Software engineer. Songwriter. Worship leader. Building software, writing songs, leading worship, and living it all for Jesus Christ, the One True King. Serving
AI wrote the code. Great. Who checked it?
I built BORK: 8 AI skills for GitHub issues, draft PRs, reviews, fixes, and E2E/browser checks.
Because “looks good to me” needs evidence. Especially from a robot.
https://t.co/bbtfhCFJhr
@_kekule15 Check the tests against the requirements, then try to break the assumptions both code and tests share. Include permissions, failure paths, and existing behavior. The same misunderstanding can produce two perfectly agreeing files. Green is a color, not a guarantee.
@oldstackjournal Then test the old URLs against the new site, including the final destination and redirect chains. A migration can look perfect on the homepage while quietly breaking ten years of links. Screenshots are terrible at catching 404s.
@MrDaveFoy "For this particular project" is doing good work here. My first CMS question is who will maintain the content six months from now. The developer enjoying setup and the editor enjoying Tuesday are two different requirements.
@GosuCoder PR count is a brave metric when the bottleneck is understanding the diff. I’d cap work in progress and require the author to explain scope, risk, and test evidence. If nobody can explain the change, it is not ready for somebody else to approve.
@isaac_dev0 The exit condition deserves its own spot on that list. What counts as done? What happens after three failed attempts? A while loop without a sensible stop rule is how a small bug gets its own renovation budget.
@AIVerseUday Context, especially the constraints people forget to write down. "Fix checkout" and "fix checkout without changing renewal behavior" are very different jobs. The model cannot read the requirements still living in your head. Yet.
@simonhamp The offline-first part caught my eye. How do you handle a client reconnecting after the server has changed the UI schema? That version mismatch is where the interesting engineering tends to hide. Usually until Friday afternoon.
@georgehanu I’d trial it on a familiar bug first. Same task, same repo, then compare correctness, unnecessary changes, and review time. Fewer output tokens are nice. Fewer "why did you touch that file?" moments would sell me.
@ascorbic A feature can be a small change in the editor and a very large compatibility question for plugins. "Who shipped first?" misses the interesting part: what assumptions did each project have to preserve?
@Nikita26rai First, revoke or rotate the credentials. Then clean the history and coordinate with everyone using that branch. A tidy Git log does not make a leaked key secret again. The order matters more than the command.
@Gegam245074 One habit worth trying: write your diagnosis before asking the agent for its fix. Then compare the reasoning. It keeps you involved in the hard part instead of becoming the person who clicks "looks good" for a living.
@jaydrogers New PHP release ritual: read the migration guide, run the test suite, then discover which dependency has been living rent-free in 2018. Staging earns its keep on upgrade day.
@arthuryuzbashev A useful addition would be a source link and last-checked date for each rule. An AI audit is only as current as its checklist. Otherwise we just automated being confidently rejected.
@KatieKeithBarn2 The plugin can solve a very specific problem while the product page says "powerful and flexible." Show the exact problem, the fix, and a working example. Nobody searches for "powerful and flexible" when checkout is broken.
@hany99dev I want a benchmark where the model fixes a legacy bug without rewriting half the app. Bonus points if the diff fits on one screen. That leaderboard would help me choose.
If an agent keeps fixing the wrong thing, give it one real input, the output you got, and the output you expected. Then ask it to explain the gap before editing.
That example often does more work than another paragraph of instructions.