HNRY

Building intelligent systems and ambitious ideas.
AI • maths • software • systems
Learning, building, documenting.

[HNRY @HNRY234]

@ambermac The flip side matters for anyone shipping software: if a coding agent can help find holes in a console, it can help find them in your app too. Worth pointing one at your own code with a "find ways to break this" prompt before someone else does.

[HNRY @HNRY234]

@KirkDBorne One Claude habit that pays off whatever book you learn from: before any code, ask it to restate the task and list what it would check to call it done. Misreadings show up in that restatement, which is much cheaper than finding them in the output.

[HNRY @HNRY234]

Quick check worth doing on any business website:

Fill in your own contact form as if you were a customer. Then check the message actually arrived, who it went to, and that it didn't land in spam.

Forms break quietly. Nobody complains about an enquiry they never knew was lost.

[HNRY @HNRY234]

@Tony_Stef_ For a fair comparison, give all three the same small task plus a check for done, like a test that must pass. And if they share one folder, a separate git branch or worktree each stops them overwriting each other's edits mid task.

[HNRY @HNRY234]

@MotivationalMan The delegation point holds. What fixed it for me was writing the brief like a handover note: what done looks like, what it must not touch, and how it should check its own work before saying finished. Vague briefs get vague work from people and agents alike.

[HNRY @HNRY234]

The most useful AI habit I have is reading the output before I trust it. Not glamorous. Saves the most time.

[HNRY @HNRY234]

@DAIEvolutionHub Agent to loop is the jump I'd slow down on. A loop needs two things the agent step doesn't: a check it can run on its own output, and a clear stop condition. Without those it just retries the same mistake and burns tokens.

[HNRY @HNRY234]

@saranormous Feels like the gap is undo more than model quality. Coding agents took off partly because repos have tests and git to roll back a bad change. Most enterprise workflows have neither, so the jump likely comes when agents get a sandbox copy of the process and a clean undo.

[HNRY @HNRY234]

If enquiries keep slipping through the cracks, try this before buying any new tool:

1. Add a "Last contacted" date column to your leads sheet
2. Make a filter view: status is open, last contacted more than 3 days ago
3. Open that view first thing every morning

It turns "I'll remember to reply" into a list of who is waiting on you.

[HNRY @HNRY234]

@Mho_23 The part I'd love to hear more about is QA. At hundreds a day, how do you catch the bad cuts: a second model checking frames and captions against the brief, or a person spot checking a sample? That review step seems to be where the editors' time goes now.

[HNRY @HNRY234]

@trevin The test I'd use for any agent: can you export its memory, instructions and tool setup as plain files today, without a support ticket? If not, keep your own copy in markdown or a git repo as you go, so switching costs a weekend instead of your history.

[HNRY @HNRY234]

Before a long AI task, ask for two things first:

"Restate the goal in one sentence, and list what you will not do."

If the restated goal is off, you've lost ten seconds instead of an hour.

[HNRY @HNRY234]

@TeksEdge A cheap check before trusting any judge: run each pair twice with the order swapped and a reworded prompt. If the verdict flips, that pair is noise, not signal. Reporting how often it flips next to the score tells readers how much the ranking can be trusted.

[HNRY @HNRY234]

@podcastnotes The plan step is where English stops being enough. If the model can't say how it will test the thing, the build loop just iterates on guesses. Asking for the acceptance checks up front, in plain words, is what keeps describe the app from turning into debug the app.

[HNRY @HNRY234]

Small fix that saves hours of cleanup later: if a spreadsheet column only has a few valid answers, make it a dropdown.

In Google Sheets: select the column, then Data, Data validation, and add a dropdown rule.

No more "Done", "done " and "DONE" counting as three different statuses.

[HNRY @HNRY234]

@taro_f I'd want to see revert rate next to that 74%. Share of merged PRs can rise just because the agent opens lots of small ones, so how often its PRs need a follow-up fix or get rolled back says more about how much of the work it's really taking on.

[HNRY @HNRY234]

Grok Bot skills: when a multi-step task keeps coming up, save the steps once as a skill and any of your bots can run it by name.

One of mine watches long Codex builds on my PC and flags when one stalls or finishes.

Keep skills generic. Put the specifics in the ask.

[HNRY @HNRY234]

@gregisenberg Pricing the agent is where it gets hard. Charge per task and customers start arguing over what counts as done. Charging per outcome they already track, like a booked call or a closed ticket, makes the bill much easier to trust.

[HNRY @HNRY234]

@codyschneider I'd watch the auto negatives most. An LLM judging intent from the search term alone can block a phrase that looks vague but converts. Logging every negative it adds and rechecking them against conversions each week catches that before it quietly shrinks reach.

[HNRY @HNRY234]

@shcallaway Removing the dropped ones from Slack is only half of it. Their GitHub app installs, OAuth grants and any API tokens or deploy keys they were given tend to outlive the Slack integration, so worth revoking those too and checking nothing is still calling in with them.