aisettings
Log InAdd your setup

Andrej Karpathy

Last updated: Aug 2, 2026
RoleAI researcher and educator
Sources13 cited fields · high confidence
@karpathykarpathy.ai

Stitches Cursor completion, Claude Code or Codex, and GPT-5 Pro into a layered workflow, while keeping an IDE nearby for review.

“This is easily the biggest change to my basic coding workflow in ~2 decades of programming and it happened over the course of a few weeks.” x.com

Share cardShare your setupCompare

Loadout

aisettingsShare on X

Andrej Karpathy’s loadout

Harness
Claude Code
Claude Code or Codex on the side of Cursor for larger, prompt-specifiable chunks; never YOLO mode.
Model
Claude Opus 5
Used Opus 5 for a roughly two-hour, 1M-token Lord of the Rings Three.js rendering experiment.
Also uses
Claude Fable 5
Praised for difficult long problem-solving sessions and ambitious tasks.
EditorCursor
TerminalGhostty
ToolsBrowser MCP

5 views · 0 equipped

Models

Claude Opus 5
Used Opus 5 for a roughly two-hour, 1M-token Lord of the Rings Three.js rendering experiment.
x.com, 2026-08 ↗
Claude Fable 5
Praised for difficult long problem-solving sessions and ambitious tasks.
x.com, 2026-06 ↗
Cursor tab completion
About 75% of his August 2025 LLM assistance was Cursor tab completion.
x.com, 2025-08 ↗
GPT-5 Pro
Final layer of defense for the hardest bugs and research-heavy tasks.
x.com, 2025-08 ↗

Editor & terminal

Cursor
Uses Cursor for tab completion, inline edits, and as the IDE beside agent sessions.
x.com, 2026-01 ↗
Ghostty
Runs a few Claude Code sessions in Ghostty windows or tabs beside the IDE.
x.com, 2026-01 ↗

Tools

Browser MCP
Uses browser MCP in a success-criteria loop for tasks that benefit from repeated browser verification.
x.com, 2026-01 ↗

Workflow

  1. 1Use a long /voice stream-of-consciousness ramble or turn it into a short interview so the model reconstructs the goal.x.com, 2026-07
  2. 2Stitch several workflows instead of searching for one perfect tool: completion, inline edits, agents, then a hard-problem model.x.com, 2025-08
  3. 3Avoid YOLO mode; run one agent instance and stop it when it goes off track.x.com, 2025-08
  4. 4Give agents success criteria and tests, then let them loop until the criteria are met.x.com, 2026-01
  5. 5Keep a large IDE beside agents for manual edits and review because conceptual mistakes remain common.x.com, 2026-01

Timeline

  1. 2026-08-02ModelClaude Opus 5x.com, 2026-08
  2. 2026-06-09Also usesClaude Fable 5x.com, 2026-06
  3. 2026-01-26EditorCursorx.com, 2026-01
  4. 2026-01-26TerminalGhosttyx.com, 2026-01
  5. 2026-01-26ToolsBrowser MCPx.com, 2026-01
  6. 2025-08-24HarnessClaude Codex.com, 2025-08
  7. 2025-08-24Also usesCursor tab completionx.com, 2025-08
  8. 2025-08-24Also usesGPT-5 Prox.com, 2025-08

Comments

Log in to comment