A Practical Method: Build Verification Into the Loop

Agent output is a first draft, not a conclusion. In many cases, the fastest way to trust it is to attach clear acceptance criteria and have the agent produce evidence: lint results, tests, and targeted checks. This approach keeps iteration cheap, makes failures obvious, and preserves human accountability with artifacts you can review.

Habits:

• Add explicit acceptance criteria to agent prompts (inputs, expected behavior, and “done” conditions).
• Require the agent to run linters/formatters and report results (or explain constraints when it can’t).
• Make unit tests the default for logic changes; have the agent add or update tests alongside code.
• For behavior that spans components, use integration or end-to-end checks where useful (smoke tests beat vibes).
• Ask for a short verification summary: what was tested, what passed/failed, and what remains unverified.
• Keep cost bounds: cap iterations/compute, and escalate to human review when the loop isn’t converging.

Static sites are great until you need users to log in. Then you’re suddenly reading OAuth specs at 2am, questioning your life choices, and wondering if maybe you should just use a “real” backend after all.

Good news: you don’t need a server. You need about 50 lines of code and a Cloudflare Worker.

The Problem

I wanted my static site tools to authenticate with GitHub so they could read and write Gists. OAuth requires a server-side component to exchange authorization codes for tokens - the client secret can’t live in browser JavaScript (unless you enjoy getting hacked).

Continue reading...

I built a skill and a skill marketplace today. It only took 12 iterations, three test projects, and one existential crisis about what “markdown format” actually means.

It Started With a Problem (As It Always Does)

I have this product at work. It’s fine. It works. But it was built before anyone thought AI agents would be poking around in our codebases, so its observability is… not great. Imagine trying to debug something while blindfolded. That’s what agents experience when they work with it.

Continue reading...

A Practical Method: Invest in the Agentic Layer

In many cases, the fastest way to ship reliably with AI is to improve the layer around your systems that agents use: prompts, tool interfaces, guardrails, and verification. Instead of treating each AI interaction as a one-off, aim to make common work repeatable and measurable. When appropriate, spend a meaningful share of engineering time strengthening this layer so the next feature takes less effort and carries less risk.

Continue reading...

A Practical Method: Keep Agents Narrow, Make Results Verifiable

When an agent is asked to do everything, it usually does nothing particularly well. A small, purpose-built agent with a focused prompt can produce a cleaner first draft, with clearer assumptions and fewer hidden side-quests. In many cases, splitting work by purpose makes it easier to apply constraints, verify results, and hold the system accountable. Use this when the workflow has distinct steps with different inputs, tools, or risk.

Continue reading...