Insights

Your golden prompts have a median of 89 chars vs 71 for all prompts — 1.3× longer.

The prompts that unlock long autonomous runs are meaningfully longer than your baseline. Terseness isn't what's getting you leverage — specificity is.

median chars, 536 golden vs 7,392 total prompts

When you include "critique" in a prompt, correction rate drops 100% — 0.0% vs your 11.4% baseline.

This is a meta-instruction you should keep using. Other meta-instructions with strong effects are shown below.

n=11 prompts containing "critique"

Your correction rate climbs from 6.6% in the first 3 turns to 11.9% after turn 26 — drift is real.

You start sessions crisp but grow more corrective deeper in. Two plausible fixes: shorter sessions, or a mid-session reset ritual (summarize, re-plan, clear stale context).

Turn 1-3: 380 prompts • Turn 26+: 5,258 prompts

Prompts in the 800+ char range are 5.8× more likely to become golden than your shortest ones.

Golden rate by length: <40: 2.3% · 40-120: 9.5% · 120-300: 9.5% · 300-800: 7.1% · 800+: 13.5%

384 prompts in the winning bucket

code/remote-browser-sandbox has your cleanest convergence (94%); sandbox has your worst (83%).

Big gaps across projects usually mean something environmental: unfamiliar codebase, bad tooling, missing context docs, or fatigue. Worth asking what the clean projects have that the messy ones don't.

Best: 3 sessions · Worst: 11 sessions (min 8 user turns each)

Your convergence is holding steady — last 2 weeks averaged 89.5%, earlier weeks averaged 88.6%.

Convergence is the share of your user turns that didn't trigger a correction. Trend over time tells you whether your prompting is improving.

6 weeks of data

5 recent sessions started with no corrections at all, for 6+ user turns.

These are the opening prompts. Steal from yourself.

Perfect-convergence sessions with ≥6 user turns