# Product Radar

> An opinionated, weekly radar of the models, tools, infrastructure and practices that change what one AI-enabled person can build, ship and run.

Edition 01, 12 Oct 2026. Rings: Adopt (a default choice), Trial (use where risk is manageable), Assess (watch), Hold (avoid for new work). Full data: https://productradar.org/radar.json

Note: this is a preview. Draft: built from the tools and practices I use every day. I am confirming each verdict this month.

## Adopt

- [Claude Fable](https://productradar.org/e/claude-fable/) (Models & Agents): Anthropic's Claude model aimed at the hardest knowledge work and coding problems. Fable is my main model. It plans the work with me and hands the building to other agents.
- [Claude Opus](https://productradar.org/e/claude-opus/) (Models & Agents): Anthropic's Claude model family for complex coding and long agent tasks. Opus writes the code in my agent builds. Each ticket gets a fresh Opus agent.
- [Claude Sonnet](https://productradar.org/e/claude-sonnet/) (Models & Agents): Anthropic's mid-size Claude model family for everyday coding and agent work. Sonnet gathers evidence for my builds and reviews risky changes with a fresh context.
- [agent-browser](https://productradar.org/e/agent-browser/) (Tools & Harnesses): A browser automation CLI built for AI agents. My agents use agent-browser for scripted browser work. Pages that need my logins run in a separate Chrome profile kept for agents.
- [Claude Code](https://productradar.org/e/claude-code/) (Tools & Harnesses): Anthropic's coding agent that works in the terminal, the IDE and the desktop app. I run my code, email, calendar and task board through Claude Code every day, with my own skills and guard hooks.
- [FluidVoice](https://productradar.org/e/fluidvoice/) (Tools & Harnesses): An open-source macOS dictation app with on-device speech recognition. I dictate most of my prompts with FluidVoice, which runs locally on my Mac.
- [Linear](https://productradar.org/e/linear/) (Tools & Harnesses): An issue tracker and product development tool for software teams. Kaufmann Health tickets live in Linear. Agents pick up a ticket and build it end to end.
- [Obsidian](https://productradar.org/e/obsidian/) (Tools & Harnesses): A Markdown note-taking app that stores notes as local files. My strategy docs, decisions and task board live in an Obsidian vault. Agents read and update it as plain Markdown.
- [GitHub Actions](https://productradar.org/e/github-actions/) (Infrastructure): GitHub's CI/CD and workflow automation service. Every Kaufmann Health change runs through GitHub Actions, with staging tests and a daily security audit.
- [Parallel](https://productradar.org/e/parallel-web/) (Infrastructure): Web search, extraction and research APIs built for AI agents. Parallel is my agents' default for web search, page extraction and research.
- [Resend](https://productradar.org/e/resend/) (Infrastructure): An email sending API for developers. Kaufmann Health sends its product email through Resend.
- [Stripe](https://productradar.org/e/stripe/) (Infrastructure): A payments platform. Stripe Connect handles payments and payouts for platforms and marketplaces. Kaufmann Health takes payments and pays out therapists through Stripe Connect.
- [Supabase](https://productradar.org/e/supabase/) (Infrastructure): A hosted Postgres platform with auth, storage and APIs. Kaufmann Health runs its Postgres database and file storage on Supabase in production.
- [Vercel](https://productradar.org/e/vercel/) (Infrastructure): A cloud platform for deploying web apps, with first-class Next.js support. Kaufmann Health runs on Vercel, including its scheduled jobs.
- [Agent skills](https://productradar.org/e/agent-skills/) (Practices): Packaged instructions and scripts that an agent loads when a task matches them. Every repeated job gets a skill, so the next agent does it right without asking me. Claude Code and Codex share the same skills.
- [AGENTS.md instruction files](https://productradar.org/e/agent-readable-repos/) (Practices): A plain Markdown file that tells coding agents how to work in a repository. One git-tracked instruction file serves both Claude Code and Codex, and each repo adds its own, so both agents follow the same rules.
- [Ask-first gate](https://productradar.org/e/ask-first-gate/) (Practices): Letting agents ship work end to end while a short list of change types waits for the owner's approval. Agents ship to production on their own, except changes to money, payouts or personal data. Those wait for one yes from me.
- [Company brain](https://productradar.org/e/company-brain/) (Practices): Keeping a company's strategy, decisions and facts in documents that agents read before they act. My strategy, standing decisions and facts live in documents that agents read before they act. They ask me only what the documents do not answer.
- [Guard hooks for agents](https://productradar.org/e/guard-hooks/) (Practices): Scripts that run before or after an agent's tool calls to block or flag risky actions. Hooks check commands, file reads and email sends before they run, and the agent cannot switch them off.
- [Model tier per task](https://productradar.org/e/model-tiers/) (Practices): Assigning each agent task to the smallest model that can do it reliably. Each agent gets a pinned model. Haiku does small checks, Sonnet gathers evidence and Opus writes code.
- [Parallel agents in worktrees](https://productradar.org/e/parallel-agents/) (Practices): Running several coding agents at once, each in its own git worktree and its own context. A lead agent splits the work, and a fresh agent builds each ticket in its own worktree. Its context ends with the ticket.

## Trial

- [Claude Haiku](https://productradar.org/e/claude-haiku/) (Models & Agents): Anthropic's smallest and fastest Claude model family. Haiku runs small checks inside my agent builds and drives well-defined browser tasks.
- [Gemini Flash](https://productradar.org/e/gemini-flash/) (Models & Agents): Google's fast Gemini model tier for high-volume tasks. Gemini Flash picks the quoted passages on isaiconsciousyet.com. It copies text and never writes a quote itself.
- [Nano Banana Pro](https://productradar.org/e/nano-banana-pro/) (Models & Agents): Google's Gemini image generation and editing model. I make brand and ad images with Nano Banana Pro.
- [Codex](https://productradar.org/e/codex/) (Tools & Harnesses): OpenAI's coding agent, available as a terminal CLI and inside ChatGPT. Codex reads the same instructions and skills as Claude Code. I run it as a second agent next to Claude.
- [T3 Code](https://productradar.org/e/t3-code/) (Tools & Harnesses): An open-source control plane for coding agents, with desktop, web and mobile apps. I run agent sessions in the T3 Code desktop app and check their work in its preview.
- [Cloudflare Pages and Wrangler](https://productradar.org/e/cloudflare-pages/) (Infrastructure): Cloudflare's static and full-stack hosting, deployed with the Wrangler CLI. My personal site and this radar run on Cloudflare Pages.
- [MLX Whisper](https://productradar.org/e/mlx-whisper/) (Infrastructure): A Whisper speech recognition port that runs on Apple silicon with Apple's MLX framework. I transcribe video footage locally on my Mac with MLX Whisper.
- [Model Context Protocol](https://productradar.org/e/mcp/) (Infrastructure): An open protocol for connecting AI applications to tools and data sources. I connect a few apps through MCP, like T3 Code and a video editor. For email, calendar and Drive I use skills, because raw MCP tools failed too often.
- [Compilers as agent feedback](https://productradar.org/e/compiler-feedback/) (Practices): Choosing languages and checks that catch agent mistakes before a human reviews the work. I pick languages and checks that catch agent mistakes before I see them. This radar is written in Rust for that reason.
- [Cross-model review](https://productradar.org/e/cross-model-review/) (Practices): Having a different model, with a fresh context, review an agent's change before it merges. A second model with a fresh, read-only context reviews risky changes before they ship. The author's model never reviews its own risky work.
- [Mask personal data before a model sees it](https://productradar.org/e/privacy-masking/) (Practices): Replacing personal identifiers with placeholders before text reaches a language model. Kaufmann Health support mail is masked before agents read it, so patient identifiers stay out of model context.
- [Models select, code verifies](https://productradar.org/e/models-select-code-verifies/) (Practices): Letting a model choose from source material while code checks the choice against the source. On isaiconsciousyet.com a model picks a passage and code checks that the exact text is on the source page.

## Assess

- [TypeSafe Jev](https://productradar.org/e/typesafe-jev/) (Models & Agents): A System One model from TypeSafe AI that returns typed judgments and probabilities for code to act on. isaiconsciousyet.com uses Jev where code needs a typed answer from a model. One project is not enough for a verdict.
- [Railway](https://productradar.org/e/railway/) (Infrastructure): A cloud platform for deploying web apps, servers and databases. I have run side services on Railway, including my old Metabase. I have no firm verdict yet.

## Hold

- [Computer use on macOS](https://productradar.org/e/computer-use/) (Models & Agents): A Claude tool that lets a model operate a desktop through screenshots, mouse and keyboard. I stopped using it for Mac apps. The permission grants never took effect on my Mac, so agents use scripts instead.
- [Metabase](https://productradar.org/e/metabase/) (Tools & Harnesses): An open-source business intelligence and dashboard tool. I shut down my Metabase dashboards in September 2026 to save cost. Agents run the KPI queries directly.
- [Notion](https://productradar.org/e/notion/) (Tools & Harnesses): A workspace app for docs, wikis, databases and project boards. I moved my docs and task board from Notion to Obsidian in April 2026. Agents now look in one place.
- [Playwright MCP](https://productradar.org/e/playwright-mcp/) (Tools & Harnesses): An MCP server that lets agents drive a browser through Playwright. I removed it in September 2026 and moved my agents to agent-browser.
- [Shipping untested agent code](https://productradar.org/e/untested-agent-code/) (Practices): Merging code that an agent wrote without tests that check it. I do not ship agent code without tests. Each push runs a typecheck and lint, staging runs end-to-end tests, and a bug I report gets a failing test first.

## Recent changes

- AGENTS.md instruction files: New. One rule set for all my agents.
- Agent skills: New. Every repeated job has a skill.
- Ask-first gate: New. Part of my definition of done.
- Claude Code: New. My daily agent harness.
- Claude Fable: New. Main model in my daily sessions.
- Claude Haiku: New. In real use for small, narrow checks.
- Claude Opus: New. Default implementer in my agent builds.
- Claude Sonnet: New. Used daily for evidence, browser work and review.
- Cloudflare Pages and Wrangler: New. In real use for smaller sites.
- Codex: New. In real use as a second harness.
- Company brain: New. How my agents decide without me.
- Compilers as agent feedback: New. In use on this site and in typechecks.
- Computer use on macOS: New. Dropped: macOS permission grants never took effect.
- Cross-model review: New. Recent, used on risky changes.
- FluidVoice: New. My daily dictation tool.
- Gemini Flash: New. In real use in one publishing pipeline.
- GitHub Actions: New. CI for all my active repos.
- Guard hooks for agents: New. Every session runs behind guard hooks.
- Linear: New. Ticket source for my agent builds.
- MLX Whisper: New. In real use in one film project.
- Mask personal data before a model sees it: New. In use since October 2026.
- Metabase: New. Retired on 2026-09-25.
- Model Context Protocol: New. In real use for a few apps only.
- Model tier per task: New. Pinned on every agent by a hook.
- Models select, code verifies: New. In use in one publishing pipeline.
- Nano Banana Pro: New. In real use for occasional image work.
- Notion: New. Moved to Obsidian on 2026-04-20.
- Obsidian: New. Home of my company brain and task board.
- Parallel: New. Default research tool for my agents.
- Parallel agents in worktrees: New. How I build every ticket.
- Playwright MCP: New. Removed on 2026-09-27, replaced by agent-browser.
- Railway: New. Used for side services, no verdict yet.
- Resend: New. Production email for Kaufmann Health.
- Shipping untested agent code: New. Avoid: tests gate every commit.
- Stripe: New. Production payments for Kaufmann Health.
- Supabase: New. Production database for Kaufmann Health.
- T3 Code: New. In recent daily use, still alpha.
- TypeSafe Jev: New. Tried in one project.
- Vercel: New. Production hosting for two sites.
- agent-browser: New. Only browser tool since September 2026.