Documentation Asset Manager

Documentation Asset Manager

Purpose

Automates the complete workflow for Tallyfy documentation asset management with Claude Code native vision:

1. Upload assets to Cloudflare R2 storage (screenshots bucket)
2. Generate AI captions using Claude Code native vision (alt text, descriptive, SEO)
3. Maintain CSV inventory at $GITHUB_ROOT/documentation/documentation_assets.csv
4. Build-time integration - Automatic alt text injection during Starlight builds
5. Return public URLs immediately (screenshots.tallyfy.com domain)
6. Track asset usage across documentation files

Recent Updates (v2.0.0)

✅ Completed:

• 287 images captioned with vision-based analysis (100% coverage for native images)
• Claude Code native - No external API key needed
• Build-time integration - Remark plugin auto-injects alt text from CSV
• Parallel Task agents - Batch processing support
• Path portability - Auto-detection, works on any developer machine

Activation Triggers

• "upload documentation asset"
• "upload screenshot to R2"
• "add image to docs"
• "manage documentation media"
• "generate image caption"
• "update documentation screenshot"
• When user provides image/asset path for documentation

Prerequisites

Required Credentials (in parent directory):

• $GITHUB_ROOT/cloudflare_credentials.json with fields:
  • account_id - Cloudflare account ID
  • api_token - Cloudflare API token with R2 access

Required Packages (auto-check):

pip3 install pillow requests

Claude Code Native (Recommended):

• No external API key needed - Uses Claude Code's native vision capabilities
• For batch operations: Use Task agents (see Batch Vision Captioning section)
• For single images: image_captioner.py uses subprocess to call claude native

Instructions

Step 1: Validate Prerequisites

# Check credentials exist
test -f $GITHUB_ROOT/cloudflare_credentials.json || echo "❌ Credentials missing"

# Check Anthropic API key
test -f ~/.anthropic/api_key || echo "⚠️ Set ANTHROPIC_API_KEY or create ~/.anthropic/api_key"

# Check Python packages
python3 -c "import anthropic, PIL, requests" 2>/dev/null || echo "⚠️ Run: pip3 install anthropic pillow requests"

Step 2: Upload Asset to R2

Use the R2 uploader script to upload any file:

# Single file upload
python3 ~/.claude/skills/documentation-asset-manager/scripts/r2_uploader.py \
    --file /path/to/screenshot.png \
    --key "tallyfy/pro/desktop-light-new-feature.png" \
    --credentials $GITHUB_ROOT/cloudflare_credentials.json

# Returns JSON with public URL:
# {
#   "success": true,
#   "url": "https://screenshots.tallyfy.com/tallyfy/pro/desktop-light-new-feature.png",
#   "key": "tallyfy/pro/desktop-light-new-feature.png",
#   "size": "125.4KB"
# }

Step 3: Generate AI Captions

Generate all three caption types using Claude Code native vision:

For Single Images:

# Generate captions using Claude Code native (no API key needed)
python3 ~/.claude/skills/documentation-asset-manager/scripts/image_captioner.py \
    --url "https://screenshots.tallyfy.com/tallyfy/pro/desktop-light-new-feature.png" \
    --output /tmp/captions.json

# Returns:
# {
#   "captions": {
#     "alt_text": "Task assignment interface showing user selection dropdown",
#     "descriptive": "Tallyfy task assignment screen with dropdown menu...",
#     "seo": "Tallyfy task assignment feature showing how to assign tasks..."
#   }
# }

For Batch Operations (Recommended for >10 images):

# Use Task agents for parallel processing
# See "Batch Vision Captioning" section below

Step 4: Update CSV Inventory

Add asset to master inventory CSV:

# Add to inventory
python3 ~/.claude/skills/documentation-asset-manager/scripts/asset_inventory.py \
    --action add \
    --r2-key "tallyfy/pro/desktop-light-new-feature.png" \
    --url "https://screenshots.tallyfy.com/tallyfy/pro/desktop-light-new-feature.png" \
    --captions /tmp/captions.json \
    --article-ids "pro-tasks-assign,pro-getting-started"

Step 5: Complete Workflow (Orchestrator)

For one-command complete workflow:

# Upload + Caption + Inventory in one go
python3 ~/.claude/skills/documentation-asset-manager/scripts/orchestrator.py \
    --file /path/to/screenshot.png \
    --key "tallyfy/pro/desktop-light-new-feature.png" \
    --article-ids "pro-tasks-assign" \
    --credentials $GITHUB_ROOT/cloudflare_credentials.json

# Complete output:
# ✅ Uploaded: https://screenshots.tallyfy.com/tallyfy/pro/desktop-light-new-feature.png
# ✅ Alt text: "Task assignment interface showing user selection dropdown"
# ✅ Added to inventory: $GITHUB_ROOT/documentation/documentation_assets.csv

Build-Time Integration

Automatic Alt Text Injection:

The master CSV is integrated with Starlight documentation builds via a remark plugin:

Location: /support-docs/src/plugins/remark/inject-alt-text.ts

How It Works:

1. During build, remark plugin reads /documentation/documentation_assets.csv
2. Matches image URLs in markdown to CSV entries
3. Auto-injects ai_caption_alt as alt text
4. Handles URL encoding/decoding automatically
5. Falls back gracefully if CSV unavailable

Local Development:

cd /support-docs
# CSV symlinked to ../documentation/documentation_assets.csv
npm run dev  # Alt text injected automatically

CI/CD (Netlify):

# package.json prebuild script copies CSV
npm run prebuild  # Copies CSV before build
npm run build     # Alt text injected during build

Example:

<!-- Before build: -->
![](https://screenshots.tallyfy.com/tallyfy/pro/desktop-light-tasks.png)

<!-- After build: -->
![Task list showing assignees and due dates](https://screenshots.tallyfy.com/tallyfy/pro/desktop-light-tasks.png)

Batch Vision Captioning

For Large-Scale Caption Regeneration:

When regenerating captions for many images (>10), use parallel Task agents for efficiency:

Process:

1. Prepare images - Download and resize to max 1800px (Claude vision constraint)
2. Launch Task agents - Process in parallel batches of ~48 images
3. Consolidate results - Merge all batch outputs
4. Update CSV - Apply new captions to master inventory

Example (287 images completed in ~60 minutes):

# See /temporary/asset-inventory/ for reference implementation:
# - prepare_images_for_vision.py (download + resize)
# - Task agents launched in parallel (6 batches of ~48 images)
# - consolidate_and_update.py (merge results → CSV)

Why Task Agents:

• Native Claude Code vision (no API keys)
• Parallel processing (6x speedup)
• Automatic retry and error handling
• Real-time progress tracking

Workflow Patterns

Pattern 1: Upload New Screenshot

# User: "Upload this screenshot to docs"
# Skill detects image file path, generates reasonable R2 key from filename
# Executes: upload → generate captions → update inventory → return URL

Pattern 2: Replace Existing Screenshot

# User: "Replace screenshot at /tallyfy/pro/desktop-light-tasks.png"
# Skill: uploads with same key (overwrites), regenerates captions, updates inventory

Pattern 3: Batch Upload

# User: "Upload all screenshots in /docs/images/"
# Skill: processes each file sequentially with delays to avoid rate limits

Pattern 4: Generate Caption Only

# User: "Generate caption for existing image at https://screenshots.tallyfy.com/..."
# Skill: skips upload, generates captions using Claude Code native, updates inventory

Pattern 5: Regenerate All Captions (Completed)

# All 287 native images now have vision-based captions (completed 2025-11)
# Location: /documentation/documentation_assets.csv (columns: ai_caption_alt, ai_caption_descriptive, ai_caption_seo)
# Build integration: Auto-injected during Starlight builds via remark plugin

Temporary Assets

Working Directory: $GITHUB_ROOT/temporary/asset-manager/

Files Created:

• upload_result.json - R2 upload response
• captions_alt.json - Alt text caption
• captions_descriptive.json - Descriptive caption
• captions_seo.json - SEO caption
• asset_metadata.json - Combined metadata
• processing.log - Processing log

Cleanup: Temporary files auto-cleaned after successful completion

Master Asset Inventory

Location: $GITHUB_ROOT/documentation/documentation_assets.csv

CSV Columns:

• filename - Asset filename
• r2_key - Full R2 bucket key
• production_url - Public URL (screenshots.tallyfy.com)
• source_type - native|orphaned|external|missing
• file_type - File extension
• file_size - Human-readable size
• file_size_bytes - Size in bytes
• url_exists - Boolean (URL accessible)
• article_ids - Comma-separated article IDs
• article_count - Number of articles referencing
• last_modified - R2 last modified timestamp
• etag - R2 ETag
• ai_caption_alt - Alt text caption
• ai_caption_descriptive - Detailed caption
• ai_caption_seo - SEO-optimized caption
• needs_caption - yes|no (based on file type)

Auto-Update: CSV automatically updated on any upload/delete/caption operation

Error Handling

Scenario 1: Credentials Missing

ERROR: Credentials file not found
ACTION: Check $GITHUB_ROOT/cloudflare_credentials.json exists
PROMPT: "Cannot find Cloudflare credentials. Ensure file exists with:
         - account_id
         - api_token"

Scenario 2: R2 Upload Fails

ERROR: R2 upload failed (API error)
ACTION: Retry once after 5s, then notify user
PROMPT: "Upload to R2 failed: {error_message}
         Retrying once..."

Scenario 3: Caption Generation Fails

ERROR: Claude API error (rate limit, network, etc)
ACTION: Complete upload, skip caption, mark for later
PROMPT: "File uploaded successfully to {url}
         Caption generation failed: {error}
         You can generate captions later with:
         python3 scripts/image_captioner.py --url {url}"

Scenario 4: Rate Limit Hit

ERROR: Anthropic API rate limit
ACTION: Wait 60s, retry once, then defer
PROMPT: "Rate limit reached. Waiting 60s before retry...
         If persistent, use --skip-captions flag to upload without captions"

Scenario 5: Inventory CSV Locked

ERROR: CSV file in use (another process)
ACTION: Wait 2s, retry 3 times, then warn
PROMPT: "Inventory CSV busy. Retried 3 times. Upload succeeded but inventory not updated.
         Run: python3 scripts/asset_inventory.py --sync to update manually"

Advanced Usage

Skip Caption Generation

# For quick uploads without AI captions
python3 scripts/orchestrator.py --file screenshot.png --skip-captions

Regenerate Captions for Existing Assets

# Regenerate all captions for specific asset
python3 scripts/image_captioner.py \
    --url "https://screenshots.tallyfy.com/tallyfy/pro/desktop-light-tasks.png" \
    --update-inventory

Bulk Operations

# Process all images in directory
for img in /docs/screenshots/*.png; do
    python3 scripts/orchestrator.py --file "$img" --batch-mode
    sleep 3  # Delay between uploads
done

Verify Inventory Integrity

# Check all URLs in inventory are accessible
python3 scripts/asset_inventory.py --verify-urls

Generate Statistics

# Get inventory statistics
python3 scripts/asset_inventory.py --stats

# Output:
# Total assets: 812
# Active (used in docs): 290
# Orphaned (not referenced): 499
# Missing (404s): 23
# Images with captions: 288

R2 Bucket Configuration

Bucket Name: screenshots Account ID: ac9046789c9ad74c8704a143b442f918 Public Domain: screenshots.tallyfy.com API Endpoint: https://api.cloudflare.com/client/v4/accounts/{account_id}/r2/buckets/screenshots/objects CDN: Cloudflare global edge cache

Naming Conventions

Recommended R2 Key Patterns:

1. Structured (Preferred):

   tallyfy/pro/desktop-light-{action}-{context}.png
   tallyfy/pro/mobile-dark-{feature}.png
   manufactory/{feature}-{version}.png
   

2. Legacy (Auto-generated):

   file-{random_id}.png
   

Examples:

• tallyfy/pro/desktop-light-assign-task.png
• tallyfy/pro/desktop-light-sidebar-numerated.png
• manufactory-triggers-v2.png

Security Notes

• Credentials stored in parent directory (not in skill)
• No API keys embedded in skill files
• Anthropic API key from environment or user's ~/.anthropic/
• R2 bucket is private with public domain mapping
• All uploads go through Cloudflare API authentication

Troubleshooting

"Module not found" errors:

pip3 install pillow requests

"Credentials not found":

cat $GITHUB_ROOT/cloudflare_credentials.json
# Should show account_id and api_token

"Claude Code not available":

# Ensure Claude Code CLI is installed and accessible
which claude  # Should return path to claude executable

For external API usage (legacy):

# Not recommended - use Claude Code native instead
export ANTHROPIC_API_KEY="your-api-key-here"

"R2 upload succeeds but URL 404":

• Check bucket name is "screenshots"
• Verify custom domain screenshots.tallyfy.com is configured
• Wait 30-60s for CDN cache

"CSV inventory out of sync":

# Rebuild inventory from R2 + documentation
python3 scripts/asset_inventory.py --rebuild

See Also

• references/r2_api_reference.md - Complete Cloudflare R2 API docs
• references/supported_formats.md - Supported file types and limits
• references/examples.md - Usage examples and tutorials
• $GITHUB_ROOT/documentation/CLAUDE.md - Documentation repo guidelines

Maintenance

Regular Tasks:

1. Weekly: Verify inventory CSV integrity
2. Monthly: Review orphaned assets (consider cleanup)
3. Quarterly: Check for missing images (404s) and fix
4. As needed: Regenerate captions for improved quality

Inventory Location:

• Primary: $GITHUB_ROOT/documentation/documentation_assets.csv (version controlled)
• Working Directory: $GITHUB_ROOT/temporary/asset-inventory/asset_inventory_master.csv (generated master)