Academic Deep Research

Academic Deep Research

Use this skill when the user asks for:

• academic reports or literature reviews
• experiment reports with professional references
• references in a specific user-requested style or richer scholarly citations
• innovation-point mining based on papers
• a research knowledge base, evidence map, or local literature store

Core Rules

1. Prefer the academic_research tool over ad hoc web browsing when the task is literature-heavy.
2. When the user asks for latest datasets, benchmarks, leaderboards, baselines, or experiment-ready evidence, call dataset_benchmark_discovery before drafting conclusions.
3. For large academic tasks, delegate to task with subagent_type="academic-researcher" so the main thread stays clean.
4. Do not invent references, DOI metadata, datasets, benchmark scores, or claims that are not grounded in generated evidence. In Synthetic Experiment Mode, simulated personal experiment outputs may support paper workflow completion, but third-party literature, baseline, leaderboard, benchmark, DOI, and dataset facts still cannot be fabricated.
5. Treat the generated report.md, references.md, references.bib, evidence_map.json, and dataset_benchmark_map.json as source material with different roles: literature support is not the same as executed experiment support.
6. When writing final prose in chat, use the artifact bundle first and only then polish wording.
7. Follow the user's requested reference style. Use APA 7 only when no style is specified.
8. For manuscript-style requests, default to LaTeX + PDF and prefer manuscript_export so writing, citation audit, PDF compilation, and artifact presentation happen in one enforced step.
9. Read or audit references.bib before inserting inline LaTeX citations. Use exact BibTeX keys only; do not use \nocite{*} unless the user explicitly asks to list every reference without inline citation placement.
10. If citation or PDF generation fails, report the exact failed tool and error. Do not claim tools are unavailable when file tools, manuscript_export, citation_audit, or present_files are available.

Recommended Workflow

1. Start Structured Research

If the user gives a research topic, run:

• academic_research(topic=..., scope=...)

If the task depends on datasets, benchmarks, or baselines, first run:

• dataset_benchmark_discovery(topic=..., scope=...)

This will automatically:

• expand queries
• retrieve and normalize papers
• deduplicate and rank them
• produce a report bundle

The benchmark discovery bundle records candidate datasets, license/access status, metrics, baseline/SOTA hints, and risks. It does not download gated data or make leaderboard claims final.

2. Build Manuscript Bundle When Requested

For a paper, manuscript, review article draft, or experiment paper:

• create or reuse references.bib
• use dataset_benchmark_map.json to name candidate datasets/benchmarks only after checking access and version/date
• use experiment_lab outputs for executed results, ablations, robustness checks, and error analysis
• keep experimental claims unsupported unless claim_support_matrix.json marks them supported_by_experiment; in Synthetic Experiment Mode, supported_by_simulation is acceptable only when simulation assumptions/disclosure artifacts are present
• write LaTeX with exact BibTeX keys from references.bib
• call manuscript_export(tex_content=..., bibtex_content=..., filename_stem="manuscript")
• if manuscript_export reports missing keys, unsupported claims, blocked \nocite{*}, or compile errors, fix the inputs and retry before final delivery
• use citation_audit and present_files only as fallback tools for narrower manual checks or existing files

3. Escalate to Subagent When Needed

If the task is large, multi-part, or the user wants a full report, use:

• task(description=..., prompt=..., subagent_type="academic-researcher")

The subagent should use academic_research as its primary tool and then summarize the bundle.

4. Output Discipline

When summarizing the generated academic bundle:

• mention project_id
• mention how many core papers and formatted references were retained
• mention citation audit status for manuscript outputs
• mention any evidence gaps
• point the user to the artifacts instead of rewriting everything inline

What Good Looks Like

• the report is evidence-backed rather than generic
• the references list is materially richer than a normal web-search answer
• reference entries are normalized, deduplicated, and formatted in the requested style
• innovation directions are tied to literature gaps, not speculation