@PrajwalTomar_: This Reddit post just dropped the Claude Code skill that takes you from 70% to 90% accurate on the first try. A builder…

X AI KOLs Following Tools

Summary

A Reddit post introduces a Claude Code skill called /grill-me that extracts all context from users by asking iterative questions and saving decisions to a knowledge doc, improving initial accuracy from 70% to 90%.

This Reddit post just dropped the Claude Code skill that takes you from 70% to 90% accurate on the first try. A builder broke down how to use Matt Pocock's /grill-me skill to extract every piece of context out of your head before you start prompting. The loop: → It asks ONE question → You answer → It writes your answer into a knowledge doc → It keeps going until there are no gaps left Some sessions take 5 questions. Some run past an hour. The long ones are the ones worth having. His modification is the actual unlock. Checkpointing after every question. It creates a `brainstorms` folder at your project root and drops a markdown file with every decision, the full Q&A log, and the highlights. Stops the context window from dropping an answer you gave 40 minutes earlier. The math: → Old way lands at 70% on iteration one. You grind 10-30 passes to cap at 95%. → Run /grill-me first and iteration one starts at 90%. You refine instead of clawing up from the bottom. Most builders are still prompting their agents from scratch every session. The smart ones are extracting every business decision into a knowledge doc first, then letting Claude Code build on top of it.
Original Article
View Cached Full Text

Cached at: 06/09/26, 01:30 AM

This Reddit post just dropped the Claude Code skill that takes you from 70% to 90% accurate on the first try.

A builder broke down how to use Matt Pocock’s /grill-me skill to extract every piece of context out of your head before you start prompting.

The loop:

→ It asks ONE question → You answer → It writes your answer into a knowledge doc → It keeps going until there are no gaps left

Some sessions take 5 questions. Some run past an hour. The long ones are the ones worth having.

His modification is the actual unlock. Checkpointing after every question. It creates a brainstorms folder at your project root and drops a markdown file with every decision, the full Q&A log, and the highlights. Stops the context window from dropping an answer you gave 40 minutes earlier.

The math:

→ Old way lands at 70% on iteration one. You grind 10-30 passes to cap at 95%. → Run /grill-me first and iteration one starts at 90%. You refine instead of clawing up from the bottom.

Most builders are still prompting their agents from scratch every session.

The smart ones are extracting every business decision into a knowledge doc first, then letting Claude Code build on top of it.

Similar Articles