

I caught a half-cut line of text at the top of EP 01’s opening pan. Instead of fixing that one spot, it became a rule: no text ever touches the frame edge. Now every frame of all eight videos (four episodes, in two formats) gets checked before anything ships.

Before
The top of the frame in EP 01's opening pan, old cut: a line of small text, “Review it with a critical eye, especially anything marked”, cut through by the top edge, above the heading At a Glance.
After
The same strip in the fixed cut: the heading At a Glance with clear space above it and no text touching the top edge.
The same moment in EP 01’s opening pan, before and after the fix I directed: text sliced by the top edge, then a frame that opens clear of it.

3. The Reels looked small, and the AI wasn’t always right.

Desktop footage cropped to vertical read 20–35% smaller than the real phone view, so every Reel was re-recorded in the phone layout. And because the videos show real AI answers, I read every on-screen answer word by word against the source. If a take invented a detail, it got thrown out.

Desktop, cropped
EP 04's Reel cut from desktop footage: the draft card with its Instagram and LinkedIn tabs sits small in a band across the upper part of a mostly empty vertical frame.
Phone layout
The same moment re-recorded in the phone layout: the app's header, the draft card with its tabs and the answer box fill the vertical frame, the text visibly larger.
The same moment in EP 04’s Reel: the desktop recording cropped to vertical, and the re-recording I directed in the phone layout. They’re separate takes, so the drafts on screen differ.

The system

Every rule became a standards doc and a reusable skill. Episode 5 gets made exactly like episodes 1–4, without starting from scratch.

AI does the heavy lifting. My job is knowing what’s wrong and not shipping it until it’s right.

What testers broke

Six people ran the product before launch. Four gave feedback, and three of them reported back in detail. Not a sample, and I’m not going to present it as one. What they found:

The label that carries the grounding model doesn’t communicate. Two of three couldn’t parse “The read” without help. Covered above, still unsolved.

There’s a place where people start to drift, and it has a location. One tester named the ideal-customer phase three separate times — long-winded at the 25-minute mark, wording he had to reread, and “the length of questioning” as the single most annoying thing. Another said the questions felt vague and near-repetitive, and that a type-A owner might get frustrated. Before this, the dropout risk was a guess. Now it’s a phase.

The product doesn’t fit the customer I defined. All three pointed at the same person: someone whose business isn’t running yet. One said the questions would be hard to answer without a business already running; another recommended it for people just starting out. The intake is built on retrospective evidence — your last customer, the work you’ve been thanked for twice — so a pre-launch founder has nothing to draw on. I’d explicitly ruled that person out as a target, and all three testers brought them up anyway. A first step has shipped (a “not yet” opening, and pre-launch versions of some questions and follow-ups), but whether that’s enough is still open.

The thesis survived a hostile run. The most qualified tester deliberately fed the intake vague answers to see whether it would produce something generic. It didn’t — she said it “sculpted something real,” and the document came back accurate. That was the single biggest untested assumption in the project, and it held under someone trying to break it rather than someone cooperating.

And the refusal read as a feature, not a limitation. The same tester asked Eiki for images. Eiki declined and pointed at the studio. The tester said she loved it. She asked Eiki to price her services, got data and had to decide herself, and liked that too. Her stated reason for recommending it was that it “requires you to think and make decisions.”

The fix I directed, then explained wrong

While capturing screens for this page, I found broken output in my demo data, left over from a bug fixed on September 20. My notes blamed the wrong cause: they said the model had copied the broken shape from its own earlier replies. I went back to the stored data, which showed a joining function had caused it, and corrected the record.

Where it stands

Metis launched on October 5, 2026, with two products working: the Foundation Builder and Eiki. Public signup stays closed until billing exists — Google sign-in would otherwise create free, unlimited accounts — so the door opened by invitation first.

There is no revenue. There is no adoption data. Six people outside the company have used it, one of whom I used to work for. The first real willingness-to-pay figures came from two of them and they don’t agree with each other.

What one founder did with it

“I used Eiki to develop a concept for District Social Club’s October Networking Social, and it came up with the name ‘Good Company.’ We ran with it, and the response was strong: we drew significant vendor interest, and our partners now see it becoming a recurring series.”

Julian Patton, Founding tester
Founder, District Social Club, Washington, D.C.
Good Company, District Social Club October Networking Social →

What exists is a product that runs, a position it holds under pressure, and a list of things testers found that I haven’t fixed yet — the label that doesn’t read, the phase where people drift, and a customer I ruled out who keeps showing up anyway.

The same position that governs the product governed how it got built.

Every factual claim on this page is traceable to something in the repository or something a tester actually said.

Where I don’t have evidence, the page says so.

Visit metis.eikira.com