ai-research-workflow
AI Research Workflow
基于既有科研工作流调研报告组织的跨平台入口。历史源文档保留在飞书: <FEISHU_OR_LARK_URL>
All durable document writes must also use research-doc-workflow plus the matching platform adapter (feishu-doc-workflow, notion-doc-workflow, or obsidian-doc-workflow). The source of a workflow may be Feishu, but the destination is selected from Feishu, Notion, or Obsidian according to the user's current URL or explicit instruction.
Observation Before Method Design
When a user reports a research phenomenon, failure case, unexpected result, or desired method improvement, first follow references/observation-to-problem.md. Verify the observation, inspect related work and existing explanations/solutions, and distinguish competing causes before selecting a method change. Related-work investigation is required before recommending a concrete improvement as the next implementation step; use relevant previously verified sources or perform a targeted search and inspect primary evidence. A list of titles, an uninspected repository, or a generic citation does not satisfy this step.
If the cause remains uncertain, proceed with a diagnostic check/experiment that separates explanations. Do not substitute an architectural proposal or an improvement MVE for cause analysis and literature investigation. A short evidence-based diagnosis is sufficient; no full survey, paper translation, or repeated user approval is required. Explicit requests to defer investigation may be handled as provisional brainstorming, with assumptions and unresolved literature clearly identified rather than treated as established findings.
Treat cause analysis, related work, and diagnostic evidence as an iterative loop. Locating a failing module is not yet a mechanism explanation. Each recommended algorithm change must identify its hypothesis, evidence, predicted effect, and falsifier. Keep research priorities human-led; external autonomous-research systems are references, not an automatic stage machine or authorization to expand experiments.
Project-Centered Workflow
For a new project, an existing project's organization, baseline reproduction, or dataset intake, read references/project-workflow.md. It owns the project-level relationship between Related Work, Reproduction, Datasets, Ideas, Evaluation, Experiments, and Handoff. Use the linked templates as starting points and fill only evidence-backed fields.