{"_id":"@arkone_ai/autoresearch","name":"@arkone_ai/autoresearch","dist-tags":{"latest":"1.0.0"},"versions":{"1.0.0":{"name":"@arkone_ai/autoresearch","version":"1.0.0","description":"Claude Code skill: apply the autoresearch methodology to any LLM application or model training project","type":"module","scripts":{"postinstall":"node install/install.js"},"bin":{"autoresearch":"install/install.js"},"keywords":["claude","claude-code","claude-skill","llm","ai","autoresearch","prompt-engineering"],"license":"MIT","engines":{"node":">=18"},"_id":"@arkone_ai/autoresearch@1.0.0","gitHead":"c2864d7023bd90b810bff0e6dde7f227a5c0710f","_nodeVersion":"22.22.1","_npmVersion":"10.9.4","dist":{"integrity":"sha512-MjXJINsI3f/s1uXbC0usohTmPAJpHqhEeI9oXgB2+ygYB6RO1B7e/MFyDRdLM0Ao0Cqb0JNMEOiZXDqtRu9xqg==","shasum":"bc1060f96c4687cb3516f667cf9dbdc481a48502","tarball":"https://registry.npmjs.org/@arkone_ai/autoresearch/-/autoresearch-1.0.0.tgz","fileCount":7,"unpackedSize":18438,"signatures":[{"keyid":"SHA256:DhQ8wR5APBvFHLF/+Tc+AYvPOdTpcIDqOhxsBHRwC7U","sig":"MEQCIQCiwnVn0btW1im13KzgnETQVPvtwuEDc1UB8vuIsQr5nAIfft5B7BWVAL+KcHNzVLg7MegWTv46oNi33sXnS8DSEw=="}]},"_npmUser":{"name":"sgthomas","email":"sobingt@bitbrothers.in"},"directories":{},"maintainers":[{"name":"sgthomas","email":"sobingt@bitbrothers.in"}],"_npmOperationalInternal":{"host":"s3://npm-registry-packages-npm-production","tmp":"tmp/autoresearch_1.0.0_1774262769323_0.22784108161164873"},"_hasShrinkwrap":false}},"time":{"created":"2026-03-23T10:46:09.201Z","1.0.0":"2026-03-23T10:46:09.449Z","modified":"2026-03-23T10:46:09.761Z"},"maintainers":[{"name":"sgthomas","email":"sobingt@bitbrothers.in"}],"description":"Claude Code skill: apply the autoresearch methodology to any LLM application or model training project","keywords":["claude","claude-code","claude-skill","llm","ai","autoresearch","prompt-engineering"],"license":"MIT","readme":"# claude-skill-autoresearch\n\nA [Claude Code](https://claude.ai/code) skill that applies the [autoresearch](https://github.com/karpathy/autoresearch) methodology to any LLM-powered project — API routes, prompt iteration, agents, evaluation pipelines, or full model training.\n\n## Install\n\n```bash\nnpm install -g claude-skill-autoresearch\n```\n\nThe skill is automatically copied to `~/.claude/skills/autoresearch.md` on install.\n\n## Usage\n\nIn any Claude Code session:\n\n```\n/autoresearch\n```\n\nClaude will apply the autoresearch discipline to whatever AI work you're doing.\n\n## What it does\n\nThe autoresearch methodology was originally designed for autonomous LLM training experiments (run overnight, keep/discard based on a single metric). This skill generalizes those principles to any AI work.\n\n**The universal structure** — every AI task maps to the same three parts:\n\n| Role | LLM Application | Model Training |\n|---|---|---|\n| **Modifiable** | Prompt, context, schema, model params | Architecture, optimizer, hyperparameters |\n| **Locked** | Evaluation function, test set, metric | Dataloader, tokenizer, time budget |\n| **Findings log** | `findings.json` | `results.tsv` |\n\n**Core principles** (apply to everything):\n- Single metric, chosen upfront, never changed mid-experiment\n- One change at a time — isolate what caused the improvement\n- Keep/discard via `git reset` — no exceptions\n- Simplicity criterion — equal metric with less code is a win\n- Autonomous loop — never stop to ask, run until manually halted\n\n**Mode A: LLM Application** — for API routes, prompt chains, agents:\n- `findings.json` pattern: log every run's change, metric, status, observations\n- History injection: inject prior runs into the prompt so the LLM tracks its own progress\n- Context accumulation: failures → system prompt constraints; successes → few-shot examples\n\n**Mode B: Model Training** — for PyTorch/JAX training loops:\n- Modern transformer architecture defaults (RoPE, Flash Attention, GQA, RMSNorm, softcap)\n- Heterogeneous optimizer (Muon for matrices, AdamW for embeddings/scalars)\n- Training loop hygiene (GC management, loss explosion fast-fail, time budget)\n\n## Origin\n\nBased on Andrej Karpathy's [autoresearch](https://github.com/karpathy/autoresearch) project — a single-GPU LLM training setup designed for autonomous overnight experimentation by AI agents.\n\n## License\n\nMIT\n","readmeFilename":"README.md","_rev":"1-74e10ec3f5df11ab0dab12a6034916ea"}