pdf-simplifier

PDF-to-Web-Reader Skill

Converts a PDF into a self-contained static web reader with three reading modes (ELI5, Medium, Full), sidebar navigation, KaTeX math rendering, and extracted images.

Pipeline Overview

Stage	What	Tool	Output
1	Extract text + fonts	scripts/extract_text.py	.state/pages.json
2	Extract images	scripts/extract_images.py	images/ + .state/images.json
3	Detect structure	scripts/detect_structure.py	.state/structure.json
4	Build content shell	scripts/build_content_shell.py	.state/content-shell.json + .state/sections/*.txt
5	Generate content	Claude (subagents)	.state/sections/*.json
6	Assemble site	scripts/validate_sections.py + scripts/assemble_content.py + scripts/render_template.py	content.json + index.html + styles.css + app.js

Quick Start

1. Get Inputs

The PDF path is provided as the skill argument (pdf_path). Verify the file exists:

test -f "{pdf_path}" && echo "OK" || echo "NOT FOUND"

If the file doesn't exist, report the error and stop.

Ask the user where they want the output directory. Do not assume a default — wait for their answer.

Optionally, ask if they have a manual structure file (a structure.json) they'd like to use.

2. Check Dependencies

python3 -c "import fitz; print(fitz.version)" 2>/dev/null || echo "MISSING"

If pymupdf is missing, ask the user:

pymupdf is required for PDF extraction. Install it with pip install pymupdf?

3. Check for Resumability

If the output directory already exists, check .state/progress.json:

cat {output}/.state/progress.json 2>/dev/null

If it exists, report what's already done and offer to resume or start fresh.

Stage Execution

Stage 1: Extract Text

python3 {skill_dir}/scripts/extract_text.py "{pdf_path}" "{output_dir}"

Read the output JSON to confirm page count and outline entries. Report to user:

Extracted text from {N} pages. Found {M} outline/bookmark entries.

Stage 2: Extract Images

python3 {skill_dir}/scripts/extract_images.py "{pdf_path}" "{output_dir}"

Report:

Extracted {N} images.

Stage 3: Detect Structure

If the user provided a manual structure file, copy it to .state/structure.json instead of running detection:

cp "{manual_structure_path}" "{output_dir}/.state/structure.json"

Otherwise, run detection:

python3 {skill_dir}/scripts/detect_structure.py "{output_dir}/.state/pages.json" "{output_dir}"

USER INTERACTION POINT: Display the detected structure and ask for confirmation:

Detected structure (method: {method}):

Title: {title}

Chapter	Sections	Pages
{ch.title}	{len(ch.sections)}	{first_page}-{last_page}
...	...	...

Does this look correct? You can:

1. Proceed with this structure
2. Provide a manual structure file (see format with /pdf-simplifier help structure)
3. Ask me to adjust specific chapters/sections

If the user wants adjustments, read references/manual-structure-format.md for the schema and help them create one.

Stage 4: Build Content Shell

python3 {skill_dir}/scripts/build_content_shell.py "{output_dir}"

Report:

Built content shell: {N} chapters, {M} sections. Math detected: {yes/no}.

Stage 5: Generate Content

This is the core stage where Claude generates the three reading modes.

USER INTERACTION POINT: Before starting, report:

Ready to generate content for {N} sections. This will use subagents to process sections in parallel. Proceed?

Content Generation Loop

1. Read .state/progress.json to find which sections are already complete
2. Read .state/content-shell.json to get the full section list
3. For each incomplete section, dispatch a subagent

Subagent dispatch - process 3-5 sections concurrently:

For each section, spawn an Agent with subagent_type: "general-purpose" and model: "sonnet".

Read the prompt template from references/subagent-prompt.md and fill in the variables:

• {skill_dir} — path to this skill's directory
• {section.id}, {section.title}, {section.startPage}, {section.endPage} — from the content shell
• {hasMath} — from .state/content-shell.json meta
• {output_dir} — the user's chosen output directory

After each batch of subagents completes, update .state/progress.json yourself (do not have subagents write to this file — concurrent writes will corrupt it):

python3 -c "
import json
p = json.load(open('{output_dir}/.state/progress.json'))
p['completedSections'].extend([{list of completed section IDs from this batch}])
p['stage'] = 'generating'
json.dump(p, open('{output_dir}/.state/progress.json', 'w'))
"

Report progress periodically:

Generated {completed}/{total} sections...

Stage 6: Assemble Site

Once all sections are generated:

1. Validate sections:

   python3 {skill_dir}/scripts/validate_sections.py "{output_dir}"
   

   If errors are reported, log them. Retry failed sections or proceed with what's available.

2. Assemble content.json:

   python3 {skill_dir}/scripts/assemble_content.py "{output_dir}"
   

3. Render template: Read .state/content-shell.json to get the title and meta.hasMath flag, then:

   python3 {skill_dir}/scripts/render_template.py \
     "{skill_dir}/assets/template" \
     "{output_dir}" \
     "{title}" \
     "{has_math}"
   

   This copies template files, safely substitutes the title, and bundles KaTeX locally when math is detected (falls back to CDN if download fails).

4. Update progress:

   python3 -c "
   import json
   p = json.load(open('{output_dir}/.state/progress.json'))
   p['stage'] = 'complete'
   json.dump(p, open('{output_dir}/.state/progress.json', 'w'))
   "
   

5. Report completion:

   Reader generated at {output_dir}/

   To view it:

   cd {output_dir} && python3 -m http.server 8000
   

   Then open http://localhost:8000

Resumability

When invoked on an existing output directory:

1. Check .state/progress.json
2. If stage is "complete" → offer to regenerate specific sections or start fresh
3. If stage is "generating" → find incomplete sections and resume from there
4. If stage is "content-shell-built" → resume from Stage 5
5. If .state/pages.json exists but no structure.json → resume from Stage 3
6. If .state/pages.json exists and structure.json exists → resume from Stage 4

Notes

• The {skill_dir} placeholder refers to the directory containing this SKILL.md file. Resolve it relative to the skill's location in the plugin.
• All Python scripts are self-contained and only require pymupdf (pip install pymupdf).
• KaTeX is bundled locally into the output when math is detected, making the reader work offline. Falls back to CDN if the download fails.
• For very large documents (100+ sections), consider increasing subagent parallelism to 5-8.
• If a subagent fails on a section, log the error and continue. The section can be retried by re-running the skill.