Skill: Export

Purpose

Export analysis results in different formats for different audiences. Converts pipeline outputs into ready-to-share deliverables.

Invocation

/export slides — generate/refresh Marp slide deck from latest analysis /export email — write an executive summary email (markdown) /export slack — write a concise Slack update (markdown) /export brief — write a 1-page decision brief (markdown) /export data — export analysis data tables as CSV /export gdoc — create a Google Doc with full Analysis Readout (charts, SQL, bookmarks) /export docx — generate a local .docx Word document (no Google upload) /export notion — export analysis to a Notion page (with charts, data stamps, provenance toggles) /export receipt — generate full analysis receipt (Reproduce-level audit trail) /export all — generate all text formats + data (does NOT include gdoc, notion, or receipt — use /export gdoc, /export notion, and /export receipt separately)

Instructions

Step 1: Find Source Material

Primary source detection (use exactly ONE of these, in order):

1. outputs/narrative_*.md — if exists, use the most recent by date (check filename date suffix)
2. outputs/analysis_*.md — if narrative absent, use most recent analysis
3. working/pipeline_summary.md — if no outputs/ files exist
4. working/storyboard_*.md — last resort if above are all missing

How to choose when multiple files exist: Sort by filename date (YYYYMMDD or YYYY-MM-DD suffix), use the latest. If no date in filename, use file modification time.

Supporting materials (collect when available, but NOT primary source):

• outputs/charts/*.png — chart images
• outputs/validation_*.md — confidence score
• outputs/close_the_loop_*.md — success tracking + action items
• working/sql_queries/*.sql — SQL queries

If no outputs exist:

• Check working/ for partial results
• If nothing found: "No analysis results to export. Run an analysis first or use /run-pipeline."

Read the primary source in full before generating any export.

Step 2: Generate Requested Format

Format: slides

• If deck already exists, ask: "Deck found at {path}. Regenerate or export as-is?"
• If no deck, invoke Deck Creator agent with latest narrative + charts
• Output: outputs/slides_{DATE}.md

Format: email

• Structure: Subject line + 3-paragraph body (context, key finding, recommendation)
• Tone: Executive-friendly, no jargon, action-oriented
• Include: 1-2 key numbers, the "so what", and a clear ask
• Output file: Write the email content to outputs/email_summary_{DATE}.md where {DATE} is today's date in YYYY-MM-DD format (e.g., email_summary_2026-04-04.md). This specific file path is important for consistent organization.

Format: slack

• Structure: ONE focused update with bold headline + 3-5 bullet points
• Short enough to read in a channel without expanding; one headline, a few bullets
• Use emoji sparingly (checkmarks, arrows only)
• Include: key metric, direction, and recommended action
• Include a data stamp for each key finding (abbreviated format): 50K | Jan-Mar 2026 | EVENTS | B (82)
• Output file: Write to outputs/slack_update_{DATE}.md where {DATE} is today's date in YYYY-MM-DD format (e.g., slack_update_2026-04-04.md)

Format: brief

• Structure: Title + Executive Summary (3 sentences) + Key Findings (numbered) + Recommendation + Next Steps + Appendix (data sources, methodology)
• One page: a reader should get the decision and the evidence without scrolling
• Output file: Write to outputs/decision_brief_{DATE}.md where {DATE} is today's date in YYYY-MM-DD format (e.g., decision_brief_2026-04-04.md)

Format: data

• Export all DataFrames from working/ as CSVs to the outputs/data/ directory
• Filename pattern: Use descriptive names based on what the data represents (e.g., conversion_by_platform.csv, funnel_steps.csv)
• Create outputs/data/README.md documenting each CSV file: columns, row count, use cases, data quality notes
• Export only the source DataFrames that were actually used in the analysis
• Output location: All CSV files in outputs/data/ directory plus outputs/data/README.md manifest

Format: gdoc

Creates a formatted Google Doc from the analysis with embedded charts, styled headings, internal bookmark links, and SQL code blocks. Follows the Analysis Readout template: Summary (30-second read) → Analysis (30-minute read) → Resources.

Formatting source of truth: the google-doc-export skill owns Doc formatting standards (heading hierarchy, image placement, spacing). This path produces that same look by building a local .docx (via gdoc_builder) and uploading with convert_to_google_doc=True — a distinct method from native MCP construction, not a restatement of it.

Step 0: Auth Check

Check if mcp__google-docs__* tools are accessible:

1. Run the auth-preflight skill (it probes with a create call, not a read).
2. If auth works: Proceed to Step 2a. Say nothing about auth.
3. If MCP tools unavailable or auth expired: Say "Connecting your Google account..." and run authorize_google_docs. Follow the browser OAuth flow.
4. If auth still fails after one attempt: Say "Google Docs connection failed. Generating a local Word document instead." Fall through to the docx format (generate .docx only, skip upload). Provide manual upload instructions: "You can upload {docx_path} to drive.google.com manually."

Step 2a: Re-export Detection

Check if outputs/gdoc_export.yaml exists:

import yaml, os
from helpers.pipeline.file_helpers import content_hash

yaml_path = "outputs/gdoc_export.yaml"
if os.path.isfile(yaml_path):
    with open(yaml_path) as f:
        state = yaml.safe_load(f)

    # Check if source has changed
    narrative_path = state.get("source_narrative")
    if narrative_path and os.path.isfile(narrative_path):
        current_hash = content_hash(open(narrative_path).read())
        if current_hash == state.get("source_hash"):
            # Source unchanged — ask user
            url = state.get("document_url", "")
            # Say: "Google Doc already exists at {url}. No analysis changes
            #        detected. Open it, or force re-create?"
            # If user says open: done. If force: continue to Step 2b.

Step 2b: Parse and Build .docx

from helpers.export.gdoc_narrative_parser import parse_pipeline_outputs
from helpers.export.gdoc_builder import build_readout

# Say: "Building document from analysis..."

data = parse_pipeline_outputs(base_dir=".")
docx_path = build_readout(data, output_dir="outputs")

# docx_path is now something like: outputs/report_mobile_checkout_20260403.docx

Confidence gate: If data.confidence_grade is D or F, pause and ask: "This analysis has low confidence (grade {grade}). The document will include a prominent caveat. Create anyway?" If the user says no, abort. If yes, continue.

Step 2c: Upload to Google Drive

# Say: "Uploading to Google Drive..."

result = mcp__google-docs__upload_file_to_drive(
    file_path=docx_path,
    convert_to_google_doc=True
)
# result: {"file_id": "...", "url": "...", "name": "..."}

If upload fails: fall back to .docx. Say: "Drive upload failed. Your analysis is saved locally at {docx_path}."

Step 2d: Write State

import yaml
from datetime import datetime
from helpers.pipeline.file_helpers import content_hash

state_path = "outputs/gdoc_export.yaml"

# Read existing state for version history
existing = {}
if os.path.isfile(state_path):
    with open(state_path) as f:
        existing = yaml.safe_load(f) or {}

version = existing.get("version", 0) + 1
history = existing.get("versions", [])
if existing.get("document_id"):
    history.append({
        "version": existing.get("version"),
        "document_id": existing.get("document_id"),
        "document_url": existing.get("document_url"),
        "created_at": existing.get("created_at"),
    })

narrative_path = _find_latest("narrative_*.md", "outputs")
state = {
    "document_id": result["file_id"],
    "document_url": result["url"],
    "title": data.title,
    "created_at": datetime.now().isoformat(),
    "source_narrative": narrative_path,
    "source_hash": content_hash(open(narrative_path).read()) if narrative_path else None,
    "local_docx": docx_path,
    "charts_embedded": sum(
        1 for f in data.findings
        for sf in f.sub_findings
        if sf.chart_path and os.path.isfile(sf.chart_path)
    ),
    "version": version,
    "versions": history,
}

with open(state_path, "w") as f:
    yaml.dump(state, f, default_flow_style=False)

Step 2e: Post-Upload Verification

After upload, do a quick read-back check:

doc_text = mcp__google-docs__read_document(document_id=result["file_id"])

Quick checks:

• Does the text contain the analysis title?
• Does the text contain "Summary" and "Analysis" headings?
• Does [Chart: appear? (indicates missing chart placeholders made it in)
• Is the document length reasonable (> 500 chars)?

If any check fails, append a note to the report: "Note: some formatting may not have survived conversion. Review the document for any issues."

Step 3: Report to User

Say:

Your analysis is ready:
  Google Doc: {url}
  Local backup: {docx_path}

  Sections: {N findings} findings + recommendations
  Charts: {N} embedded
  Version: {version}

The .docx file is saved locally as a backup. You can share the Google Doc
link with your team.

If this was a re-export, also mention: "This is version {N}. Previous versions are tracked in outputs/gdoc_export.yaml."

Format: docx

Same as gdoc Steps 2a-2b only (parse + build .docx). Skip auth, upload, and state tracking. Report the .docx path to the user.

Say: "Word document saved at {docx_path}. You can upload it to Google Drive manually or share it directly."

Format: notion

Creates a Notion page from the analysis with charts, data stamps, and provenance toggle blocks. Follows the Notion Export skill (.claude/skills/notion-export/SKILL.md).

Step 0: Notion Auth Check

Check if mcp__notion__* tools are accessible. If not: "Notion MCP is not configured."

Generation

Invoke the notion-export agent with:

• NARRATIVE: latest narrative
• CHART_FILES: chart PNGs from outputs/charts/
• DATASET: active dataset
• PROVENANCE_BLOCKS: from cross-verification (if available)
• ANALYSIS_RECEIPT: receipt path (if Tier 3)

The agent handles Analysis Gallery detection, chart hosting, toggle blocks, and self-check.

Output: outputs/notion_url_{{DATASET}}_{{DATE}}.txt

Say: "Analysis exported to Notion: {url}. {N} findings with charts and provenance."

Format: receipt

Generates a full analysis receipt — the Reproduce-level audit trail. Contains every query, methodology decision, cross-verification result, and confidence factor breakdown.

Prerequisites

• Query log must exist (working/query_log_*.jsonl)
• Validation report must exist (outputs/validation_*.md)
• If neither exists: "Cannot generate receipt. Run a full analysis first."

Tier 3 Pre-Export Gate

If the analysis was run at Tier 3, the receipt is generated automatically at step 18.5. Check if outputs/analysis_receipt_*.md already exists:

• If exists and source unchanged: "Receipt already generated at {path}. Open it, or force re-create?"
• If exists but source changed: Regenerate

Generation

Invoke the receipt-generator agent with:

• QUERY_LOG: most recent query log JSONL
• VALIDATION_REPORT: most recent validation report
• CROSS_VERIFICATION_REPORT: cross-verification YAML (if available)
• PIPELINE_STATE: pipeline state JSON (if available)

Output: outputs/analysis_receipt_{{DATASET_NAME}}_{{DATE}}.md

Say: "Analysis receipt generated at {path}. Contains {N} findings, {N} queries, and full validation breakdown."

Format: all

• Run email + slack + brief + data sequentially
• Skip slides if already exists
• Does NOT include gdoc or receipt (external resource creation must be explicit)
• After completion, suggest: "Want to also create a Google Doc? Run /export gdoc. Need a full audit trail? Run /export receipt."

Step 3: Post-Export

• List all exported files with paths
• Suggest: "Copy the email to your clipboard?" or "Want to adjust the tone?"
• For gdoc/docx: suggest sharing or other export formats

Rules

1. Never fabricate findings — only use data from actual analysis outputs
2. Always cite the source analysis date and dataset
3. Adapt detail level to format (email = high-level, brief = medium, data = raw)
4. Apply Stakeholder Communication skill for all text outputs
5. If the analysis had confidence scores, include them in brief format
6. The gdoc, notion, and receipt formats create external resources or audit trails — never include them in /export all
7. Always generate the .docx before attempting Google upload (local fallback)
8. If outputs/gdoc_export.yaml shows unchanged source, offer to open existing doc
9. Every format is one version, built only from numbers and findings in the source narrative; simplify wording, never add derived metrics or alternate versions

Edge Cases

• Partial analysis: Export what's available, note gaps: "Note: validation step was not completed."
• Multiple analyses in outputs/: Use the most recent by date, or ask user which one
• Charts missing: Text formats still work, note: "Charts not available for this export."
• User requests unknown format: List available formats and ask to choose
• gdoc with no Google auth: Fall back to docx format automatically
• gdoc re-export, source unchanged: Offer to open existing doc or force re-create
• gdoc re-export, source changed: Create new doc, preserve version history
• Confidence D/F on gdoc: Warn user before creating doc (only checkpoint in gdoc flow)
• Offline / no network: Generate .docx only. Say: "Saved as Word doc. Upload to Drive when you're back online."