Experiment · 2026Parked until the next spark

Pocket

Pocket is an app I built to remember things for me. I could throw anything into it, photos, voice notes, links, bits of text, and it would work out what each one was and connect it to what I’d saved before. It was the equivalent of asking someone: what have I eaten this week? What was that exhibition with the flashy pink portrait of a unicorn?

Pocket app: a 'night market walks' note from Bangkok with a photo, a voice note transcript and its connections
// Why it exists

Everything physical had a place. The digital didn’t.

Everything I collect has its spot: the dinosaur shelf, the rocks, the business cards. My digital things didn’t. Photos, notes, Substack saves and voice memos each live in a different “pocket” of the internet.

Every tool I tried wanted me to organise things as I saved them, and I never kept it up. So I imagined one place to save everything, where meaning builds up from what piles up.

It’s designed for one person: me.

// The premise

Meaning emerges
from accumulation
not filing

// Language

A travel notebook, not a filing cabinet

Pinang Peranakan Mansion in Georgetown, Malaysia gave me the inspiration for the colour palette and the mood. I wanted Pocket to feel like I was there, browsing through this 19th century mansion surrounded by Chinese carved woodwork, Nyonya shoes and antique British glassware.

The model’s voice follows the same line. Natural sentences, specific details, no clichés, no rhetorical questions. I tested different versions of the tone that could be adjusted via the settings panel, only keeping what read right.

// How it works

Save first, make sense of it later

Saving takes one tap. The AI reads, compares and judges; fixed rules filter; I approve.

The aim was to teach Pocket how I associate things, not necessarily what they are on paper: a photo of a dog and a voice note recorded at a market belong together, because of where I was that day and the sentiment behind it.

MEI decide
POCKETReads and judges
RULESFixed logic, principles
01Save somethingA photo, voice note, link or text. One tap, nothing to file.
02Read itWrite a short summary, add tags and categories, read text in images, transcribe voice.
03Capture its meaningUnderstand what the thing is, and how it fits with all the collected nodes.
04Find nodesPull the saved nodes closest in meaning and any that share tags.
05Judge each connectionWith my taste profile as a guide, score it and write a one-line reason.
06Automatic filterKeep strong and moderate matches. Checks the reason is specific.
07Approve or rejectEach suggestion arrives with its reason. I can rewrite the reason before approving.
IF APPROVEDThe connection joins the graph
08Count decisionsEvery eighth decision updates what Pocket has learned about me.
09Rewrite my taste profileSum up what I approve and reject. Saved as a log I can read.
SKIPPEDAlready connected, or archivedNever reaches the AI.
DROPPEDWeak matchNever shown to me.
DROPPEDVague reasonLike “both are about travel”. Duh.

Refines the next judgment

Whether a connection is approved or rejected, it still counts toward the learning feedback loop.

// Decisions. Decisions. Decisions.

Trade-offs

Every fork below shows the path I took and didn’t, its cost, and where it led to. Below, the two walls the decisions added up to → temporarily parking the Pocket.

Build

How it was made

01How do I build it?One HTML file on the web, changed through conversations with Claude
→ ALTERNATIVENative iOS appSlower to change, no iterating from my phone
LED TOFast iteration. Later, a speed ceiling
Capture

Saving nodes

02What happens when I save a node?No decision on where to file. One tap and Pocket adds a summary, tags and connections after
→ ALTERNATIVEFolders or tags at captureThe friction that killed every tool I tried before
LED TOEvery save waits on several AI steps in a row
05What if Pocket gets it wrong?My edits always win. Edited summaries won’t automatically be changed. My tags are merged in.
→ ALTERNATIVEThe AI reprocesses freelyMy fixes would get overwritten, so I’d stop correcting
LED TOEdits become the strongest signal
Connect

Linking nodes

03Who makes a connection?Pocket suggests, I approve. Each suggestion comes with a reason
→ ALTERNATIVEPocket connects on its ownOne wrong link and I stop trusting the graph
LED TOFriction at every connection. But justified.
04How does it get to know me?It learns from my decisions. The taste profile keeps getting updated and I can keep an eye on it.
→ ALTERNATIVEHard-coded rulesGeneric suggestions, the way that I associated topics could not have been learned
LED TOSuggestions get more accurate with use
Search

Finding nodes

06What did I save last week?Dates and places first, then related items
→ ALTERNATIVESearching only by what feels relatedOnly returned one result for last week
LED TOComplete answers for time and place
07What does the answer look like?The app picks the right display depending on the source: a list, a gallery, a map.
→ ALTERNATIVEOne list for everythingPhotos and reflections buried in the same list
LED TOOne model call per search
Voice

How Pocket speaks

08How does Pocket sound?One voice for every piece of communication with me
→ ALTERNATIVEEach prompt with its own voiceThree tones, clichés, rhetorical questions
LED TOThe app reads as one cohesive experience, increasing trust

Which eventually led to..

Wall 1 · from build, capture, searchSpeed

Several model calls per capture, all in the browser. Faster after every fix, never as fast as a camera.

Wall 2 · from constant useThe bigger question

Most of what I save is already in my photo library, with a date and a place. Should this be a feature of the iOS gallery, not an app?

Hence..Parked. For now. Until the next spark

I got to version 7.84.

// Me and Pocket

What the robot does, and where I stopped it

The model’s good at reading and capturing, but it bluffs. And misses context. So the work is split: it reads everything I save, and I decide what joins the graph.

It does
Connections across formats

It links things by what they’re about, so a photo can connect to a voice note with no tags in common.

Photos become text

Every image gets a short summary, so a photo of a street stall can be found later by typing “noodles”.

Voice notes in two languages

Whisper transcribes without being told the language, so a note that switches between Italian and English comes back intact.

I hold it back
Nothing joins without me

Every connection is a suggestion until I approve it. One wrong link and the whole graph feels untrustworthy.

It has to explain itself

Each suggestion comes with a one-line reason, not a score, so I can agree, disagree or rewrite it.

Weak guesses never reach me

Every match is scored strong, moderate or weak. Strong and moderate come to me. Weak ones are dropped before I see them.

// Learnings

What I’m taking with me

  1. 01

    Designing for one person is a fast start, not an answer

    I could make every decision in minutes because I was the user. I still don’t know if anyone else needs it.

  2. 02

    What the AI leaves out matters as much as what it suggests

    Dropping weak connections did more for trust than improving the strong ones.

  3. 03

    Speed can’t be added later

    The choices that made Pocket fast to build are the same ones that capped how fast it could run.

  4. 04

    Cheap building makes it easy to skip the hard question

    I got to version 7 before I seriously asked whether this should be an app at all.