Accessibility Annotator

Accessibility Annotator

Make technical documents fully understandable by a brilliant reader who lacks computer science training.

Target reader: A top-quintile intelligence mechanical engineer. Sharp, process-oriented, comfortable with measurements and thresholds, experienced with quality systems and structured problem-solving. Has never taken a CS class.

Inputs

Two required arguments (positional or prompted):

1. Word document path -- the .docx file to annotate
2. Project directory path -- the codebase the document describes (used for accurate, contextual explanations)

Optional flags:

• --generate-images -- Generate diagrams via Google Gemini and insert them directly into the document (default: OFF — costs money). When off, insert teal-bordered placeholder boxes with full generation prompts instead.
• --style-json PATH -- Path to the image style JSON for Nano Banana Pro prompt construction (default: $IMAGE_STYLE_JSON env var if set, otherwise provide via --style-json flag. Author's default: ~/dev/brand-assets/clients/cfa/styles/clean-style-sanitized.json)

Parse from the user's message. If missing, ask for the two required arguments.

Phase 1: Analysis

Step 1 -- Read the document

Extract text via pandoc:

pandoc "document.docx" -t markdown

Read the full output carefully. Note the document's voice, tone, and how it already explains concepts.

Step 2 -- Survey the project

Read the project directory structure, README, CLAUDE.md, and 2-3 key source files. Build enough understanding to write accurate explanations later. Don't go deep yet -- that happens in Phase 2.

Step 3 -- Identify opaque concepts

Walk the document paragraph by paragraph. For each technical term or concept, apply the filter below.

The ME WOULD understand (do not flag):

• Process flows, staged pipelines, architecture as "how parts connect"
• Quality gates, thresholds, measurements, calibration, scoring
• Business logic, operational rationale, cost/benefit arguments
• Tables, metrics, percentages, scoring frameworks
• Procedures and step-by-step instructions
• "Database" as "a place that stores and retrieves information"
• Files, records, fields, filtering, sorting
• Statistical concepts: mean, standard deviation, precision, recall, accuracy, confidence, distributions, correlation
• Hardware: GPU, servers, CPU, memory, network, storage
• Acronyms that are expanded and self-explanatory in context

The ME would NOT understand (flag these):

• ML/NLP concepts: embeddings, vector similarity, transformers, tokenization, NER, NLI, cross-encoders, bi-encoders, attention, fine-tuning, inference temperature, context window, token budget, model parameters
• CS data structures and algorithms: inverted indexes, graph traversal, hash maps, B-trees, approximate nearest-neighbor, clustering algorithms (HDBSCAN, DBSCAN)
• Software architecture: REST API, microservices, SSE streaming, CORS, middleware, async processing, event-driven
• Database internals: vector databases, graph databases, Cypher queries, collections, upsert, indexing strategies (IVF, HNSW)
• Data formats beyond basic: JSON/JSONL structure, regex, serialization
• DevOps: Docker, CI/CD, containerization
• Named tools/frameworks with non-obvious purpose: ChromaDB, Neo4j, FAISS, FastAPI, GLiNER, vLLM, pandas, NumPy
• Techniques: TF-IDF, BM25, cosine similarity, Reciprocal Rank Fusion, PCA, anisotropy correction, Platt scaling, embedding dimensions

Critical -- check for self-explanation: If the document already explains a concept well enough for the target reader, do NOT flag it. Many good technical documents include "In plain English" sections or parenthetical clarifications. Only flag concepts that remain opaque after reading surrounding context.

Critical -- flag the concept, not just the term: "Cosine similarity" is a term. The concept is "measuring how similar two things are by comparing their numeric representations." Flag the concept only if the document doesn't convey this.

Step 4 -- Map dependencies

Some concepts require others. Build a dependency list:

• "Cross-encoder reranking" needs "embeddings" explained first
• "Cosine similarity" needs "vector representation" explained first
• "Knowledge graph traversal" needs "graph database" explained first
• "BM25" needs "inverted index" or at minimum "keyword matching" explained first

Record dependencies by item number so the analysis table shows the chain.

Step 5 -- Assess section density

For each document section, count flagged concepts per paragraph:

• Low (0-1 per paragraph): inline treatment works
• Medium (2-3 per paragraph): mix of inline and footnotes
• High (4+ per paragraph): glossary/appendix required -- inline would destroy readability

Step 6 -- Select explanation mechanism

For each flagged concept, select ONE mechanism using these heuristics:

Density override: If a section has 4+ flagged concepts per paragraph, do NOT use inline parentheticals in that section. Escalate all items to glossary entries, with footnotes or sidebars for foundational concepts.

Dependency rule: If Concept B depends on Concept A, and Concept A is assigned to the glossary, Concept A must appear before Concept B in the glossary. If Concept A gets a sidebar, place it before the first section that uses Concept B.

Step 7 -- Present the analysis

Output a markdown table with these columns:

• #: Sequential number
• Concept/Term: The flagged term or concept
• Location: Section name and approximate position (e.g., "Executive Summary, para 3")
• Why opaque: One sentence on why the ME wouldn't understand this
• Depends on: References to other items by # (e.g., "#3, #7") or "None"
• Mechanism: One of: inline, footnote, sidebar, glossary, appendix-section, web-link
• Draft explanation: The proposed explanation text at target depth

Order rows by appearance in the document.

After the table, include:

Density summary: List each document section with its density rating (low/medium/high).

Dependency map: Show chains (e.g., "#3 vector representation -> #7 cosine similarity -> #12 cross-encoder reranking").

Mechanism distribution: Counts per mechanism type plus a sentence on whether the distribution feels right.

Appendix sections recommended: If any, describe the unified explainer topics.

Then ask:

Review the analysis above. You can approve as-is, adjust items (change mechanisms, edit explanations, add/remove items), or ask questions. When ready, say implement and I'll modify the document.

Phase 2: Implementation

After user approves or says "implement":

Step 1 -- Deep-read the project

For each concept that needs explanation, read the relevant source code, configs, or docs in the project directory. Every explanation must be accurate to what the system actually does -- not a generic textbook definition.

Step 2 -- Write final explanations

For each item:

• Match the document's existing voice and tone. If the document is direct and conversational, annotations should be too. If it's formal, match that.
• Target depth: "A standard way software systems talk to each other over the internet" -- enough to follow the argument, not enough to implement the concept.
• Use analogies from the ME's world when they fit naturally: manufacturing processes, measurement instruments, inspection systems, assembly lines, material testing, quality control. Don't force analogies -- only use them when they genuinely clarify.
• Glossary ordering: By dependency (foundational first), NOT alphabetical. Group related concepts when it aids comprehension.

Step 3 -- Modify the Word document

Use the docx skill workflow. Invoke the document-skills:docx skill for the actual document modification.

What to add:

1. Inline parentheticals -- insert directly after the term in the text. CRITICAL: annotate at FIRST USE, not just in the glossary. A reader encountering "cosine similarity" on page 4 cannot be expected to know it's defined in a glossary on page 22. Add a brief parenthetical like "(a 0-to-1 measure of meaning similarity -- see glossary)" at the first occurrence in the body text. Every glossary term that appears in the body must have at minimum a brief inline definition or "(see glossary)" pointer at its first use.
2. Footnotes -- use Word's native footnote fields (FootnoteReferenceRun)
3. Sidebars -- styled bordered paragraphs or callout boxes with a distinct background. Use CFA Red (#DD0031) left border, #F5F5F5 background.
4. Glossary appendix -- new section at end titled "Technical Glossary" with dependency-ordered entries
5. Appendix explainer sections -- new sections at end, before the glossary, with narrative explanations
6. Web links -- hyperlinks with descriptive link text
7. Developer section labels -- insert italic "Technical Reference" labels before sections that contain implementation details aimed at developers rather than the target reader.

Clean up developer jargon: Replace code-level terms (class names like AsyncLLMProcessor, GradientBoostingClassifier, FormalRule) with plain-language equivalents. Replace technical shorthand ("SQLite-based LLM cache stores responses keyed on the prompt hash") with accessible language ("a local database caches previous results so identical requests skip the LLM entirely").

Remove mathematical formulas: Replace formulas like P(correct) = sigmoid(a * raw_confidence + b) with plain-language descriptions of what the formula accomplishes.

Preserve everything: All existing formatting, styles, headers, footers, images, tables. Additions must feel native to the document.

Output file: Save as {original-name}-annotated.docx alongside the original. Never overwrite the original.

Step 4 -- Images (3-5 per document)

Identify 3-5 locations where a generated diagram or illustration would most help the non-CS reader. Prioritize:

• System architecture / data flow diagrams (especially replacing existing ASCII-art or text-based diagrams)
• Conceptual visualizations of foundational techniques (how vector search works, how a knowledge graph looks)
• Process flow diagrams showing how data moves through stages
• Dashboard/UI mockups when the document describes interactive tools

For each image location, write a detailed content-specific description and a full generation prompt incorporating the style JSON from the --style-json flag value (or $IMAGE_STYLE_JSON env var).

If --generate-images is ON:

1. Load the Google API key from Bitwarden: bws secret get 74f9abaf-d730-4c4e-b689-b3df01843d65
2. Use google-genai SDK with model gemini-3-pro-image-preview (verify with provider if errors occur)
3. Dispatch one context: fork subagent per image (max 3 concurrent — see concurrency note above)
4. Each subagent: generate image with response_modalities=["IMAGE"], aspect ratio 16:9, 300s timeout; save PNG to word/media/<filename>.png
5. After all subagents complete, parent skill inserts each image into the document:
   • Add relationship entry to word/_rels/document.xml.rels
   • Add content type for PNG to [Content_Types].xml if not present
   • Insert image XML using wp:anchor with <wp:wrapSquare wrapText="bothSides"/> (NOT wp:inline)

If --generate-images is OFF (default):

1. Insert teal-bordered (#3EB1C8) placeholder callout boxes at each image location
2. Each placeholder contains the full generation prompt (copy-paste ready for Nano Banana Pro)
3. Label: "Image Placeholder: [Title]" in teal bold, 9pt
4. Dimensions: "Landscape 16:9 -- 3200 x 1800px" in grey, 8pt
5. Content description in grey, 8pt; full prompt in grey, 7pt

If replacing an existing text-based diagram (ASCII art, blockquote flowcharts), remove the original.

Step 5 -- Report

After modification, report:

• Count of each mechanism type added
• Count of images generated or placeholders inserted, with titles
• Count of inline first-use annotations added
• Count of developer section labels added
• Count of jargon replacements made
• List of appendix sections added
• Any concepts where the project directory didn't have enough context (and what you did instead)

Depth Calibration

Too shallow: "Vector search (a type of search)" -- useless, tells the reader nothing.

Right depth: "Vector search finds information by meaning rather than keywords. The system converts each piece of text into a numeric fingerprint that captures what it means, then finds other texts with the most similar fingerprints -- like identifying matching paint formulas by comparing their chemical compositions rather than their brand names."

Too deep: "Vector search works by encoding text into high-dimensional dense representations using transformer-based neural networks, then computing cosine similarity in the embedding space..." -- this IS the CS training the reader doesn't have.

The test: after reading your explanation, could the ME follow the document's argument about WHY this technique is used and WHAT it accomplishes? They don't need to know HOW to build it.

What NOT to annotate

• Concepts the document already explains adequately for the target reader
• Business and operational concepts (even if they sound technical)
• Proper nouns that are just names -- only annotate their function, and only if unclear from context
• Statistical concepts an ME would know
• Acronyms that are expanded and self-explanatory
• Concepts where the document's own "why it matters" section makes the concept clear enough

Lessons Learned

1. Always annotate at first use, not just in the glossary. The glossary may be 20+ pages away. Every glossary term needs at minimum a brief parenthetical or "(see glossary)" at its first appearance in body text.

2. Replace developer jargon, don't just annotate it. Code-level class names (AsyncLLMProcessor, GradientBoostingClassifier) should be rewritten as plain language, not parenthetically defined. The reader gains nothing from knowing the class name.

3. Remove mathematical formulas. Replace with plain-language descriptions of what the formula accomplishes. "P(correct) = sigmoid(a * raw_confidence + b)" becomes "the system recalibrates raw confidence scores against actual outcomes."

4. Label developer sections explicitly. Sections containing port numbers, concurrency tuning, caching internals, and configuration parameters should be labeled "Technical Reference" in italic text so the target reader knows they can skip without losing the conceptual thread.

5. Each document must be self-contained. If documents share concepts (embeddings, LLM, etc.), each document needs its own sidebar explanations. Don't assume the reader has seen another document.

6. Sidebar selection matters per document. Choose sidebars for the 2-3 foundational concepts that THIS document relies on most heavily. A document focused on duplicate detection needs TF-IDF and embedding sidebars. A document focused on knowledge extraction needs LLM and embedding sidebars.

7. Tables with many columns need fixed layout in Word XML to prevent column truncation. Add <w:tblLayout w:type="fixed"/> to <w:tblPr> for any table with 4+ columns.

8. Physical-world analogies work best. Paint samples for embeddings, ingredient lists for TF-IDF, book indexes for inverted indexes, quality inspection stages for cross-encoder reranking, filing systems for FAISS. The ME thinks in physical processes.

9. Dashboard/UI descriptions need mockup images. When a document spends multiple pages describing an interactive dashboard, include at least 1-2 mockup images so the reader can visualize what's being described.

10. The "Why it matters" / "Why Not Just..." pattern is the strongest rhetorical device for this audience. These sections earn trust by explaining trade-offs through consequences. Support and extend them, don't compete with them.

Error Handling

• Missing document path or project directory: ask for both required arguments before starting Phase 1 (see Inputs) rather than assuming defaults.
• pandoc extraction fails or the Word doc is corrupt/unreadable: report the error and stop before Phase 1 analysis — do not fabricate document content.
• --generate-images is ON but the Bitwarden secret fetch or the Gemini call fails (bad key, rate limit, timeout): fall back to a teal-bordered placeholder box for that image, and note the fallback in the Step 5 report.
• A per-image subagent still fails after its 3 retries (8-second backoff): insert the placeholder for that image instead of leaving a gap, and list which images fell back in the report.
• Project directory doesn't have enough context to write an accurate explanation for a flagged concept: use a more general but still accurate explanation, and note the gap explicitly in the Step 5 report.
• User rejects or requests changes to the Phase 1 analysis table: revise and re-present it — do not proceed to Phase 2 implementation without approval.