Stephen J. Hayes AI coding research • Claude Code • LLM workflows • build notes

Notes on AI-assisted software engineering.

I test the newest AI coding workflows (Claude Code, copilots, agent patterns) and write down what actually improves speed without sacrificing correctness. This is a simple, fast, text-first site: fewer announcements, more receipts.

Dublin • hello@stephenjhayes.ie • GitHub (add) • LinkedIn (add) • RSS (add)

Latest

short posts • change logs • links

2026-01-xx
Claude Code workflow: “small PRs, high confidence” My default loop: scope → constrain → diff → tests → review. What I changed after 30+ runs.
2026-01-xx
A practical eval harness for AI-assisted refactors How I keep reliability: golden tests, regression checks, and “stop conditions” for agents.
2026-01-xx
Tool note: when Cursor beats Claude Code (and vice versa) Where each shines: repo navigation, patch hygiene, long context, and terminal loops.
2026-01-xx
Link: A great write-up on coding agents (add link) My takeaway: the “fast path covers 99%” principle applies to agent design too. (Short commentary.)

Research log

what I’m testing right now

Claude Code patterns

Making outputs reviewable and safe in real repos.

  • “Patch hygiene”: tiny diffs, clean commits, explicit file scope
  • Context strategy: repo map, pinned constraints, rolling summaries
  • Verification: tests first, static checks, “prove it” prompts

Agent reliability

Where agents help vs. where they cause damage.

  • Routing: classify → choose toolchain → execute
  • Guardrails: permissions, allow-lists, audit trails
  • Evals: golden tasks + regression suite + cost/latency budgets

LLM “workbench”

A repeatable environment for comparing tools and models.

  • Same tasks, same repo, fixed constraints
  • Track: time-to-first-solution, bug rate, PR review time
  • Notes: what breaks, what scales, what’s hype

Personal rule

Borrowed from the best blogger-engineers: write as you go.

  • Capture prompts + diffs + outcomes
  • Post the useful bits, not the marketing
  • Prefer “here’s what happened” over “here’s my opinion”

Projects

things I’m building

ongoing
AI migration accelerators Automation + coding agents to reduce repetitive migration work. (Writeups + tools soon.)
ongoing
Prompt + eval library A small set of “default” scaffolds for reliable AI-assisted changes.

About

why this exists

I’m a software leader in P&C insurance technology, obsessed with practical automation. I use AI daily and I’m documenting what works: Claude Code runs, tool comparisons, repeatable agent workflows, and how to adopt this stuff safely in teams.

(Golf + Ireland posts may show up later, but this front page stays AI-focused.)

Subscribe

optional

If you want updates, I can add a simple newsletter signup here (Buttondown/Beehiiv/Mailchimp), plus an RSS feed when the post structure is in place.

For now: email me if you want a heads-up when new notes drop.