ワンクリックで
karpathy-autoresearch
|-
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
メニュー
|-
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | karpathy-autoresearch |
| description | |- |
Andrej Karpathy's way of working, distilled into four principles for LLM coding agents. (1) Think Before Coding — state assumptions, surface ambiguity, push back when warranted. (2) Simplicity First — minimum code, no speculative abstractions, DRY+NRtW. (3) Surgical Changes — touch only what you must, trace every line to the request. (4) Goal-Driven TDD — failing test first, loop until verified, never claim solved without testing.
Applied in b00t: soul-first (learn before implement), lfmf for lessons, autolearn loop for goal-driven knowledge loading, and OODA for autonomous execution.
Apply Karpathy's principles in every coding session. Think before implementing, keep changes minimal and surgical, and always write the failing test first.
Canonical b00t reviewer capability — adversarial multi-framework code/datum review. Shared across all harnesses (Claude Code, opencode, Hermes, b00t-cli). Loaded via: b00t learn reviewer, or by harness-specific role supplements.
Remove AI-generated artifacts from code. Three-phase certainty-graded cleanup. Use after any AI implementation session or before PR creation.
8-phase agile workflow — strategy→ideate→brainstorm→plan→work→review→compound→pulse. State machine with FOL-guarded transitions, gh-issues backlog, executable just harness. Inspired by everyinc/compound-engineering-plugin.
Australian crypto tax treatment per ATO guidance QC 53725. Crypto assets are CGT assets under ITAA 1997 s 108-5. Each disposal triggers CGT event A1. Covers personal-use exception, trading stock test, and the 50% CGT discount.
Australian R&D Tax Incentive (RDTI) — ITAA 1997 Division 355. Covers eligibility criteria for core and supporting R&D activities, registration with AusIndustry/IP Australia, expenditure categories, offset rates, and the company size threshold test.
Evidence graph construction for tax audit trails. Each Satisfies check emits EvidenceNode structs with Blake3 content hashes. Evidence chains are tamper-evident and link legislative citations to factual findings.