Founder project · 2026

AI gives brilliant answers. Then buries them.

PACE (Personalized Adaptive Cognitive Experience) turns long AI answers into something you can scan, follow or listen to, without losing the substance.

✦Founder project · Claude plugin in build
Cognitive Mode: an AI answer in layered summary form beside a Reading Preferences panel
Project
Founder project · in build
Domain
Consumer AI · Accessibility
Role
Founder, product designer & builder
Tools
Figma, FigJam, Claude
01

The problem

The insight is in there, buried in one long block. Most people skim it, abandon it, or re-prompt: “shorter,” “use bullets,” “explain simply.”

~20%of the words on a page are actually read. People scan.Nielsen Norman Group ↗
76assistant models tested; models systematically over-answer, even when short is ideal.YapBench, 2026 ↗
MapsChatGPT users swamped by long answers asked for a view of how ideas connect.Graphologue, UIST 2023 ↗

How might we keep the depth of AI answers, but make the point findable in seconds?

02

Who it's for

The skimmer

Between meetings. Needs the point first, detail on demand.

The listener

Walking, commuting, cooking. Needs to hear it, not read it.

The low-load reader

Tired, neurodivergent, or reading in a second language. Needs one idea at a time.

Design ruleNever cut the substance, only change how it arrives. The full answer is always one tap away.

03

One toggle: Cognitive Mode

Same question, same model, two answers. No regenerating, nothing lost.

Standard · 10 min read
Standard mode: a long, dense answer about buying a house in California, with a 10-minute reading time
Cognitive · 20 sec read
Cognitive Mode: the same answer as a one-line process summary, key takeaways and a deep-dive option, with a 20-second reading time
04

Shape follows the task

A summary, a research brief, a plan and a brainstorm are different shapes of thinking. So each gets its own layout instead of one chat bubble.

01

Research & analysis

“Pros and cons of four-day work weeks.” Each finding becomes a card you can filter, instead of reading linearly.

Sources and confidence stay on every card, so trust doesn't cost extra scrolling.

Research findings on four-day work weeks shown as filterable pro and con cards with sources and confidence
02

Learning a concept

A six-step plan, one step at a time, with a progress bar and a checklist.

The mini-map answers “where am I?”, the structure Graphologue's users asked for.

Step 2 of 6 in a concept plan, with a progress bar and checklist
03

Creative brainstorming

Eight campaign ideas as tagged cards on a grid: bold, safe, weird.

Continue with FigJam turns the cards into sticky notes, so ideas leave the chat ready to sort.

Eight campaign ideas shown as tagged cards with a Continue with FigJam button
05

Tune it, or hear it

We can already choose how hard an AI thinks. Why not how it speaks? Reading Preferences reuse the effort-level pattern people already know, pointed at the output.

Reading Preferences panel with reading mode, complexity, summary length, font size and line spacing controls
Reading Preferences: mode, complexity, length, font size, spacing
An existing effort-level menu with Low, Medium, High, Extra and Max options
The familiar pattern: effort level
Narration player with playback speeds and chapters: instant answer, key takeaways, critical info, deep dive
Narration follows the layers. An 8-scroll answer becomes a 4-minute listen.
06

Where it's going

From concept to a working Claude plugin. PACE runs as an MCP server with an MCP Apps UI: the assistant writes the answer, PACE renders it in Standard or Cognitive mode. The same open standard lets it follow to ChatGPT.

  1. Now

    A manual toggle

    Cognitive Mode in any chat interface.

  2. Next

    Preferences that persist

    Your depth, your format, your voice settings.

  3. Beyond

    Output that adapts itself

    Shorter when you're skimming, audio when you're moving, visual when you're comparing.

Behind the designEvidence, readers & what I'll measure

The evidence

People scan. NN/g found users read about 20% of the words on an average page, and only half of pages with 111 words or fewer.

Models over-answer. YapBench traces it to preference-based training that favors longer outputs, and links lower cognitive load to stronger intent to keep using a chatbot.

Users ask for structure. In the Graphologue study, people found long, unstructured answers hard to navigate and wanted to see how ideas connect.

Who needs what

Skimmer → Cognitive Mode
Key line up top, sections collapsed
Listener → Audio player
Chapters per section, jump back a step
Low-load reader → Preferences
Spacing, font, chunk size, focus

Next, measured

  • Ship the Claude plugin to a small group of testers
  • Measure time to find the key point, Standard vs Cognitive
  • Bring the same build to ChatGPT
Next case study

From AI answers to the club floor.

Kiosk Returns: self-service returns that skip the line.

Kiosk Returns

Want to talk about PACE?

I'm building it now, and I'd love your feedback or to show you a live demo.

Quick read · about a minute

PACE AI, in a minute.

Founder, product designer & builder · 2026 · Claude plugin in build

ProblemAI answers bury the key point in long blocks of text that people skim or abandon.

SolutionOne Cognitive Mode toggle that reshapes the same answer to scan, follow or listen to.

Jump to a chapter

  1. 01The problem
  2. 02Who it's for
  3. 03Cognitive Mode
  4. 04Use cases
  5. 05Tune or hear it
  6. 06Where it's going