Daily Research Review

Daily Research Review

Core Contract

Treat the review as a personal research-company operating review, not a diary. The review must answer two questions:

1. Did today turn time, AI, papers, code, experiments, and communication into reusable research assets?
2. Did today make the user more like a researcher worth betting on: better at defining problems, testing hypotheses, building systems, and judging direction?

Default language is Chinese. Use English terms when they are the natural technical labels.

Also use chinese-technical-writing for reader-facing Chinese review prose and TODOs: keep method names, acronyms, project names, and simulation terms when they are names, but translate ordinary technical concepts into Chinese instead of mixing raw English phrases into Chinese sentences.

For weekly literature picks / Week Summary pages, use weekly-research-summary instead of this skill.

Platform Placement

Use research-doc-workflow and the matching *-doc-workflow adapter for destination selection and durable writes. Preserve an existing review hub's platform; do not create a parallel Feishu copy when the active review system is Notion or Obsidian.

If the user brings a research-method article or web source into the review context and asks for 原文, 提取原文, or 英文原文与中文译文, use bilingual-source-archive as a side path. Do not fold the source into the daily review as a summary unless the user explicitly asks for a review takeaway.

Keep this logical hierarchy on every platform:

• Parent context page: Entrepreneurial Research Mindset
• Review hub: 科研经营复盘｜Daily Research Operating Review
• Dated entries under or linked from the hub: YYYY-MM-DD 科研经营复盘

Map it natively:

• Feishu/Lark: wiki parent, review hub, and dated child page through feishu-doc-workflow.
• Notion: review hub plus dated subpage or database record with date/status properties through notion-doc-workflow.
• Obsidian: review MOC/index plus dated Markdown note in the established review folder through obsidian-doc-workflow, with frontmatter and links following neighboring notes.

Before writing, fetch or read the hub/index and dated entries. If today's entry exists, update it; otherwise create it in the same system. Verify after writing that:

• The title is not Untitled / 无标题 and the date is correct.
• The entry is linked to or contained by the review hub, not left orphaned.
• The final page contains 明日三项关键 Todo.

Review Flow

Start by asking only the next useful questions. Do not ask the whole template at once if the user is already answering naturally.

Use this order:

1. Establish today's date and the user's current stage or context.
2. Capture project-level progress.
3. Capture assets created today.
4. Identify hypotheses validated, weakened, or still unknown.
5. Evaluate AI / GPT / Codex ROI.
6. Identify risks and bottlenecks.
7. Derive tomorrow's role-based TODOs, then reduce them to exactly three key TODOs.

If the user gives a long free-form answer, first reflect it back as structured notes, then ask for missing high-value information.

Good Research Loop Lens

Use the review to train research taste, not just record activity. When relevant, extract these signals and place them into the existing required sections:

• Problem ownership: did the user choose the problem from their own reasoning, or merely absorb an advisor's task, a trend, or a paper's framing?
• Important and attackable: is the current question both worth solving and small enough to test with the next concrete action?
• Forecast before result: what did the user expect before reading a paper, asking for advice, running an experiment, or seeing a metric?
• Belief update: after the result, what changed in the user's model of the project, method, baseline, data, or evaluation?
• Loop speed: how quickly did the user discover they were wrong, and what tooling, command, config, plot, or AI workflow could shorten that loop tomorrow?
• Output inspection: did the user inspect raw outputs, failure cases, visualizations, transcripts, samples, or geometry artifacts instead of only trusting aggregate metrics?
• Failure clustering: what is the largest pile of failures, and what is the smallest attack on that pile?
• Input quality: did today's reading include primary papers, appendices, limitations, old or underpriced sources, and not only social summaries?
• Focus versus exploration: did the user protect the main project line while giving exploratory ideas a bounded budget?
• Generous asset: did today's work produce something reusable for future self, collaborators, public writing, code, Notion/Obsidian/Feishu documentation, or paper writing?

Map these signals into the template rather than adding new daily sections. For example, forecast and belief update belong in 今日假设验证; loop speed and output inspection belong in 今日资产沉淀, GPT / Codex / AI 投资回报, or 当前最大风险; problem ownership and importance belong in 项目经营判断.

Required Sections

Write each daily page with these sections:

# YYYY-MM-DD 科研经营复盘

## 0. 今日语境
## 1. 今日最重要的推进
## 2. 今日资产沉淀
## 3. 今日假设验证
## 4. 项目经营判断
## 5. GPT / Codex / AI 投资回报
## 6. Simulation-Ready 方向检查
## 7. 当前最大风险
## 8. 明日角色化 Todo
## 9. 明日三项关键 Todo

Use ## 6. Simulation-Ready 方向检查 every day. The user has stated that simulation will be a future must-have capability, so projects should be reviewed for whether they move toward simulation-ready / simulation-capable outputs.

Project Triage Lens

For every active project mentioned, record:

• Current priority: short-term critical / medium-term important / long-term thesis line / parked.
• Today's concrete movement.
• Next evidence needed.
• Whether it creates a reusable asset.
• Whether it improves simulation-readiness.

Current known project lines:

• UniGAvatar: short-term critical. Baseline experiment documentation, result analysis, geometry loss, semantic experiments.
• 3DGS4SMPL: document and asset consolidation; needs deeper analysis.
• GaussianAvatar: July-August delivery line. Need to complete the SMPL-X pipeline and reuse mature ExAvatar components; tied to paper completion and submission.
• Home-scene world model for SMPL-X planning: long-term PhD thesis line. Needs sustained learning and literature grounding.
• Single-pose simulation-ready avatar reconstruction: emerging idea. Inspired by PGC-like thinking; multi-view single-pose reconstruction may be a better route to accurate simulation-ready meshes than monocular-video fitting.

These are context anchors, not fixed facts. Update them when the user corrects priorities.

Hypothesis Language

Phrase hypotheses explicitly:

• 假设：...
• 今日证据：支持 / 反驳 / 不确定
• 下一步：继续 / 收缩 / 转向 / 停止
• 预测：...
• 结果：...
• 更新后的判断：...

Do not convert ideas into conclusions too early. For new ideas, preserve the user's intuition and mark what evidence is still missing.

AI ROI Lens

Evaluate AI usage by outputs, not interaction volume:

• Time saved.
• Judgment improved.
• Assets created or cleaned.
• Experiments, code, papers, or durable research pages advanced.
• Whether AI use became unproductive chatting.
• Whether AI shortened the research loop by making experiments, plots, comparisons, source reading, or failure inspection faster.
• Whether the user kept ownership of the key judgment instead of outsourcing taste, problem choice, or claims to AI.

Role-Based TODOs

Use these roles to find blind spots:

• CEO: direction, priority, and strategic bet.
• CTO: technical route and module choices.
• Research Scientist: hypothesis and novelty.
• Research Engineer: smallest executable experiment or code step.
• PM: milestone, scope, and acceptance criteria.
• CFO: time, GPU, GPT, and attention ROI.
• Data / Infra: data, environment, logs, visualizations, failure cases.
• Reviewer: strongest attack on novelty, evaluation, claims, and baselines.
• Writer / Communicator: Notion/Obsidian/Feishu documentation, paper, slides, mentor communication.
• AI Manager: what to delegate to AI and what judgment the user must own.
• Research Taste Coach: forecast before seeing answers, compare with reality, and convert the gap into a sharper next experiment or reading question.

After role scanning, keep exactly three execution TODOs:

1. CEO Todo: the one action most affecting direction survival.
2. Research / Engineering Todo: the smallest experiment, code, paper, or analysis action.
3. Asset Todo: the one thing that must become a reusable research asset.

Writing Style

Be concrete and decision-oriented. Avoid motivational fluff.

Good wording:

• 今天真正推进的是...
• 这仍然不是结论，但它把下一步验证问题收缩为...
• 明天的验收标准是...

Avoid:

• Generic diary prose.
• Overconfident conclusions from weak evidence.
• Long project descriptions that do not end in an action or decision.