How I work
How I build products, straight from my second brain: what I believe about testing, deployment and state, how I set up a project for AI agents, and the tools I build with. AI assistants can read it too, through my MCP server.
- Decide where a thing lives before choosing a tool.State goes in a layer, a test at a boundary, an event at a moment that matters.
- Make it verifiable.A failing test, a landed event and a build that fails on a bad migration beat a green light that might lie.
- Keep feedback cycles fast.Find out in seconds whether something worked: fast tests, quick builds, a preview per PR. A slow loop costs the focus that never comes back.
- Keep it runnable by an agent.CLI over dashboard, done-conditions as files. Whatever an agent can run, it can check.
Testing
A test is proof that behaviour holds, written to fail when it breaks. Judge a suite by what it catches, not its coverage.
- FewEnd-to-end. Critical paths: auth, payments, onboarding. Playwright
- MostIntegration. Features with their real dependencies, used the way a user would. Vitest, Testing Library
- ManyUnit. Pure logic: validation, transforms, calculations. Vitest
- AlwaysStatic. Types and lint rules, on every save. TypeScript, ESLint
- 1
Test behaviour, not implementation. Test what users see: text, buttons, forms. Those tests survive refactors.
- 2
Fake only at the boundaries. Every mock is a claim about the other side. Keep the claims at the edge and verify them with contracts.
- 3
Write mostly integration tests. Kent C. Dodds' testing trophy: integration tests are the bulk, because they run your real code and the wiring is where bugs hide. On the backend, that bulk is scenario tests on the core with fakes at the edge.
- 4
A bug fix starts red. Failing test first, then the fix. In review, a failing test beats a comment.
- 5
Coverage doesn't prove a test works. A test can run every line and still pass on broken code. Mutation testing checks the tests.
- 6
Test states production can reach. Hand-built objects create states that never exist for real.
- 7
With agents, tests are the constraints. A comment asks, a failing test enforces. Write the done-condition before the agent writes code.
- 8
Spend checks on purpose. Fast checks early and often, heavy ones near the PR. Drop a check that never catches anything.
- 9
Tests are docs that can't rot. Integration tests show how the system works today. Point the docs at them.
Deployment
A deploy should be boring. Code reaches production all the time; users see a change only when I release it.
- 1
Deploy is not release. Feature flags keep them apart, so every deploy is low-stakes.
- 2
Merging to main is the deploy. No release ceremony. A branch and a PR per change, even solo.
- 3
Every PR gets a preview, kept out of search. Preview URLs are public, so they're marked noindex.
- 4
Migrations run on the build. A failed migration fails the build, so a broken schema never goes live.
- 5
Config lives on the platform, not in git. Secrets and env vars are set per environment.
- 6
CLI over dashboard. Repo, deploy, env vars and merge run from the terminal, so an agent can run them too.
- 7
Know where the runtime differs from local. Edge runtimes, deployment protection and cron billing break things that work on your machine.
- 8
Assess risk before a release. Give an AI the diff and the incident history, and ask what's most likely to hurt.
State
Where state lives matters more than which library manages it. URL, then server, then local, then global: each layer catches what it can, so the global store stays nearly empty.
- 1
The URL is the first store. If copying the link should reproduce the view, the state belongs in the URL.
- 2
The path names a thing, the query adjusts it. Records get a path. Filters, sort and view mode get params.
- 3
The server owns server data. Render and cache it in Server Components. TanStack Query only when the client must refetch live.
- 4
One owner per piece of data. Never copy server data into useState or a store. The copy goes stale.
- 5
Compute derived state, don't store it. A filtered list or a total is calculated during render.
- 6
Keep client state in the leaves. Local first, then a common parent, then global.
- 7
Global state is what falls through. Jotai inside one React tree, Nanostores when state crosses roots or frameworks.
- 8
On the server, a module is shared between users. A module-level store leaks one request's state into another.
- 9
Sometimes there's no React state at all. A data attribute and CSS can filter a list without a re-render.
Agent-ready setup
Set a project up so AI agents can do most of the work and check it themselves. The setup around the model matters as much as the model.
- 1
The harness matters as much as the model. Instructions, tools, hooks and a sandbox turn a model into an agent. A decent model in a strong harness beats a great one in a weak harness.
- 2
Start from a scaffold, not a blank folder. One command sets up a new project with my default stack. The Adopt ring of my tech radar decides what goes in.
- 3
Write knowledge down where the agent reads it. Instruction files in the repo, and a second brain the agent searches before it debugs or picks an approach.
- 4
Write the done-condition first. A failing test or a check the agent can run, before it writes any code.
- 5
Make the architecture a lint error. When an agent takes a shortcut across modules, the linter fails and the agent fixes its own code.
- 6
Let the agent run what it builds. Give it a scriptable way to start, use and screenshot the app, so only work that survived its own run reaches me.
- 7
Turn repeated work into skills. A script wrapped as a named skill is a capability an agent can rerun forever, under version control.
- 8
Agents speak, workflows act. In a pipeline the agent only comments. A deterministic workflow moves the labels and merges. The agent has opinions; the workflow has permissions.
- 9
From issue to preview, with a human at the gates. A one-line issue becomes a spec, a tested PR and a preview. I answer questions, sign off and review.
- 10
Earn autonomy per kind of change. Low-risk changes merge on evidence, the rest get read. Clean PRs raise the bar for what merges unread; escaped bugs lower it.
Tech radar
The tools I build products with. Each tool sits in a ring: how confident I am in it, not a recommendation for you.
The Adopt ring is not just an opinion. My project scaffold reads it, so a tool that moves to Adopt ends up in the next project I start. Updated October 30, 2026.
- AdoptBattle-tested. The default in new projects.45
- TrialPromising. Tried on non-critical paths.6
- AssessWorth watching. Not ready to commit.4
- HoldUsed before. Moving away from it.2
All 57 tools
Languages & frameworks
- 1TypeScriptAdoptNon-negotiable. strict: true, no exceptions.
- 2PythonAdoptFor scripts, data work and backends where the libraries live in Python.
- 3Next.jsAdoptFile-based routing, Server Components, Server Actions and built-in optimizations.
- 4FastAPIAdoptPython APIs where type hints define validation, serialization and the OpenAPI docs.
- 5OpenNextAdoptRuns Next.js on Cloudflare Workers.
- 6TanStack StartTrialType-safe full-stack React, and the portable alternative to Next.js.
- 7TanStack RouterTrialType-safe React router. The foundation of TanStack Start.
- 8TauriAssessDesktop apps with web technology, a few megabytes instead of Electron's 150.
UI
- 9shadcn/uiAdoptCopy-paste components you own, built on Base UI or Radix, and Tailwind.
- 10Tailwind CSSAdoptUtility-first CSS. No context-switching between files.
- 11class-variance-authorityAdoptDeclarative, type-safe component variants.
- 12LucideAdoptClean, tree-shakeable icons. The shadcn default.
- 13Tailwind VariantsTrialThe cva idea with class merging and multi-part slots built in, and still shipping.
- 14MobbinTrialScreens and flows from real apps, searchable by pattern. Agents can search it too.
- 15Base UIAssessUnstyled, accessible React primitives. shadcn's default since 2026, so you likely have it already.
State
- 16JotaiAdoptAtomic, bottom-up state. No providers wrapping providers.
- 17NanostoresAdoptTiny atoms outside any component tree. Shares state between islands, in any framework.
- 18nuqsAdoptType-safe search params. State that survives a refresh.
- 19TanStack QueryAdoptFetching, caching and syncing server state, so you don't have to.
Forms & validation
- 20ZodAdoptSchema-first and TypeScript-native: runtime validation plus type inference.
- 21PydanticAdoptZod for Python: declare a model, parse input into it or get a clear error.
- 22React Hook FormAdoptPerformant, uncontrolled forms. Pairs with Zod.
Utilities
- 23validator.jsAdoptBattle-tested string validators. Pairs with Zod via .refine().
- 24date-fnsAdoptModular, tree-shakeable, immutable dates.
Auth
- 25Better AuthAdoptEmail, social login, sessions and roles, with plugins for the rest.
Testing
- 26VitestAdoptVite-native, fast, ESM-first. Jest without the config pain.
- 27React Testing LibraryAdoptTest like your users: query by role, not implementation.
- 28PlaywrightAdoptCross-browser, reliable, auto-waiting. Replaces Cypress.
- 29MSWAdoptNetwork-level interception. The same handlers for dev and test.
Tooling
- 30pnpmAdoptFast, disk-efficient, strict. No phantom dependencies.
- 31uvAdoptOne Rust binary that replaces pip, virtualenv and pyenv. Fast enough to forget.
- 32BiomeAdoptOne Rust binary for linting and formatting.
- 33RuffAdoptBiome for Python: linting and formatting in one Rust binary.
- 34GitHub CLIAdoptPull requests, issues, releases and API calls from the terminal.
- 35ESLintHoldReplaced by Biome. Kept only where a custom rule has no Biome equivalent.
- 36PrettierHoldReplaced by Biome: one tool for linting and formatting.
AI tooling
- 37Claude CodeAdoptClaude in the terminal, driving files, shell, git and MCP tools.
- 38Claude Agent SDKAdoptClaude Code's agent loop, under your program's control.
- 39MCPAdoptThe standard way agents connect to tools and data. The connection, not the knowledge.
- 40Claude Code GitHub ActionAdoptClaude Code inside GitHub: reviews PRs, implements issues, answers @claude.
- 41Claude ArtifactsAdoptPages Claude publishes from the terminal. Private by default, shareable by link.
- 42HerdrAdoptA terminal multiplexer that knows which panes run agents and lets you steer them.
- 43HunkTrialA diff viewer built for reviewing, not paging. Made for reading what agents changed.
- 44WorktrunkAssessGit worktrees made easy, for running several agents in parallel on their own branches.
Database
- 45SQLiteAdoptA database that is just a file. No server, no ops, and enough for most products.
- 46TursoAdoptManaged SQLite over HTTP. A local file in dev, a remote database in production, same code.
- 47DrizzleAdoptSQL-like queries, schema as code, zero overhead.
- 48SQLAlchemyAdoptDrizzle for Python: a SQL expression language with an ORM on top.
- 49AlembicAdoptMigrations for SQLAlchemy, generated by comparing your models with the database.
- 50FTS5AdoptFull-text search built into SQLite. No Elasticsearch.
- 51LitestreamTrialStreams SQLite changes to S3: durable backups for a single server.
- 52Cloudflare D1AssessManaged SQLite on Cloudflare, reached as a Worker binding instead of a URL.
Hosting
- 53GitHubAdoptRepositories, CI and code review. The default home for code.
- 54VercelAdoptDeploys, env vars, logs and domains, all from the CLI.
- 55Cloudflare WorkersAdoptDeploys with Wrangler, with storage and databases as bindings.