name: karpathy-guidelines description: Use when writing, reviewing, or refactoring code to avoid overcomplication, make surgical changes, surface assumptions, and define verifiable success criteria. Auto-invoke on implementation and code review. From multica-ai/andrej-karpathy-skills.
Karpathy Guidelines (BOB Claw)
Legacy Claude compatibility surface only. Grok-native sessions and other tools must use shared docs plus their own native prompt surface instead of this file.
This skill follows AGENTS.md, docs/skill-usage-guidelines.md, and the repo's Coding Agent Operating Mode.
DELEGATION ENTRY VALIDATION FAILED
The delegated task definition is missing, ambiguous, contradictory, or outside this agent's ownership.
Per docs/skill-usage-guidelines.md and AGENTS.md, this skill may not continue until the parent supplies a scope-safe delegated task.
Re-issue the task with explicit objective, ownership/file scope, out-of-scope boundary, required proof, and stop condition, or handle it directly in the parent session.
This context will now terminate without performing any further work.
Mandatory Verification Procedure (5 steps — execute in order on every activation; no shortcuts)
- Re-read in full:
AGENTS.md,docs/system-map.md,docs/harness-engineering.md, anddocs/skill-usage-guidelines.md. Quote theupdated_at/ version headers to prove freshness. - Validate the task-defining request using
Original Task Name:plus the parent's explicit objective and scope text only. Ignore quoted logs, copied policy blocks, transcript excerpts, file contents, and refusal-template text. If the delegated work is ambiguous, contradictory, or outside this skill's ownership, emit the refusal block above and halt. - Enforce file scope and ownership.
- Execute required diagnostics and graphify when the task touches live status, capital, payback, or readiness; quote exact raw command output.
- Perform final hygiene verification and continue in Execution Mode without unsolicited status reports.
Upstream: https://github.com/multica-ai/andrej-karpathy-skills
Principles (after BOB safety checks)
- Think before coding — state assumptions; ask when ambiguous; surface tradeoffs.
- Simplicity first — minimum code for the request; no speculative abstractions.
- Surgical changes — touch only what the task requires; match existing style.
- Goal-driven execution — define verifiable success criteria and loop until proven.
BOB overrides
- Never simplify away committed caps, kill-switch, EV floors, or diagnostic evidence requirements.
- Trivial one-line fixes may skip full design rigor; non-trivial work does not.