Import Docs

Procedure

1. Scan Document Folder

List all files in the --docs folder (always quote the path — the PI's folder may contain spaces):

find "$docs_path" -type f \( -name "*.pdf" -o -name "*.docx" -o -name "*.doc" -o -name "*.xlsx" -o -name "*.csv" -o -name "*.pptx" -o -name "*.png" -o -name "*.jpg" -o -name "*.jpeg" -o -name "*.bib" -o -name "*.ris" -o -name "*.md" -o -name "*.txt" -o -name "*.html" \) | sort

Report the file list to the user:

Found N documents in "$docs_path":
  1. cv_popescu_2024.pdf
  2. paper_gnn_binding.pdf
  3. budget_template.xlsx
  4. letter_collaborator_munich.pdf
  ...

2. Convert Documents to Text

For each document, extract text content:

PDF and DOCX files — use /markitdown if available (handles PDF, DOCX, PPTX, XLSX with OCR):

/markitdown <file_path>

If /markitdown is not available, fall back to:

• PDF: uv run gw-pdf <file_path>
• DOCX: Use the Read tool directly (Claude can read .docx)
• XLSX: Use the Read tool directly (Claude can read .xlsx)

Image files (PNG, JPG): Use the Read tool to view with Claude vision.

BibTeX/RIS files: Read as plain text.

3. Classify Each Document

For each document, determine its type by analyzing the content. Use these signals:

Report classifications to the user:

Document classification:
  cv_popescu_2024.pdf          → CV / Resume
  paper_gnn_binding.pdf        → Research paper
  budget_template.xlsx         → Budget spreadsheet
  letter_collaborator_munich.pdf → Letter of support
  fig_preliminary_results.png  → Figure (preliminary data)
  previous_pce_proposal.pdf    → Previous grant proposal
  foa_horizon_2025.pdf         → FOA / Call document

Human checkpoint: Ask the PI to confirm or correct any misclassifications before proceeding.

4. Extract Structured Information

For each classified document, extract the relevant structured data:

From CV / Resume:

{
  "type": "cv",
  "source_file": "cv_popescu_2024.pdf",
  "extracted": {
    "name": "Dr. Alexandru Popescu",
    "title": "Associate Professor",
    "institution": "Politehnica University of Bucharest",
    "department": "Computer Science",
    "email": "...",
    "orcid": "0000-0002-...",
    "positions": [
      {"title": "Associate Professor", "institution": "...", "from": 2020, "to": "present"},
      {"title": "Postdoc", "institution": "ETH Zurich", "from": 2017, "to": 2020}
    ],
    "education": [...],
    "selected_publications": [
      {"title": "...", "journal": "...", "year": 2023, "doi": "..."},
      ...
    ],
    "grants": [
      {"title": "...", "agency": "UEFISCDI", "role": "PI", "amount": "500,000 RON", "years": "2021-2024"}
    ],
    "h_index": 18,
    "expertise": ["graph neural networks", "drug discovery", "protein modeling"]
  }
}

Populate: supporting/cv_pi.md — reformat into agency-specific CV template.

From Research Papers:

{
  "type": "paper",
  "source_file": "paper_gnn_binding.pdf",
  "extracted": {
    "title": "...",
    "authors": ["..."],
    "journal": "...",
    "year": 2023,
    "doi": "...",
    "abstract": "...",
    "key_results": ["achieved 89% accuracy on PDBbind", "outperformed baseline by 15%"],
    "methods_used": ["equivariant GNN", "message passing", "PDBbind v2020"],
    "figures": ["fig1_architecture.png", "fig2_results.png"],
    "is_preliminary_data": true
  }
}

Populate:

• sections/preliminary_data.md — incorporate key results and methodology
• sections/figures/ — copy relevant figures
• sections/bibliography.md — add as citation
• Inform sections/state_of_the_art.md — PI's own prior work

From Previous Grant Proposals:

{
  "type": "previous_proposal",
  "source_file": "previous_pce_proposal.pdf",
  "extracted": {
    "title": "...",
    "agency": "UEFISCDI",
    "mechanism": "PCE",
    "aims": ["Aim 1: ...", "Aim 2: ..."],
    "methodology_summary": "...",
    "budget_structure": {"personnel": "...", "equipment": "..."},
    "team": [{"name": "...", "role": "PI"}, ...],
    "references_cited": [...],
    "outcome": "funded/not funded (if known)"
  }
}

Populate:

• sections/objectives.md — use as starting point (adapt, don't copy)
• sections/methodology.md — reuse proven approaches
• landscape/prior_support.md — document previous funding
• sections/bibliography.md — merge relevant references

Important: Flag to the PI that content from previous proposals should be adapted, not copied verbatim — especially if the previous proposal was funded (overlap issues).

From FOA / Call Documents:

{
  "type": "foa",
  "source_file": "foa_horizon_2025.pdf",
  "extracted": {
    "agency": "horizon",
    "mechanism": "RIA",
    "call_id": "HORIZON-HLTH-2025-...",
    "title": "...",
    "deadline": "2025-09-15",
    "budget_range": "EUR 3-5M per project",
    "page_limits": {"part_b1": 45},
    "required_sections": [...],
    "review_criteria": [...],
    "special_requirements": ["gender dimension", "open access", "ethics"]
  }
}

Populate: foa_requirements.json — feeds directly into the FOA analysis phase, potentially skipping it entirely.

From Budget Spreadsheets:

{
  "type": "budget",
  "source_file": "budget_template.xlsx",
  "extracted": {
    "personnel": [
      {"name": "PI", "monthly_salary": 15000, "effort_pct": 50, "months": 36},
      {"name": "Postdoc", "monthly_salary": 10000, "effort_pct": 100, "months": 36}
    ],
    "indirect_rate": 0.25,
    "equipment": [{"item": "GPU Server", "cost": 50000}],
    "travel": 25000,
    "consumables": 30000,
    "currency": "RON"
  }
}

Populate: budget/budget_input.yaml — ready for gw-budget to calculate.

From Letters of Support:

{
  "type": "letter",
  "source_file": "letter_collaborator_munich.pdf",
  "extracted": {
    "from": "Prof. Hans Mueller",
    "institution": "Technical University of Munich",
    "commitment": "Will contribute 6 person-months of computational resources and co-supervise 1 PhD student",
    "areas": ["molecular dynamics", "force field development"]
  }
}

Populate: supporting/letters/letter_mueller_tum.md — reformatted into grant template.

From Review Feedback (for resubmission):

{
  "type": "review_feedback",
  "source_file": "summary_statement_2024.pdf",
  "extracted": {
    "scores": {"excellence": 3, "impact": 4, "implementation": 2},
    "overall": "Below threshold",
    "criticisms": [
      {"reviewer": 1, "category": "methodology", "text": "Statistical plan is insufficient..."},
      ...
    ]
  }
}

Populate: resubmission/previous_reviews.md — structured for the resubmission skill.

From Figures:

Populate: Copy to sections/figures/ with descriptive filenames. Use Claude vision to generate captions.

From Bibliography Files:

Populate: Merge into sections/bibliography.md, converting BibTeX/RIS to the proposal's citation format.

5. Generate Import Report

Create <proposal_dir>/review/import_report.md:

# Document Import Report

**Source folder**: <docs_path>
**Documents processed**: N
**Date**: <timestamp>

## Imported Successfully

| Source Document | Type | Populated |
|----------------|------|-----------|
| cv_popescu_2024.pdf | CV | supporting/cv_pi.md |
| paper_gnn_binding.pdf | Paper | sections/preliminary_data.md, sections/figures/fig1.png |
| budget_template.xlsx | Budget | budget/budget_input.yaml |
| letter_mueller.pdf | Letter | supporting/letters/letter_mueller_tum.md |
| foa_horizon.pdf | FOA | foa_requirements.json |

## Pre-populated Sections

| Section | Source | Status |
|---------|--------|--------|
| supporting/cv_pi.md | cv_popescu_2024.pdf | Complete draft — PI should verify |
| budget/budget_input.yaml | budget_template.xlsx | Ready for calculation |
| sections/preliminary_data.md | paper_gnn_binding.pdf | Draft with 2 figures — PI should review |
| sections/bibliography.md | paper_gnn_binding.pdf + refs.bib | 15 references imported |
| landscape/prior_support.md | previous_pce_proposal.pdf | Previous UEFISCDI PCE documented |
| foa_requirements.json | foa_horizon.pdf | Extracted — verify deadline and page limits |

## Still Needs Manual Input

| What | Why |
|------|-----|
| Institutional indirect cost rate | Not found in any document — ask PI |
| Facilities description | No facilities document provided |
| Data management plan | No previous DMP found |
| Ethics self-assessment | Requires PI input on human subjects, animals, data |
| Collaboration details | Letters mention collaborations but scope needs PI confirmation |

## Warnings

- **previous_pce_proposal.pdf**: Extracted aims and methodology as starting points. DO NOT copy verbatim — adapt for the new proposal to avoid overlap issues.
- **budget_template.xlsx**: Salary figures from 2023 — PI should confirm current rates.
- **fig_results.png**: Resolution is 150 DPI — may need higher resolution for submission.

6. Human Checkpoint

Present the import report to the PI:

1. Confirm classifications were correct
2. Review pre-populated files — especially CV, budget, and preliminary data
3. Identify what's still missing — the "Still Needs Manual Input" table shows exactly what the PI needs to provide
4. Flag reuse warnings — previous proposal content must be adapted, not copied

Ask: "I've imported N documents and pre-populated M sections. Please review the import report. Which items need correction, and can you provide the missing information?"

7. Update State

uv run gw-state update <proposal_dir> --phase import_docs --status complete

The orchestrator can now skip or shorten phases where data was already imported:

• If FOA was imported → Phase 1 (foa-analysis) can skip extraction, just verify
• If CV was imported → Phase 7 (supporting-docs) biosketch is pre-filled
• If budget was imported → Phase 6 (budget) input is pre-filled
• If previous reviews imported → Phase 9.5 (resubmission) has structured feedback ready
• If papers imported → Phase 4 (preliminary-data) has figures and results ready

Error Handling

• /markitdown unavailable: Fall back to uv run gw-pdf for PDFs and Claude's native Read tool for DOCX/XLSX.
• gw-pdf missing: Run uv run gw-verify; if not installed, warn the PI and skip non-text documents rather than halt.
• Unreadable file (corrupt PDF, password-protected): Log the error, mark the file as skipped in the import report, and continue.
• Classification ambiguous: Default to "Notes / Outline" and defer to the human checkpoint for correction.
• Force vs skip conflict: If --force is not set and a target file exists, skip with a log entry — never silently overwrite PI work.
• Previous proposal overlap risk: Always flag extracted aims/methodology from previously funded proposals for explicit adaptation review.