메뉴
Auto-improve any agent prompt using Karpathy's autoresearch method with iterative test-measure-improve loops
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Use when implementing features, fixing bugs, refactoring code, optimizing performance, resolving deprecations, or any multi-step development task requiring planning, implementation, testing, and review. Orchestrates specialized agent teammates through research, architecture, coding, QA, code review, adversarial review, and documentation phases. Triggers on: "implement", "build", "fix bug", "refactor", "add feature", "develop this", "help me build", "add functionality", "optimize performance", "resolve deprecation", "systematic development". Do NOT trigger on: simple questions ("what does this code do?"), file searches ("where is the auth function?"), one-off commands ("run the tests"), code explanations, quick edits, or non-development tasks.
Multi-lens adversarial review with Skeptic, Architect, and Minimalist lenses and 7+ attack vectors
Safety guardrail that blocks destructive commands for the current session
Core development rules and philosophy for coding standards, git practices, and quality guidelines
Directory freeze guardrail that restricts file edits to a specific directory for the current session
Comprehensive security review skill for authentication, user input, secrets, API endpoints, and OWASP Top 10
| name | autoresearch |
| description | Auto-improve any agent prompt using Karpathy's autoresearch method with iterative test-measure-improve loops |
| author | Jennings Liu |
| version | 1.0.0 |
| license | MIT |
Based on Andrej Karpathy's autoresearch method. Instead of manually improving agent prompts, run an iterative loop: try a small change, score the result, keep improvements, revert regressions. Systematically increase agent output quality.
Triggers on: "autoresearch", "auto-improve", "optimize agent", "tune prompt", "improve skill quality"
Run agent on test input. Score output against checklist (yes/no). Record baseline score (average of 3 runs). Analyze weakest checklist items. Make ONE small change to agent prompt. Re-run agent on same test input (3 runs). Score again. If improved → KEEP change. If dropped → REVERT. Repeat until 95%+ score 3 times in a row OR max rounds reached (default: 10). User specifies agent. Read from `agents/.md`. 3-6 yes/no questions. Each checks one specific aspect. Good: "Does the code review identify at least one production-risk bug?" Bad: "Rate the quality 1-10" (subjective). Run agent on test input 3 times. Score each. Average = baseline. For each round: analyze failing items → hypothesize one change → apply (save backup) → test 3 runs → score → KEEP if improved, REVERT if same/worse → log round. Produce summary (baseline/final score, rounds, changes kept/reverted) and changelog per round (score change, what changed, why, effect). Add specific gotcha, add worked example, add banned-patterns list, add rule targeting weakest item, restructure output template Rewriting entire prompt, adding vague instructions, adding 5 rules at once, removing sections without testing Using only one test input optimizes for that case. Use 2-3 diverse inputs. More than 6 items may cause gaming individual checks at expense of quality. Short-term score drops may enable bigger gains. Try "creative rounds" for structural changes. Results stored in `${CLAUDE_PLUGIN_DATA}/autoresearch/`: `[agent-name]-results.json` and `[agent-name]-changelog.md`.