Maggie Appleton
"What app is that?" is always the wrong question

A strange thing

The two most-linked-to idea essays on maggieappleton.com are "A Brief History & Ethos of the Digital Garden" and "Tools for Thought as Cultural Practices, not Computational Objects" — the site's two load-bearing walls. Open the second one today, scroll to the middle of the piece, and you hit a grey box:

Draft in Progress. The quality of writing below this point is haphazard, disjointed, and nonsensical. It's probably a good idea to come back later.Tools for Thought as Cultural Practices, not Computational Objects

A few paragraphs further down, a note she plainly wrote to herself and forgot to delete is sitting in the body text: [need evidence for this claim]. And below that: "Obsidian comes out in xxxx. Logseq is released in xxxx." Both xxxxs are still there, waiting for years to be filled in.

The essay was published in July 2021. Its metadata says "Last tended almost 4 years ago." A core argument that her own site links back to repeatedly has been publicly hanging in a state of "needs evidence" and "xxxx" for four years. Two other long essays on the site end with the line "Coming Soon. Feel free to bug me on Bluesky to finish writing this." And the January 2020 piece "How to Draw Invisible Programming Concepts" opens by announcing a four-part series — across all 141 site updates from January 2019 to August 2026, Part II has never appeared.

If you read that as an abandoned site, you have it exactly backwards. Every page carries a status badge at the top: 🌱 seedling / 🌿 budding / 🌳 evergreen, plus a "planted" and a "last tended" date. Unfinishedness here is not a failure; it is a declared state, designed into the layout. Those two xxxxs are not oversights — they are the direct consequence of her position. They also happen to be its most expensive bill.

Who she is

Maggie Appleton is from London and grew up in international schools in Hong Kong, Vietnam, Thailand and Singapore. She studied cultural anthropology at what she describes as "a small, hippie, liberal arts college" in the US, and then — in her own words — anthropology "is not terribly employable (unless you want to be an academic or a military advisor)," so she "promptly switched into freelance design and illustration to pay rent."

One detail with a picture in it: she started learning visual design at fourteen, because she bootlegged a copy of Photoshop to make her own skeuomorphic desktop wallpapers and icon sets, and "never realised you could get paid for that." In her early twenties she country-hopped: working with web developers in Vietnam, training with feature-film illustrators in Los Angeles, learning typography and brand design in Prague. Her summary: "I made a lot of hideous stuff, but figured out what I liked along the way."

Then five years as illustrator and art director at Egghead, making 150+ course illustrations — a job that is literally turning abstract programming concepts into visible things. After that: Head of Design at HASH, founding product designer at the AI research lab Elicit, and since October 2025 a staff research engineer at GitHub Next, a team she calls "the department of fuck around and find out." Her bio is three words: designer, anthropologist, and mediocre developer. She is still writing; the most recent post is August 2026.

Glossary

What she is wrestling with

Across the five years of making those 150 illustrations at Egghead, the question she was asked most was not "how did you arrive at that metaphor?" It was:

"What app is that?"How to Draw Invisible Programming Concepts: Part I, 2020

She even wrote a post titled "What App is That?" to answer it once and for all. But her real answer is a different line: what everyone asks about is the whipped cream and cherries on top of the cake — "This happens in every industry – the most visible top layer of any creation feels like the most important part of it."

In August 2020 she turned that judgement into a harder essay: "Why You Own an iPad and Still Can't Draw." It opens with her actual situation — a pile of smart, determined friends who bought iPads in order to learn to draw and then messaged her asking which model, which app, what brush size. She says those questions "aren't wrong. In fact they're perfectly reasonable," and then splits drawing into three things: Material (the physical tools that make marks; an iPad plus Procreate is "one overtly expensive option," colored pencils on newsprint is "a far more financially reasonable choice"); Medium (visual language — composition, value, contrast, light, shadow, form, perspective, rhythm — which she says "are the equivalent of characters, verbs, sentences, introductory paragraphs, and satirical post-modern novels"); and Meat (what you're actually saying; she quotes a writing professor: "-39 points for not having a point").

Then the allocation: 2% Material, 18% Medium, 80% Meat.

That was a craftsperson's observation about drawing. But she took the same knife to the whole technology industry and found it wasn't craft experience — it was a general theorem: people systematically mistake the artifact for the practice. The app for the thinking. The model for the interface. The agent for team agreement. "We have the tool now" for "the problem is solved." Every core argument she has made since 2020 is that same knife landing somewhere new, and it lands the same way every time: concede that the artifact is genuinely useful, then point at the 98% it hides.

What lets her see this repeatedly isn't technical judgement, it's disciplinary displacement. She isn't a technologist looking at tools; she's an ethnographer who happens to live inside the programmer tribe — which is why the site contains pieces titled "The Cultural Anthropology of React" and "Fetishism & Mechanical Keyboards." Programmers are not her colleagues here. They're her field site.

Her ideas, told as how she found them

1. The digital garden: making "unfinished" a state you can publish in

On 15 April 2020 she posted a long Twitter thread collecting personal-wiki sites she and a friend had been gathering. Her own record: "It gathered a little buzz, and made clear we're in a moment where there is something culturally compelling about this concept."

What she noticed specifically was that these sites don't follow the conventions of the personal blog. Not a set of polished articles in reverse chronological order, but free-form, work-in-progress wikis; notes linked by contextual association rather than queued by date, and "published as half-finished thoughts that will grow and evolve over time."

Then she did the thing she always does: went and found out when the word was born. The result is three histories that don't know about each other. The first recorded use is Mark Bernstein's "Hypertext Gardens" in 1998 — though she says honestly that it reads "more of a manifesto on user experience flows and content organisation." The term then goes quiet for nearly a decade. The second stretch is the funniest: April 2007, Twitter barely open, and Rory Sutherland — she adds "oddly, the vice president of Ogilvy Group" — tweets "Digital gardening - ie faffing about syncing things, defragging - like pruning for young people." The next dozen mentions all mean exactly that: cleaning out folders, codebases, photo libraries. Her verdict is one line: "These people were digital puttering more than gardening." The third stretch is the real source: Mike Caulfield's 2015 talk "The Garden and the Stream" — "If anyone should be considered the original source of digital gardening, it's Caufield."

Her own contribution is collecting the scattered practice into six patterns, and the load-bearing one is the third, "Imperfection & Learning in Public." Her framing of it is the best passage in the essay:

In performance-blog-land you do that thinking and researching privately, then shove it out at the final moment. A grand flourish that hides the process.A Brief History & Ethos of the Digital Garden, 2020

The crucial part is that for her this pattern is not a licence but an obligation. Her words: "gardens make their imperfection known to readers." Hence the 🌱🌿🌳 badges and the two dates. She also abstracted it into two companion patterns: Epistemic Disclosure (state up front how you know what you know, how much effort went in, how sure you are — the precedents she names are Scott Alexander and gwern.net) and Assumed Audiences (a small box at the top saying who the piece is for — "Readers are a problem for every writer. Writers have to write for a specific person. That specific person is almost always imaginary").

Put the three together and you can see what she actually changed: not writing, but what the act of publishing means. Publishing no longer asserts "I think this is right and this is finished." It asserts only "here is my current state, labelled." And the payoff she offers the reader is the least technical-writing sentence imaginable:

They come to realise you are not a magical idea machine banging out perfectly formed thoughts, but instead an equally mediocre human doing The Work of trying to understand the world and make sense of it alongside you.A Brief History & Ethos of the Digital Garden, 2020

2. The phrase "tools for thought" contains no computer

This is the sharpest thing in her whole body of work. It starts from a corpus observation, not a position. By 2021 "tools for thought" had become a rallying point: meetups, research groups, "tools for thought is now a category tag in the databases of venture capital firms," and the phrase "appears on the landing page of every hot new knowledge management app." Those products clearly share a definition: write notes, connect them, explore dynamic views, then "experience a kind of emergent wisdom." Her one-line review: "An enchanting promise."

Then she points out the thing nobody had said, and points at it very plainly:

Taken at face value, the phrase tool for thought doesn't have the word 'computer' or 'digital' anywhere in it.Tools for Thought as Cultural Practices, not Computational Objects, 2021

She makes it stick in two steps. Step one: find out who coined it, and where that person stood. Kenneth Iverson; first public paper, "Notation as a Tool for Thought," ACM, 1979. Then she lists the field's repeatedly cited "major thinkers": Iverson, Licklider, Vannevar Bush, Alan Kay, Douglas Engelbart, Seymour Papert, Bret Victor, Howard Rheingold. Having listed them, the anthropologist's move — all male, all white, all North American, all attached to MIT, Caltech, Stanford, Berkeley, mostly in California from the 1960s through the 1990s. She adds: "This demographic is so overwhelmingly consistent that they're sometimes called the 'patriarchs' of tools for thought." And patriarchs have a side effect: most of them worked together, "or are at most one personal connection away from the next."

Tools for thought is a community scene as much as it's a concept. This gives tools for thought a distinctly computer-oriented, male, American, middle-class flavour.Tools for Thought as Cultural Practices, not Computational Objects, 2021

The force of this step is that it changes the kind of claim being made: the concept's computer-centrism is not a discovery about how people think, it is the fingerprint of one small social network.

Step two: put up the counter-list. She lists the things that actually expanded what humans can think, each with a date: written language (3200 BCE), drawing (~73,000 years ago), maps, Hindu-Arabic numerals, the Socratic method, the scientific method, Cartesian coordinates, zettelkastens, Aboriginal songlines, spreadsheets, data visualisation. Songlines are the best of them, because they push the word "tool" to its limit: actual physical paths through the Australian landscape, each with a corresponding song, encoding information about the land, political and social rules, and cultural knowledge — with her empirical footnote that they are "similar to method of loci, but shown to be more effective in memorising vast quantities of information." That is not a machine and not an app. It is a cultural technique passed down across generations. Her verdict on the whole list: "Every one of these also pre-dates digital computers by at least a few hundred years, if not thousands or tens of thousands."

What breaks the camel's back is a second observation from an entirely different corpus: around 2020, as note-taking apps proliferated, YouTube filled up with walkthroughs and comparisons — "The strange thing is, many of these guides are only superficially about the application they're presented in. Most are teaching specific cultural techniques: Zettelkasten, spaced repetition, critical thinking." The app is a delivery vehicle for a practice — and those practices cover a very narrow band of human activity: the things white-collar knowledge workers do. Hence the final rename:

If we're being honest about our current interpretation of TFT's, we should actually rename it to CMFWCKW – computational mediums for white-collar knowledge work.Tools for Thought as Cultural Practices, not Computational Objects, 2021

3. The dark forest and the reverse Turing test

On 31 December 2022, three weeks after ChatGPT shipped, she posted the most widely shared piece on the site. The structure is clean: borrow an existing theory (Yancey Strickler's 2019 dark forest theory of the internet), add a new variable, then invert the conclusion. The new variable is large language models, and her argument isn't abstract: she counts the copywriting products that appeared in the previous six months, notes they're good precisely because GPT-3's training corpus is web text — "It's intimately familiar with the style of language we use online" — and does the arithmetic: 100 thousand-word articles for two cents.

Then the inversion:

Our new challenge as little snowflake humans will be to prove we aren't language models. It's the reverse turing test.The Dark Forest and Generative AI, 2022

In the original test the machine has to prove competence. She made the consequence explicit on stage in 2023: "On the new web, we're the ones under scrutiny. Everyone is assumed to be a model until they can prove they're human."

She offers five ways to prove it — the best of them borrowed from linguistics rather than technology: what models have codified is la langue (the formal, dictionary-sanctioned language), and what's left to play with is la parole (the informal, creative speech acts of everyday life, "where language evolves"), which is why neologisms, jargon and in-group dialects become signals of humanness. And the essay's ache comes from one image and one admission. The image: proving yourself with obscure references, real-but-not-famous friends and recent events "feels eerily like a hostage holding up yesterday's newspaper to prove they are actively in danger. Perhaps a premonition." The admission is that the whole scheme simply doesn't work for some people — those with physical disabilities, those in remote rural places, those whose time is taken by caring for the very young or the very old. "They will have a harder time verifying their humanness online. I don't have any grand ideas to help solve this."

The more durable half appears only in the 2023 talk, because on stage she had to face the audience's strongest objection: models are often more accurate and more knowledgeable than most people, so why does it matter who made the content? Her words are "I'm sympathetic to this point," followed by three differences. First, a different relationship to reality — "The model cannot check its claims against reality because it can't access reality" — though not no relationship: "it's a bit like a fuzzy memory of reality." Second, a different social position: "there is no truth or knowledge outside of social context," and a model is not a person with a fixed identity — "They are not speaking from one stable social position." Third, the one nobody else says:

There is always someone on the other side of the work who you can have a full human relationship with. Some of us might argue this is the whole point of writing on the web.The Expanding Dark Forest and Generative AI, 2023

Then the deepest part of the talk, and the part only an anthropologist would say. Models represent "the generalised views of a majority English-speaking, western population who have written a lot on Reddit and lived between about 1900 and 2023. Which in the grand scheme of history and geography, is an incredibly narrow slice of humanity." She says that "If I put my anthropologist hat on and look at this problem it feels exceptionally tragic" — the whole discipline exists to expand our sense of how differently human cultures can run — and generating mass content from one worldview "funnels us down into a monoculture," shutting down the range of shapes human culture might otherwise take.

Finally she does the move she always does, shifting responsibility from the artifact back to the practice:

At this point I should make clear generative AI is not the destructive force here. The way we're choosing to deploy it in the world is. The product decisions that expand the dark forestness of the web are the problem.The Expanding Dark Forest and Generative AI, 2023

4. Barefoot developers, and the 2026 relocation of the bottleneck

In May 2024 she gave a talk at the Local-first Conference in Berlin. It begins with a deliberate choice of subject: she had spent a year watching what language models do to software development, saw a study putting Copilot users 55% faster, and decided not to talk about it — "With all due respect, we have it pretty good. Our problems are boring." The people she wants to talk about are elsewhere. The route to them is, again, genealogy first (Clay Shirky's 2004 "situated software," Robin Sloan's 2020 "An App Can Be a Home-Cooked Meal"), then collect examples and see what they have in common. The best of them is the glucose interface someone built for their diabetic partner, because the monitor's own interface "wasn't giving them the critical information they needed." The common thread is this sentence: "all of these examples of home-cooked apps deal with the kind of data you don't want leaving your personal home, like your children, bodily functions, health, and finances." All of those needs sit in the long tail: Google Maps will never support historical borders or tidal patterns, because "it doesn't make any financial sense to support the long tail."

Then the sharpest picture in the talk. What shape is the Venn diagram of people who want home-cooked software and people who can build it? Her answer: a concentric circle. Every example she had just shown was made by a professional developer.

Hence the name, and an analogy she handles carefully: the barefoot doctors of Mao-era China — villagers selected, trained, and sent back to serve the people they knew, "still barefoot peasants like everyone else, but now with more skills." By the mid-1970s there were over 1.5 million of them, and life expectancy over that period rose from 35 to 63. She adds a characteristic parenthetical: "I'm not saying everything Mao did was great, but this was a pretty good program. I'm sure he had very little to do with it."

Barefoot developers are the middle band: teachers making elaborate Notion setups to run classes, financial planning wonks producing extensive spreadsheets. They are "technically savvy and interested in solving problems for themselves and people around them, but don't want to become fully-fledged programmers." And they never get over what she calls the command line wall — the terminal is "a huge jump in complexity, usability, and frustration from using something like Airtable or Notion."

So what is actually missing? The answer is the same knife again:

Language model legos need glue.Home-Cooked Software and Barefoot Developers, 2024

The model hands you interface elements, state management, API calls and basic logic, but not how to stick them into a working application. "You still have no idea how you might deploy that to a particular web domain… Or persist data by setting up a database. You probably don't know what a database is." She also breaks with the entire end-user-programming tradition in the same talk: "I think we need to stop harassing normal nurses, teachers, and therapists to code when they don't want to."

Then the line's latest notch in 2026, which is the same move for the Nth time. Opening her April 2026 GitHub Next talk, she puts up a widely circulated demo video: a wall of terminal coding agents running in parallel on one person's machine. She calls the vision "one man, a two dozen claudes," then dismantles it in a sentence: "The main problem with this dream is it assumes software is made by one person." Believing individual output produces good software is, she says, "nine women make a baby in one month" logic.

Implementation is rapidly becoming a solved problem… The hard question is no longer how to build it. It's should we build it. Agreeing on what to build is the new bottleneck.One Developer, Two Dozen Agents, Zero Alignment, 2026

The mechanism is specific: development used to have alignment touch points scattered through it — Slack threads, issue comments, draft PRs — and they existed because implementation was slow enough to leave room for them. "But that implementation window has collapsed." The touch points all get pushed to the wrong side of implementation, which "leaves the poor pull request to carry the weight of all those checkpoints at the very end of the process when it's too late." Her claim — which she says sounds heretical coming from a GitHub employee — is that the PR and the issue are the wrong primitives. And most of the context alignment needs isn't in the codebase at all, it's "in people's heads": business context, the politics of who decides, organisational history. "Agents can never discover this context on their own."

How she thinks

1. Find out when the word was born. Every major piece starts with a dating exercise: digital garden → 1998 / 2007 / 2015; tools for thought → Iverson 1979; folk → Old English folc; bi-directional links → Bush 1945, Nelson 1965; barefoot → the barefoot doctors of the 1960s. Her stated reason: "Language is always a response to the evolving world around us… Naming is a political act as much as a poetic one." Why it works: a word's birth date and birthplace tell you whose problem it was invented to solve. The CMFWCKW punch was landed exactly this way.

2. Metaphor is a tool, not a decoration. She treats metaphors as operable parts, with notation to match: 🎯 marks the target (the thing being explained), ⛲️ the source (the thing explaining it). Then she takes her own illustrations apart on the page: "Vue-Router 🎯 is an underground railway system with train carriages travelling along it ⛲️. Each page in a Vue application 🎯 is a station ⛲️." She's wry about doing this: "I know this is a little like telling a joke, and then explaining why the joke was funny. It kind of sucks the fun out of it." But dismantling trains you to see consciously how metaphors are built — "You stop operating on autopilot." The crucial piece is the window-frame image: the source is a frame, what falls inside it is highlighted and what falls outside is hidden — "Which metaphor we pick depends on what qualities we want to highlight or hide." That part later becomes her critical weapon: because we made paper the metaphor for the web, we imported paper's limits, and left its best properties outside the frame (margin annotation, laying multiple sheets out to reorder them, carrying a page across a room to look at it with someone). Her prescription is a method: "practice seeing systems in multiple ways."

3. Count yourself in the sample. "Paleolithic Nostalgia" spends its length mocking contemporary romanticism about the Stone Age, then she brakes on herself: "This is not meant as an unkind, savage critique. I write this as someone who runs in barefoot shoes and tries not to eat things my great-great-great grandmother wouldn't recognize as food." The dark forest talk opens with "I'm Maggie and I want to lay out all my biases up front." Mid-argument in the tools-for-thought essay she inserts a disclosure that she's used Roam as her primary knowledge base since November 2019, "so my understanding of this history is certainly coloured by that affiliation."

4. Ask to be criticised. The last line of the barefoot developers talk: "I would really love constructive criticism about places you think I'm wrong. Just please be kind about it."

Together the four add up to a very consistent posture: she doesn't claim her conclusions are right, she claims her process is auditable. Which is also exactly where this approach is strongest and weakest.

Where she might be wrong

The garden ethos licenses permanent unfinishedness, and the bill has landed on her load-bearing pieces. The CMFWCKW rename is arguably her single best idea — and it sits inside the section she herself labelled "haphazard, disjointed, and nonsensical," in an essay where "[need evidence for this claim]" and two "xxxx"s have been hanging for four years. Epistemic disclosure is honest, but it is not a substitute for the argument: a reader knowing that a passage is unfinished does not make its inferences hold. Similarly, of the five reverse-Turing tactics, she herself dated the multimodal one — "This defence will probably only last another 6-12 months." The most-quoted part of an essay is often the part that keeps worst.

A conflict of position, declared every time, never reconciled. In April 2023 she asked a room: "if you are working on a tool that enables people to churn out large volumes of text without fact-checking, reflection, and critical thinking. And then publish it to every platform in parallel… please god, stop." By April 2026 she is a research engineer at GitHub Next demoing a multiplayer agent workspace whose entire premise is that a wall of parallel agents is the near-term normal. In fairness, that talk argues the opposite way — quality becomes the new differentiator, "Craftsmanship is what will set you apart from the vibe-coded slop." But she hasn't written the essay on whether the tools she now designs make the forest darker or lighter. It is the piece she is best placed to write and has the most reason not to.

The strong version of barefoot developers hasn't landed, and the analogy carries weight she doesn't check. Two years on, the visible phenomenon is "vibe coding," mostly among people who could already code; her own August 2025 note is titled "Vibe code is legacy code," relaying Steve Krouse's definition: "code that nobody understands." More importantly, the analogy itself: barefoot doctors worked because they were a program with selection, training and supervision. She explicitly notes the developer version "isn't going to be organized by a central government," and then doesn't ask what remains once you remove selection, training and supervision. In medicine, unsupervised practitioners is not a happy story.

Sample bias — diagnosed beautifully, never turned on herself. The best blow in the tools-for-thought essay is that the patriarchs' computer-centrism is the fingerprint of one small social network. Yet nearly every phenomenon she names — digital gardens, folk interfaces, barefoot developers — is generalised from a few hundred people in the same Twitter/Bluesky neighbourhood. She says so herself on stage: "I'm what we call 'very online'. I live on Twitter and write a lot online. I hang out with people who do the same." Naming a bias is not the same as correcting it. On others she uses etymology and demographic analysis; on herself she uses a disclaimer.

A necessary note about criticism. Her pages carry no comment threads, and discussion of her work happens mostly on Bluesky and Hacker News — both unreachable from the environment this page was written in (blocked at the network layer). So this page relays no critical argument it hasn't actually read. The most substantive external response I could read was Simon Willison's on the barefoot developers talk, and it is approving rather than critical — he calls it "a great talk."

Sources

• Tools for Thought as Cultural Practices, not Computational ObjectsJuly 2021. Dates the phrase, names the patriarchs, puts up a 60,000-year counter-list, and ends by renaming the field CMFWCKW. Also where the two xxxx's at the top of this page live.
• A Brief History & Ethos of the Digital GardenJune 2020. The most-linked-to piece on the site. Three etymological histories that don't know about each other, plus her six patterns; "gardens make their imperfection known to readers" is the foundation of her whole publishing ethic.
• The Expanding Dark Forest and Generative AIApril 2023 — full transcript and slides, an expansion of December 2022's The Dark Forest and Generative AI. The complete reverse Turing test, plus the three differences: relationship to reality, social position, and the possibility of a relationship.
• Home-Cooked Software and Barefoot DevelopersMay 2024 — full transcript. Situated software → the home-cooked meal → the concentric-circle Venn diagram → barefoot doctors → "language model legos need glue."
• How to Draw Invisible Programming Concepts: Part IJanuary 2020. The source of her method: metaphor source/target, the window frame, the 🎯⛲️ notation, and the four-layer cake. Read alongside the 2% / 18% / 80% in Why You Own an iPad and Still Can't Draw.

Updates