A reading list, taken apart

Software you reshape at the point of use. geoffreylitt.com

• Modification at the point of use. A user should be able to fix a mismatch while doing the work, without becoming a separate software project’s maintainer.
• A gentle slope to creation. Using, adjusting, combining, and programming should connect through small steps. A customization menu followed by a steep programming cliff breaks that progression.
• Tools compose around shared data. Small capabilities become more useful when they operate together on the same material. An isolated generated app can recreate the old silo.
• Agency can be communal. Everyone need not become a programmer. People can adapt tools with help from technically capable members of their own communities.

Malleable software · with Horowitz, van Hardenberg & Matthews · 2025

Your data lives on your machine. The network is optional. inkandswitch.com

• The local copy is primary. Work happens against data on your device. Servers can help synchronize it; access need not depend on a server granting each request.
• Offline and multiplayer belong together. Ownership need not mean isolation. Local work should synchronize across devices and collaborators when connections become available.
• Ownership is an experience. Speed, offline use, longevity, privacy, and control shape whether a creation feels like yours. An export button is only one piece.
• Sync needs something underneath it. CRDTs give a way to merge changes made at the same time on different copies. They support the idea; they do not deliver every ideal on their own.

Local-first software · Kleppmann, Wiggins, van Hardenberg & McGranaghan · 2019

Software cooked at home, by people who are not chefs. maggieappleton.com

• Local has a social meaning. Software can fit a household, classroom, or neighborhood. This describes whose situation it serves, separately from where its database lives.
• The barefoot developer. A technically curious community member can bridge ordinary users and professional engineers, applying intimate knowledge of local problems.
• Generated pieces still need glue. A convincing interface is not a complete service. Persistence, permissions, collaboration, and deployment must work together.
• Defaults determine ownership. People may never request portability or offline access. Authoring tools and agents can either provide those defaults or quietly reproduce platform dependence.

Home-Cooked Software and Barefoot Developers · 2024

An app can be a home-cooked meal. robinsloan.com

• Use value can be intimate. A tool serving a few loved people can be a complete success. Growth and monetization are optional goals.
• Known users simplify design. A specific social setting lets the maker omit features needed by a general platform. Fit can replace breadth.
• Ambient presence is a real need. Small exchanges can sustain closeness without demanding conversation. The emotional effect of software can matter more than feature count.
• Stability is a feature. Keeping a beloved interaction unchanged can be valuable. The people using the tool should influence when and why it changes.

An app can be a home-cooked meal · 2020, updated through 2026

Show the prompt. simonwillison.net

• Make a small complete tool. A single HTML file can contain interface, behavior, and styling. Fewer moving parts make distribution and revision easier.
• Use ordinary transfer surfaces. Copy/paste, file input, downloads, and URL state can connect a tool to existing work without elaborate integrations.
• Treat prior tools as ingredients. A working example gives the next tool a concrete starting point. A personal collection becomes a reusable design vocabulary.
• Publish the making process. Prompts and transcripts expose decisions and iterations, helping others learn how a tool was produced rather than only seeing its result.

Useful patterns for building HTML tools · 2025 · Tools colophon · prompts and development histories

File over app. stephango.com

• The artifact outlives its editor. Choose storage that remains usable when today’s application disappears. The work deserves a longer life than the tool.
• Control requires retrievability. A file is meaningfully yours when you can obtain, copy, and read it. Possessing opaque bytes provides weaker control.
• Plain formats reduce dependency. Simple, broadly readable formats make future access less contingent on a particular vendor or software generation.
• Future you is an audience. Preservation need not serve posterity or a public readership. Being able to revisit your own work is sufficient justification.

File over app · 2023

Personal computing from first principles. alexanderobenauer.com

• Make items first-class. Messages, tasks, events, and passages can be addressable objects in a shared environment, rather than trapped inside separate applications.
• Reference instead of duplicate. Transclusion embeds an item while preserving its connection to the original. Context survives reuse.
• Allow many paths to an item. A thing can participate in several projects and relationships. Finding it need not require remembering one folder location.
• Preserve surrounding context. Connections to people, events, and origins can provide additional routes for recall and navigation.

LN 014: The Graph OS · Lab Notes · research index

Builds the tools he thinks with. thesephist.com

• Fit the lived workflow. Building personal tools lets their boundaries follow actual habits, including the spaces between conventional app categories.
• Offer materials with consistent rules. Text, shapes, or commits become expressive when users understand how they behave. A coherent material can support unanticipated uses.
• Composition creates possibility. A few combinable primitives can enable many workflows. Adding another dedicated feature is not the only path to capability.
• Investigate representations. Interface innovation can change the objects people think with, rather than merely speeding up familiar commands.

Build tools around workflows, not workflows around tools · 2020 · Design with materials, not features · 2022

Small personal software, syndicating vibes. interconnected.org

• Abundance moves the bottleneck. When making tools becomes easier, finding, remembering, and tracking useful tools become harder.
• Follow makers as well as products. A person’s taste can organize discovery across many tiny creations. Tool publishing starts resembling an ongoing creative practice.
• Syndication can distribute discovery. Feeds and interoperable metadata could help people aggregate tools without submitting everything to one central registry.
• Software updates complicate feeds. Apps have platforms, versions, and executable consequences. Following an update is not identical to reading another blog post.

We need RSS for sharing abundant vibe-coded apps · 2026

Two people, a boat, and the tools they wrote. 100r.co

• Constraints are design inputs. A small boat-based studio explores resilience and self-reliance through low-tech solutions and minimal technological tooling.
• Custom tools are creative instruments. Making software can extend drawing, music, and games. The tool’s design becomes part of the artistic practice.
• A living archive guides work. A record of making and learning can preserve context and help determine what to explore next.
• Build continuity into distribution. Their updates document mirroring source across repositories after disruption. Preserving access is ongoing work. Choosing a license is only the start.</li>

Devine’s React Berlin interview · author-hosted transcript · 2019 · Hundred Rabbits project updates · official repository

Books do not work. Memory can be engineered. andymatuschak.org

• One note, one coherent idea. An atomic note isolates a reusable concept while retaining enough context to make sense. Smaller is not always better.
• Connections enable reuse. Separating concepts makes it easier to link an old idea into a new context. Over-fragmentation can obscure relationships.
• Retrieval is a designed activity. A good question asks for a specific, checkable act of remembering. Rereading an answer is a different activity.
• Effort must remain tractable. Prompts should require real recall without becoming repeatedly impossible. Tight scope makes difficulty easier to diagnose and adjust.

Evergreen notes should be atomic · How to write good prompts · 2020

A medium that remembers for you. michaelnielsen.org

• Integrate remembering into reading. The mnemonic medium embeds retrieval prompts inside an explanation, reducing the separate work required to create a review practice.
• Continue beyond the first encounter. Scheduled revisits make learning a repeated relationship with material rather than a one-time reading event.
• Fluency supports later reasoning. Internalized concepts and notation reduce the burden of understanding ideas that depend on them.
• Invent media, not only utilities. A medium’s elementary objects and actions can expand the kinds of thought available to its users.

How can we develop transformative tools for thought? · coauthored with Andy Matuschak · 2019 · Augmenting Long-term Memory · 2018

Learn hard things fast, on purpose. scotthyoung.com

• Self-direction concerns control. Choose what you are trying to learn and organize effort toward it. A personally chosen goal still permits structured instruction.
• Practice the target performance. Learning should include doing the kind of thing you eventually want to do. Do not assume the skill transfers to a different setting.</li>
• Drill the limiting component. Separate a troublesome subskill for focused work, then bring it back into the whole performance.
• Revise appealing learning theories. Young describes moving away from overvaluing associations and toward examples, retrieval, and substantial practice.

You Do Have Time for Ultralearning · 2019 · How My Views on Learning Have Changed Over Time · 2023

Learning science, applied at scale. justinmath.com

• Automaticity frees attention. Practiced component skills require less conscious handling, leaving more attention for the new or difficult parts of a problem.
• Prerequisites form a dependency graph. A learner’s readiness depends on specific underlying skills. Being assigned the same grade level does not imply the same readiness.
• Learning and maintenance compete for time. Choosing new tasks also requires tracking which previous skills are fading and which get refreshed indirectly.
• Adaptation is a scheduling problem. The proposed system combines knowledge structure and individual performance to select useful practice, rather than serving one sequence to everyone.

Cognitive Science of Learning: Developing Automaticity · 2024 · The Main Value Proposition of Math Academy · 2026

Always invite AI to the table. oneusefulthing.org

• Assign a pedagogical role. Tutor, coach, mentor, teammate, tool, simulator, and student are different roles. Their success criteria should differ too.
• Coaching can elicit reflection. An AI can ask for concrete experiences and follow-up reasoning, helping a learner examine their process rather than merely receive information.
• Tutor behavior must be challenged. Try wrong answers, misconceptions, and demands for solutions. A plausible conversation can conceal factual or pedagogical failure.
• Teachers retain a design role. Someone must choose goals, evaluate the interaction, and monitor failure. A prompt alone is not an educational system.

Assigning AI: Seven Ways of Using AI in Class · 2023

You can turn the crank yourself. ciechanow.ski

• Expose hidden structure. Rotation and disassembly make interior relationships inspectable. The explanation reveals what a normal view conceals.
• Build complexity progressively. Start with components and their jobs, then reconnect them into a working mechanism. Readers need not understand the whole machine immediately.
• Let readers control inspection. Sliders and viewpoint changes let readers examine a state at their own pace instead of following an unchangeable animation.
• Coordinate words and pictures. Consistent colors and nearby explanations help readers connect names, parts, and behavior.

Mechanical Watch · 2022

Play it, then believe it. ncase.me

• Make a strategy executable. Rules such as reciprocation become behaviors that can interact, making abstract game-theoretic ideas concrete.
• The interaction horizon matters. Repeated encounters change incentives. Assumptions about the final round affect what strategic reasoning predicts.
• Distinguish levels of change. A population changing its mix of strategies differs from an individual strategy changing its own behavior.
• Put caveats beside the model. The notes identify underlying books, strategy names, and modeling distinctions, giving readers a route beyond the simplified experience.

The Evolution of Trust · 2017 · The Evolution of Trust: Feetnotes

The hub for the genre. explorabl.es

• A broad family of forms. Games, sandboxes, and interactive articles can all explain. The genre is defined more by learning activity than by one visual style.
• Play should expose the subject. The interaction ought to help someone learn, rather than merely decorate otherwise static material.
• Curiosity can motivate participation. The hub emphasizes autonomy, mastery, and discovery over attaching points to routine tasks.
• Examples support a making community. A collection of works, tools, and tutorials helps new authors borrow techniques and invent their own explanations.

Explorable Explanations · FAQ · Genre hub and examples

The origin. worrydream.com

• Text can be a thinking environment. Readers can do more than accept or reject claims: they can manipulate examples and investigate consequences.
• Make assumptions adjustable. A reactive document reveals how an argument’s result changes when its inputs or premises change.
• Make abstractions concrete. An interactive example provides a manipulable instance of an abstract relationship, supporting the formation of intuition.
• Offer context at the moment of need. Nearby definitions and evidence help readers investigate without losing their place in the argument.

Explorable Explanations · 2011, postscript 2024

Culture, argued with data. pudding.cool

• Start with audience and purpose. The intended reader changes the explanation’s vocabulary, structure, and editorial choices.
• Story structure is part of analysis. Choosing what to reveal and in what order helps people interpret data. Chart selection alone does not supply a narrative.
• Coordinate visual and verbal work. Words can guide attention while graphics carry relationships that would be cumbersome to describe.
• Choose interaction for a reason. Scrollytelling can coordinate a sequence, but adds rendering and state-management costs. The form must earn those costs.

Making Internet Things, Part 3: Storytelling · Ilia Blinderman · About The Pudding

Curiosity, written down. jvns.ca

• Find the hidden fact. A basic mechanism can be surprising to someone who has never encountered it. Expertise can make teachers overlook that gap.
• Select the useful subset. A concise reference earns its place by choosing the most useful facts, not by shrinking an exhaustive manual.
• Explain what becomes possible. Useful explanations reveal either a new action or why a confusing behavior occurs.
• Clarity needs understanding. Simple drawing tools can suffice. The difficult work lies in choosing and explaining the idea accurately.

How I write useful programming comics · 2020 · About Julia Evans

Small tools, honest notes. macwright.com

• A tool can outlive its company. Placemark’s transition to open source lets useful work continue after the original business closes.
• Dependencies affect practical freedom. An application’s own license is not the whole story. The licenses and service requirements underneath it shape what others can reuse.
• Maintenance can arrive through community. An outside contribution can remove old constraints and improve a tool beyond the original maker’s effort.
• Reuse crosses project boundaries. Code can move into another tool, preserving value even when a particular product’s trajectory changes.

Placemark is now fully open source · 2026

Measure before you believe. danluu.com

• Measure an unanswered question. A commonplace assumption can remain untested for years. Curiosity can identify valuable measurement work without requiring a novel product.
• Absence of reports is weak evidence. A system may fail in ways nobody checks. Confidence based only on not noticing a failure can be misleading.
• Choose a relevant benchmark. Measure the behavior people actually care about. A convenient technical proxy can miss the experience being evaluated.
• Public evidence changes incentives. Reproducible failures and comparisons can make neglected problems harder to dismiss and give improvement a visible target.

Measurement, benchmarking, and data analysis are underrated

A site as a research practice. gwern.net

• Write for future reconstruction. Preserve why a conclusion seemed reasonable, including evidence and uncertainty, so a future self can revisit the reasoning.
• Let essays grow over time. An evolving page can accumulate additions and corrections rather than becoming an abandoned dated post.
• Collect contrary evidence. Notes provide a place for observations that complicate a thesis, not just those that support it.
• Make preservation part of authorship. Durable text sources and long-term site design support work that is worth improving over years.

About This Website