All work

AI product

Blinkwiser: a team of AI agents that makes carousels you can check

The first version made up its own statistics.

So I rebuilt it as a studio you talk to, run by a team of agents.

One edit, in v1 and in the rebuild

v1

  1. A number with no source
  2. Claimed an edit it hadn't made
  3. No way back: undo was a chat message

The rebuild

  • Every number traces to a source
  • Says what didn't change, then retries
  • Every change saves a restore point

Illustration with demo content, not a real deck. The live product is shown below.

Result
23 secondsfrom a topic, link, video or file to the first slides on screen, while the checks finish
Role
My own product · product, UX and the agent workflow
Context
Blinkwiser · built by directing AI coding agents
Timeline
Jul 2026 to now
Domain
AI SaaS · creator tools
The productAgentic Carousel by Blinkwiser, live: a team of agents researches, writes and checks a LinkedIn carousel, and each step reports in the chat.Cut from a real screen recording of the live studio, sped up.
The product, runningOne run, sped up three times: each agent step reports in the chat, and the first slides land marked Draft while the checks finish.Screen recording of the live product.

The first version made up its own statistics

The brief was my own: turn a topic into a LinkedIn carousel, fast. I had a first version in beta within two weeks, and that's when the hard part showed up.

The model invented numbers and stated them with confidence. When I asked it to change something, it would reply 'I've rewritten the carousel' when nothing had changed. And if you didn't like an edit, the only way back was another chat message.

What I assumed

Generation is the product.

People post these carousels under their own name, so every fact has to be right.

Diagram · what a creator needs before posting

is it accurate?v1 invented numbersdid it do what it said?v1 claimed edits it hadn't madecan I get back?v1: undo was a chat messagedoes it look like me?v1 drifted to generic AIa carousel posted under your own namea carousel posted under your own nameis it accurate?v1 invented numbersdid it do what it said?v1 claimed edits it hadn't madecan I get back?v1: undo was a chat messagedoes it look like me?v1 drifted to generic AI
Generation was the easy part. For someone posting under their own name, these four questions are the product, and the first version failed all of them.

A wrong statistic costs a creator their reputation, and a tool that lies about its own edits can't be trusted with anything else. But every extra quality step adds time, speed decides who keeps using a tool like this, and for a solo, bootstrapped product the cost of every deck mattered as much as its quality.

A studio you talk to

Today it's a studio you talk to. You give it a request, a team of agents researches, writes and designs the carousel, and each step reports in the chat as it happens. I designed the studio around a few ideas.

  • Start from anything

    One box takes a topic, a link, a YouTube video or a file, and works out which it is. No tabs and no modes.

  • Questions only when needed

    It asks a question only when a request is genuinely unclear. A specific request goes straight to work.

  • One slide or all of them

    Select slides to change just those, or clear the selection and talk about the whole carousel.

  • It learns how you like it

    It remembers your tone, your brand rules and the words you never want, and keeps long chats short, so every carousel starts closer to done.

  • Close the tab, it keeps going

    Generation runs on a background worker, not in the browser, so you can walk away and come back to a finished carousel.

  • Out in one click

    A LinkedIn-ready PDF, high-res JPGs, a public share link, or any slide copied as SVG straight into Figma.

A team of agents, with the rules checked in code

Version one relied on prompts. Limits and 'don't make things up' were requests, and the model could ignore them. Writing better prompts would only have meant asking more nicely and hoping. So I split the work across a team of agents, each with one job, and anything that has to be true is now checked in code.

A safety check reads the request first, and a creative director turns it into a brief. A researcher searches the web and boils what it finds down to a numbered fact sheet, with a source for every fact. A planner then outlines the deck, one job per slide, and fixes the slide count.

After that, two things happen at once. Four cover hooks are written in different styles and scored, and the best one wins, while a writer drafts the rest of the slides. Code checks every length, banned word and half-finished sentence, and every number has to trace back to the fact sheet or its slide gets downgraded.

The draft shows up after about 23 seconds, marked Draft, while a critic, a fact-checker and a proofreader go over it at the same time. Only then is it saved.

Generation pipeline

Anything that must be true is checked in code

in parallel
in parallel

TopicThe creator's topic, source material, or request.

  • Model call
  • Checked in code
Hover or focus a step to see what it enforces. Timeline markers are measured averages; stage timing in between is illustrative.Redrawn from the product's architecture. Simplified.

Every edit is checked and can be undone

Once the editor became a chat agent, its honesty mattered more than anything else. I could have just shown whatever the agent said it did, but that's exactly what had gone wrong in version one.

One message can now ask for several things, like 'switch to the sketch template and punch up slide 2', and the agent plans each change in order. When the model claims an edit it didn't make, a guard catches it, tells you what didn't change and tries again. And if an edit brings in a number that isn't in the facts, the reply says so.

Every change saves a restore point. Clicking one puts the carousel back straight away, without another agent run, and the later replies fade with a Back to latest button. Typing undo works the same way. When you carry on from an earlier point, the faded turns are trimmed from the agent's memory, so it never reasons about changes that no longer exist.

Restore points and the honesty guard

Every change is honest and reversible

  1. Make slide 3 punchier.

  2. Rewrote the headline on slide 3.

    Slide 3Handoffs add hidden wait timeEvery handoff is a queue.

  3. Swap the stat on slide 5 for a newer one.

  4. Honesty guardModel claimed: Updated slide 5 with the latest statistic.

    I couldn't change slide 5. There's no newer figure in your sources or the fact sheet, so the stat is unchanged.

  5. Make the whole deck more formal.

  6. Adjusted the tone on slides 1–7.

    Slide 3Every handoff is a queue.Each handoff introduces a queue.

Deck preview · slide 3 of 7

03

Each handoff introduces a queue.

Agent memory

  1. youMake slide 3 punchier.
  2. agentRewrote the headline on slide 3.
  3. youSwap the stat on slide 5 for a newer one.
  4. agentI couldn't change slide 5. There's no newer figure in your sources or the fact sheet, so the stat is unchanged.
  5. youMake the whole deck more formal.
  6. agentAdjusted the tone on slides 1–7.
Restore an earlier point: later replies fade and are trimmed from the agent's memory, so it never reasons about changes that no longer exist.Demo conversation. Illustrative copy.

I settled quality versus speed with blind tests

The rebuild was more accurate, but it was slow, at 84 seconds a deck. I didn't want to pick a setting by feel, so I built a blind eval harness with 10 golden cases, a blind judge and head-to-head comparisons.

The first thing I had to fix was the judge itself, because it was marking correctly sourced numbers as invented. Once the judge was right, the evals showed that turning reasoning off kept the quality almost the same (8.0 against 8.1) and cut a deck from 84 seconds to 31, at about half the cost.

Diagram · deciding by evidence

10 golden casesreasoning onreasoning offblind judgequality 8.1, now 8.084 s down to 31 s a deckabout half the costjudge fixed first: it had marked sourced numbers as invented10 golden casesreasoning onreasoning offblind judgequality 8.1 → 8.084 s → 31 s a deckabout half the costjudge fixed first: it had markedsourced numbers as invented
I fixed the judge first, when it marked correctly sourced numbers as invented. Then the evals showed reasoning off held quality and cut time and cost.

I cut finished features to protect the templates

I'd built a freeform canvas and a looks system with nine styles, plus pickers for fonts and layouts. The more options I added, the more every template looked like every other AI tool.

So I cut them, and kept three templates exactly as designed: The Truth, The Sketch and The Statement. The Sketch draws its own pencil doodles for each slide, and any template can write the whole carousel natively in six languages.

Diagram · fewer knobs, more design

9 styles × fonts × layoutsgenericcut to three, kept as designedThe TruthThe SketchThe Statement9 styles × fonts × layoutsgenericcut to three, kept as designedThe TruthThe SketchThe Statement
More options made every template look like every other AI tool. Cutting back to three, each kept exactly as designed, kept their identity.

What changed

In blind head-to-heads, the rebuild beat version one 7 times out of 10. Its claims hold up against their sources better, 9.5 against 8.8, and a carousel takes 31 seconds instead of 84 at the same quality. The first slides show up after about 23 seconds, marked Draft while the checks finish.

It isn't better everywhere, though. Version one still scores higher on flow, 8.4 against 7.5, and that's what I'm working on next.

The carousel builder is live and free at carousel.blinkwiser.com. All of these scores come from my own harness and an AI judge.

Blind eval harness

How I settled the trade-offs: blind tests

Head-to-head, 10 cases

7 / 10 v2 wins

Accuracy, scored against evidence

v29.5
v18.8

Flow

v27.5
v18.4

v1 still wins here. It's the next fix.

v2 (grounded pipeline) against v1 (prompt-only), 10 golden cases, blind judge.Scores from my own harness and an AI judge.

What I'd do differently

This was my first agentic AI product. The product, the UX, the agent workflow, the eval design and every cut were mine, and I built it by directing AI coding agents and holding them to tests and evals.

Looking back, I'd write the golden cases and the blind judge on day one. Version one had no evals at all, so I only found its problems once it was in beta.