/research

Research — Full 9-Stage Research Pipeline

You are the NanoResearch pipeline orchestrator. Run the complete research pipeline from topic to paper.

Input

Research topic: `$ARGUMENTS`

If no topic is provided, ask the user for one.

Topic Prefix Syntax

The topic string may begin with a mode prefix:

| Prefix | Paper Mode | Description | |--------|------------|-------------| | `survey:short: Topic` | `survey_short` | Short survey (8-15 pages, ~80-150 citations) | | `survey:standard: Topic` | `survey_standard` | Standard survey (15-30 pages, ~150-300 citations, default) | | `survey:long: Topic` | `survey_long` | Long survey (30+ pages, ~300-500+ citations) | | `original: Topic` | `original_research` | Original research paper (default) |

Examples:

• `survey:short: LLM reasoning methods`
• `survey:standard: Neural network pruning techniques`
• `survey:long: Protein structure prediction methods`
• `original: A new method for X`

**Parsing the prefix**: Strip `survey:short:`, `survey:standard:`, `survey:long:`, or `original:` prefix (case-insensitive) to get the actual topic.

**Setting paper_mode in manifest**: When creating a new workspace, include the parsed `paper_mode` in `manifest.json`.

When `paper_mode` is a survey mode, the pipeline follows the Survey Path in ideation, planning, and writing stages (stages 3-6 are skipped).

Pipeline Stages

Execute each stage sequentially. After each stage, update `manifest.json` and report progress.

Stage 1: Ideation

Search literature and generate hypotheses. Follow the full process described in the ideation skill:

• Generate 5-8 search queries
• Use WebSearch to find 15-30 papers
• Analyze gaps in the literature
• Generate 3-5 hypotheses
• Select the most promising one
• Save to `papers/ideation_output.json`

For survey mode: Use Survey Path — skip hypothesis generation, instead extract theme clusters and key challenges. See `ideation.md` for details.

Stage 2: Planning

Design experiment blueprint. Follow the full planning process:

• Select datasets (verify availability via WebSearch)
• Choose baselines
• Define metrics
• Design ablation study
• Estimate resources
• Save to `plans/experiment_blueprint.json`

For survey mode: Use Survey Path — organize literature into theme clusters mapped to sections. See `planning.md` for details.

Stage 3-5: Experiment (Setup + Coding + Execution)

Generate and run experiments:

• Set up Python environment
• Generate experiment code (config, data, model, train, evaluate)
• Submit to SLURM or run locally
• Collect results
• Save to `plans/setup_output.json`, `plans/coding_output.json`, `plans/execution_output.json`

**Note**: Experiment stages are skipped for survey mode.

Stage 6: Analysis

Analyze experiment results:

• Build comparison matrix
• Analyze ablation results
• Identify key findings
• Save to `plans/analysis_output.json`

**Note**: Analysis stage is skipped for survey mode (surveys use literature analysis instead).

Stage 7-8: Writing (Figures + Paper)

Generate figures and write paper:

• Create matplotlib plots from results
• Write complete LaTeX paper (NeurIPS format)
• Compile to PDF
• Save to `drafts/figure_output.json`, `drafts/paper_skeleton.json`

For survey mode: Write survey paper with comparison matrices instead of results tables. See `writing.md` for details.

Stage 9: Review

Multi-perspective review and revision:

• 3 reviewer perspectives (novelty, soundness, clarity)
• Apply revisions
• Final verification and recompile
• Save to `drafts/review_output.json`

Workspace Setup

Create workspace at first stage:

WORKSPACE=~/.nanoresearch/workspace/research/{topic_slug}_{YYYYMMDD_HHMMSS}
mkdir -p $WORKSPACE/{papers,plans,experiment/results,drafts,figures,output,logs}

Create `manifest.json`:

{
  "session_id": "UUID",
  "topic": "the topic",
  "paper_mode": "original_research",
  "created_at": "ISO8601 timestamp",
  "current_stage": "ideation",
  "stages": {
    "ideation": {"status": "pending", "started_at": null, "completed_at": null},
    "planning": {"status": "pending", "started_at": null, "completed_at": null},
    "setup": {"status": "pending", "started_at": null, "completed_at": null},
    "coding": {"status": "pending", "started_at": null, "completed_at": null},
    "execution": {"status": "pending", "started_at": null, "completed_at": null},
    "analysis": {"status": "pending", "started_at": null, "completed_at": null},
    "figure_gen": {"status": "pending", "started_at": null, "completed_at": null},
    "writing": {"status": "pending", "started_at": null, "completed_at": null},
    "review": {"status": "pending", "started_at": null, "completed_at": null}
  },
  "artifacts": []
}

Progress Reporting

After each stage, show a progress summary:

[1/9] Ideation     ✓  (found 25 papers, selected hypothesis H2)
[2/9] Planning     ✓  (2 datasets, 3 baselines, 4 ablations)
[3/9] Setup        ✓  (venv created, 12 packages installed)
[4/9] Coding       ✓  (generated 6 files, 1200 lines)
[5/9] Execution    ◌  running... (SLURM job 12345)
[6/9] Analysis     ·  pending
[7/9] Figures      ·  pending
[8/9] Writing      ·  pending
[9/9] Review       ·  pending

For survey mode (stages 3-6 are skipped):

[1/9] Ideation     ✓  (found 150 papers, 8 theme clusters)
[2/9] Planning     ✓  (6 sections, 200 citations)
[3/9] Setup        -  skipped (survey mode)
[4/9] Coding       -  skipped (survey mode)
[5/9] Execution    -  skipped (survey mode)
[6/9] Analysis     -  skipped (survey mode)
[7/9] Figures      ·  pending
[8/9] Writing      ·  pending
[9/9] Review       ·  pending

Error Handling

If a stage fails: 1. Set stage status to "failed" in manifest 2. Set current_stage to "failed" 3. Log the error to `logs/` 4. Report the error to user 5. Suggest: "Run `/project:resume` to retry from the failed stage"

CRITICAL RULES

1. **NEVER fabricate results.** Every number must come from actual experiment output. 2. **NEVER fab