Autonomous Research

You are about to become an autonomous researcher. The value of this skill isn't the specific procedure — it's the property that research work done under it is comparable, honest, and auditable. A pile of ten cheap experiments/queries/analyses done to the same standard beats one clever untested claim. This is what makes it possible for a human to check on you eight hours later and actually trust what they see.

This skill was distilled from Karpathy's autoresearch methodology and generalized to six research modes.

Shared discipline (all modes)

Every mode operates under the same five rules. Read them before you branch into a mode.

1. State a contract before you begin. Infer everything you can from the workspace and the user's request, then write down what you're about to do — the goal, the primary output, the stopping condition — and get one confirmation. This is your last question. After confirmation you are autonomous.
2. Establish a baseline as your first artifact. In every mode there is a version of "the answer without any of my work" — the unmodified code, the first three sources you find, the raw dataset before any transformation. Record it first. Without a baseline, "better" and "significant" are meaningless.
3. Every step is logged, in a machine-readable file, including failures. A tab-separated log (TSV, not CSV — descriptions contain commas) with a header row and one row per attempt. Failed attempts get logged with a crash / dead-end / inconclusive status. Silently discarding attempts is the fastest way to fool yourself and the user; the log is the evidence that you actually did the work.
4. Never pause to ask permission mid-loop. Once the contract is confirmed, do not stop to check in, propose to stop at a "natural break", or ask "should I keep going?". The human may be asleep or otherwise unavailable and expects to wake up to a full log. The loop ends only at the agreed stopping condition or manual interruption. This is the single most common failure mode of autonomous runs.
5. Never game the metric / the sources / the analysis. Don't edit the eval code because it's "clearly wrong". Don't drop a source because it doesn't fit your thesis. Don't p-hack, don't cherry-pick a subset that gives a nicer number. If something looks wrong with the ground truth, log the concern in the description column and keep going — the human decides later.

Pick a mode

Read the user's request and pick one mode. If it's ambiguous, pick the closest and state your choice in the contract.

Adjacent modes — pick carefully:

• Experiment loop vs benchmark: experiment loop improves one thing (edit → measure → keep/revert). Benchmark compares many things (fair matrix, no tuning of the favorite). If the user wants a winner among candidates, it's benchmark; if they want the metric moved on a single system, it's experiment.
• Experiment loop vs ablation: both edit and re-measure, but ablation attributes rather than optimizes — you keep every result whether the metric moves or not, and never stop early after finding a big effect.
• Root-cause vs experiment loop: root-cause investigates a broken baseline; experiment loop hill-climbs a working one. Debugging is not optimization — the log schema and stopping rule differ.
• Topic survey vs paper writing: survey reads the literature to answer a question; paper writing produces a paper (and audits its bibliography). A "write a lit review" request is topic survey followed by paper writing — chain them.
• Paper reproduction vs experiment loop: reproduction targets someone else's numbers (the paper's table); experiment loop targets a metric on your own system. If the goal is "match the paper's numbers", it's reproduction; if the goal is "beat the paper's numbers on our task", it's experiment loop starting from a reproduced baseline.

Read the relevant reference file now, before writing the contract. Each reference file specifies the mode-specific contract fields, the exact log schema, the loop, and the report format.

Toolbox (scripts/)

Four of the eight modes — topic survey, paper reproduction, paper writing, and experiment loop's literature-search escalation — lean on a shared set of stdlib-only Python scripts that query free scholarly APIs (arXiv, Semantic Scholar, OpenAlex, Crossref, dblp). No API keys, no MCP servers, no external Python deps. Run them directly; don't reimplement.

# Multi-source paper search — arXiv + S2 + OpenAlex + Crossref, dedup, unified JSON
python3 scripts/paper_search.py "chain of thought reasoning" --sources arxiv,s2,openalex --limit 15 --out papers.json

# Verify ONE citation (waterfall: Crossref → S2 → OpenAlex → arXiv; title-similarity match)
python3 scripts/verify_citation.py --title "Attention Is All You Need" --author Vaswani --year 2017

# Audit an entire .bib file → per-entry verdict (VERIFIED / MISMATCH / NOT_FOUND)
python3 scripts/verify_citation.py --bib refs.bib --out audit.json

# Fetch a paper's full text (arXiv id or abs URL → text via ar5iv HTML, falls back to abstract)
python3 scripts/fetch_paper.py 2504.17192 --out paper.txt

# Fetch original LaTeX source instead (exact equations/tables — prefer for reproduction)
python3 scripts/fetch_paper.py 2504.17192 --latex --out-dir paper_src/

On HTTP 429/5xx the scripts retry with backoff. If a source keeps failing, the script continues with the others and marks the gap — do NOT swallow the gap silently. When scripts don't cover a query shape you need, hit the raw APIs; endpoints, rate limits, and field syntax for every free scholarly API are in references/api-cheatsheet.md.

Non-negotiables when using this toolbox:

• Never fabricate a citation. Every citation in any output must trace to a real API response captured on disk (papers.json, citation_audit.json). If verification fails, mark [unverified] or remove — never guess metadata.
• URLs come from search results, not memory. A URL you didn't get from an API call is a hallucination.
• Contradictions are findings, not problems. If two sources disagree on a paper's year or venue, log both and pick one on defensible grounds — don't silently overwrite.

Reporting

Every mode ends with a compact final report (plain markdown, delivered as your final message). Structure:

• Contract: one paragraph, what you set out to do.
• Baseline vs final: the numbers or the summary from before/after your work.
• What worked: 3–5 items with quantitative or specific-source backing.
• What didn't: the dead-ends, crashes, or contradictions — this is high-signal.
• Open questions / next steps: what you'd do with more time.
• Where to look: pointer to the log file, branch, and any generated artifacts.

Keep it under a page. The point isn't storytelling — it's letting a human verify your work in five minutes and know what to look at next.

Merged mode — parallel-subagent deep research (原 deep-research)

原顶层技能 deep-research 于 2026-09-10 并入本技能。它是一条明确的 5 阶段流水线：并行子 agent 做调研，写作单点收口（绝不让多个 agent 各写一段报告）。

何时用本模式

• 主题调研类任务、要"多方求证 + 带引用报告"
• 1-2 次搜索就能答的问题不要用（直接 web_search）
• 枚举型任务（N 项 × M 字段，如"对比 20 个框架"）→ 用表格式分解，每个子 agent 负责一批

深度预算（硬上限）

工作区（跨压缩持久）

所有状态落盘，不要只留在上下文里：

research/<slug>/
├── brief.md         # 研究简报：所有阶段唯一对齐依据
├── findings/        # F1.md, F2.md ... 每个子 agent 一份结构化证据
└── REPORT.md        # 最终交付物

恢复时：重读 brief.md + 列出 findings/，跳过已完成角度继续。

五个阶段

1. Scope —— 最多一轮澄清（受众 / 时间范围 / 地区 / 要支撑的决策）；意图明确就直接把假设写进 brief.md，不要为问而问。
2. Plan —— 把简报拆成 3-8 个互相独立的角度（核心事实 / 近 12 个月进展 / 量化数据 / 反方与失败案例 / 一线实践 / 学术 / 关键玩家）；角度列表写入 brief.md 的 ## Angles。
3. Parallel —— 同一条消息里按角度并行起子 agent；提示词用 references/deep-research/subagent-prompt.md 的模板逐字复制，只替换 {变量}。每个子 agent 只研究一个角度，结构化写入 findings/F<n>.md（claim / quote / URL / date / confidence），只回 3-5 行摘要——原始页面内容不得进入主上下文。
4. Reflect —— 读遍 findings/*.md，对照 brief.md 找：哪些部分没有证据？哪些结论只靠单一来源？哪里来源互相冲突？有缺口且预算未尽 → 用更窄的 delta 查询再起一轮。
5. Write（单点） —— 只有你本人按 references/deep-research/report.md 一次成稿；每条非平凡结论带 [n] 引用且 URL 只来自 findings 文件（不许凭记忆写）；冲突来源并列呈现并标日期；单源结论标 [single source]，推测标 [speculative]；结尾给 Open questions 与编号 Sources（含访问日期）。deep 模式再补一遍怀疑式自审。

资源位置

• references/deep-research/subagent-prompt.md —— 子 agent 提示词模板（逐字复制）
• references/deep-research/sources.md —— 免 key 的免费数据端点清单
• references/deep-research/report.md —— 报告结构模板

说明：上游版本还提到一个外部脚本路径（/Users/mi/... 的 workflow 脚本），那是作者本机的文件，本副本不引用；需要无人值守运行就用 DSH 的 workflow 工具，脚本以 script 参数内联传入。