Agentic Delegation for Code Verification

Agentic Delegation for Code Verification | Generated by AI

Home 2026.05

Question: Is the prompt “try it, run it, kill after 1 min, adjust params if not suitable, then tell me it’s OK to run for hours” a great Claude Code prompt pattern? And can you explain why this workflow is so effective?

Answer:

Yes — That Is an Excellent Claude Code Prompt

Your prompt is a near-perfect example of agentic delegation. Here’s why it works so well:

Why This Prompt Pattern Is Great

1. It gives Claude a complete mini-workflow, not just a task

You’re not asking “what are the params?” — you’re saying:

“Do the experiment, judge the result, fix the problems, hand me the final answer.”

Claude Code handles the entire feedback loop:

Run → observe → diagnose → patch → confirm

2. It uses time-boxing as a sanity check

“Kill after 1 minute” is genius. It means:

No wasted GPU hours on a broken config
Claude sees real stdout/stderr (OOM errors, CUDA errors, loss=NaN, shape mismatches)
1 minute is enough to confirm the training loop is healthy

3. It ends with a clear human handoff point

“Tell me it’s OK — let me run in another terminal for hours”

This is the perfect division of labor:

Claude Code: trial run, debugging, param validation
You: execute the final blessed command in tmux

This matches exactly what the Claude Code team calls verification-first workflow — Claude performs dramatically better when it can verify its own work, like running tests and validating outputs, before handing results back to you.

The General Pattern (Reusable Template)

This prompt works for almost any long-running ML job:

"Try running [command], kill it after [short time], 
check for errors/warnings/bad params, 
adjust anything that looks wrong, 
then give me the final production command to run for real."

Apply it to:

torchrun distributed training
python tokenize.py on large datasets
pytest on a big test suite before CI
ffmpeg batch jobs
Any nohup-worthy process

What Claude Code Does During That 1 Minute

When Claude Code runs and kills your training job, it’s checking:

Signal	What It Means
CUDA OOM immediately	`batch_size` too large → halve it
`loss=NaN` at step 1	LR too high or bad data
Slow step time	Need `compile=True` or smaller model
Clean step logs	✅ Config is good
Shape mismatch error	Config/model mismatch → fix dims

Your Prompt Improved (Even Sharper Version)

Run a 1-minute smoke test of the training command. 
Watch stdout for OOM errors, NaN loss, CUDA errors, or 
abnormally slow step times. Kill it after 1 min. 
Fix any param issues you find. 
At the end, give me one clean final command I can paste 
into a tmux session to run overnight.

This version makes Claude’s checklist explicit, so it knows exactly what “looks OK” means.

Summary

Your instinct is exactly right and aligns with professional ML workflow best practices:

Claude Code (1 min smoke test)  →  You (hours-long tmux run)
     diagnosis & fixing               production execution

The Claude Code team calls verification the single most impactful practice — giving Claude a way to check its own output before handing back to you. Your “run then kill” pattern is a hand-crafted version of exactly that.

References:

Back Donate