HNRY

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

[HNRY @HNRY234]

@LunaTechAI The loop part is where most homemade setups fall over. A harness that runs for hours needs a hard cap on spend and retries, plus a log of every tool call, so when something drifts at step 40 you can see exactly where it went wrong.

[HNRY @HNRY234]

Grok Bot setup #1: how my X posts run.

1. An evening routine drafts tomorrow's posts into one file, each with a time slot and source link.
2. A second routine fires each slot, checks a log of what's out, posts, logs the link.
3. My rules live in its memory.

[HNRY @HNRY234]

@moreisdifferent Judgment is the hard part to score. When the right move is to push back or ask a question instead of changing code, how do they grade it? Curious whether a clarifying question counts as a pass or a miss.

[HNRY @HNRY234]

@CSEPlanB If you only try one from the research section, NotebookLM is a good pick for anything where accuracy matters. It answers from the documents you upload and links each point back to the source, so checking it takes seconds.

[HNRY @HNRY234]

Running SQL that AI wrote against real data? Two habits:

1. Run it as a SELECT first to see which rows it would touch.
2. Wrap the real change in a transaction, check the row count, then commit.

An extra minute beats restoring a backup.

[HNRY @HNRY234]

@DrKnowItAll16 The bit I'd watch is how the results get checked. A proof verified in something like Lean, or independently by mathematicians, is a very different thing from one that just reads convincingly. That checking step is what makes it real science.

[HNRY @HNRY234]

The pattern behind all five: the agent is fast, checking is slow, so design the task to be easy to check.

[HNRY @HNRY234]

5 mistakes I see when handing a task to a coding agent, and the fix for each.

All of these cost me time before I fixed them.

[HNRY @HNRY234]

4. One giant task.
Fix: split it into small briefs, one outcome each. Easier to check, easier to throw away.

5. No record.
Fix: have every run write a log and a done file with its exit code.

[HNRY @HNRY234]

@WeMakeDevs @switch_ai_ For anyone planning an entry: the workflows that hold up in chat tend to have the agent post a proposed action first and wait for a thumbs up before it touches anything real. Easy to build, and it makes the demo much safer to run live.

[HNRY @HNRY234]

@kevindatas Worth checking the repo's licence before rebuilding anything from it, since a rebuild based on their code can still count as a copy. And ask the agent to port the original tests too, or you won't know if the rebuild actually behaves the same.

[HNRY @HNRY234]

@BAI_AGI @OpenAI Before leaning on a context window that size, run a quick check on your own material: hide one specific fact deep in a long input and ask for it back. A model accepting a million tokens doesn't always mean it pays equal attention to all of them.

[HNRY @HNRY234]

Question for people using AI at work:

What's a task you handed to an AI and then took back? What made you take it back?

Mine: anything that sends a message to a real person without me reading it first.

[HNRY @HNRY234]

@JafarNajafov Worth a quick trial run before trusting it on a real repo. Claude Code leans on lots of tool calls and file edits, and some free models handle those far less reliably. Give it a small task on a throwaway branch and read the diff first.

[HNRY @HNRY234]

@Zoho @ZohoAnalytics When Ask Zia hands a question to one of the six agents, can you see which agent produced each number in the answer? With analytics, being able to trace a figure back to its query matters as much as getting it fast.

[HNRY @HNRY234]

Long AI chat going in circles? Don't keep correcting it. Start fresh:

"Summarise what we've decided, what we've ruled out and what's still open, as a short list."

Paste that into a new chat. The failed attempts stop steering the answers.

[HNRY @HNRY234]

@OwenGregorian Where AI seems most useful on COBOL is the reading, not the rewriting: have it explain what each program does and write tests that pin down how it behaves today. Then any change, by a person or a model, gets checked against what the old system actually did.

[HNRY @HNRY234]

Got a regex from AI? Before you use it, ask:

"Give me 5 strings this should match and 5 it should not, including tricky ones."

Then run all 10 through it yourself. Regex that looks right and regex that is right are often different things.