essays and notes

Writing

Mostly about agents, product decisions, and the parts of developer tools that quietly annoy me. I try to publish only when I’ve actually changed my mind about something.

by theme

Three things I keep coming back to.

Agent experience

What an agent is allowed to do shapes everything else about the product.

Credentials, permissions, and the ways agents fail.

Open the thread

Product judgment

Most insights are trivia. The useful ones change what you do on Monday.

Evidence, tradeoffs, and deciding without enough information.

Open the thread

Platforms and developers

You can predict what a platform will neglect from how it earns.

Developer momentum, ecosystems, and where the gaps stay open.

Open the thread

everything

All of it

13 posts, oldest at the bottom.

  1. 01the agent surface map: pick the cell before the promptChat-plus-approve works for a coding pair. It breaks a hosted operator and a custom actor. Pick the cell, then pick the credential model.
  2. 02the one line 95% of agent skills are missingAn agent uses a skill's description to decide whether to load it. I scanned 200 public skills and found that 163 of 171 valid files weaken that signal.
  3. 03you can't charm an agentAgents have no goodwill buffer or patience for a vague error. Their success rate gives product teams a direct measure of experience quality.
  4. 04your 404 page has a blind spotAn agent that follows a stale or hallucinated URL reaches a dead end. Agent 404 indexes live pages and returns ranked replacements.
  5. 05is your 'insight' trivia?Correlation can point you toward a problem. You earn the word "insight" after you find the cause and use it to choose a path.
  6. 06claude code's DX is too good. and that's a problem.Claude Code nails developer experience in ways most people can't articulate. But I'm starting to see a familiar pattern emerge - the same one that's killed great devtools before. Here's what I'm watching for.
  7. 07vercel vs cloudflare: two philosophies of building for developersTwo platforms, two theories of the web: Vercel optimizes for the developer at build time, Cloudflare for the request at runtime. I compared their pricing, edge runtimes, and platform boundaries to see where each one wins.
  8. 08why you shouldn't outsource your thinking to AIUse AI to gather context and accelerate the work. Form your own view before the model gives you a polished answer to agree with.
  9. 09building at the edgesA strong opening in a mature market sits beside the category leader: one painful workflow that its scale, incentives, or architecture neglects.
  10. 10borrowed happinessA promotion, a new car, or someone else's perfect trip can reset your idea of success. Borrowed goals keep moving the point where happiness begins.
  11. 11product strategy trapPMs can spend days making a strategy look convincing before they have learned enough to make the underlying decisions.
  12. 12the MCP anti-patternConverting an OpenAPI spec into tools gives an agent more endpoints to juggle. Start with the user's job and design the tool around that workflow.
  13. 13f1 strategists can teach you to make better product decisionsA pitwall cannot wait for complete information. It plans plausible scenarios, defines the signal for each move, and reviews the call after the race.

by email

Want an email when I post?

One email per post. No schedule, no roundups.

That’s it — one email when something new is up.