Context Lab

Start here · Why Markdown wins

What actually happens when a model reads your document — and why the input format decides the answer quality.

• 01

  LLMs don't just read Markdown — they speak it

  Ask a model anything: it answers in headings and code fences. Feeding its native register back matters.

  .md · 7 min

• 02

  Why PDFs are hostile input for LLMs

  PDFs store painted glyphs, not structure. What silently breaks when you paste one into a model.

  .pdf · 6 min

Format guides

The reference shelf — what's really inside each file type and the exact element mapping applied. All thirteen guides →

.ipynb .docx .pptx .xlsx .csv .pdf .html .json .srt .tex .eml .mbox .md

Under the hood

How the converter earns trust: structure detection element by element, and a fidelity report that never says "preserved".

• 01

  What a converter built for LLMs does differently

  Headings from styles, typed tables, figure placeholders — and a report of what was detected, not promised.

  philosophy · 8 min

• 02

  Making Jupyter notebooks LLM-addressable

  Cell addresses, execution-order confessions, dependency hints — a build log with real numbers.

  .ipynb · 9 min

• 03

  Tables LLMs can actually read

  Column types, explicit truncation, serial dates, cached formulas — spreadsheets without invented data.

  .xlsx .csv · 7 min

• 04

  Which output preset should I use?

  Pick by destination, not by file — the 30-second decision table with live sample links.

  how-to · 4 min

Real-world workflows

Measured, story-driven guides: what changes in Claude Projects, Perplexity Spaces and other AI workspaces when the uploads are Markdown.

• 01

  Case study: lecture notes that made a smaller model smarter

  Same workspace, same questions — the only change was PDF → Markdown. A first-person report with numbers.

  case study · 7 min

• 02

  Markdown is the best upload format for AI workspaces

  Storage caps, token budgets, retrieval accuracy — three bottlenecks, one format change, measured.

  workflow · 8 min

• 03

  Meeting recordings → minutes your model can quote

  The caption track is the transcript — minutes with [mm:ss] receipts and a searchable archive.

  .srt workflow · 5 min

Choosing tools

Honest comparisons — including the cases where the other tool wins.

• 01

  vs pandoc

  When you want the universal CLI, when you want the fidelity report.

  comparison · 5 min

• 02

  vs online converter sites

  "Upload" means your file lands on their servers — and the difference is checkable here.

  comparison · 4 min

• 03

  vs just attaching the file

  Platforms convert attachments with a hidden extractor. Converting first inverts the control.

  comparison · 5 min

• 04

  vs cloud parsing APIs

  They bring OCR at per-page prices; the local default handles the born-digital majority.

  comparison · 5 min

Advanced

For people building pipelines: retrieval systems, agent context, automation.

• 01

  RAG-ready Markdown: chunks, anchors, honest citations

  Most RAG quality problems are chunking problems. Fix retrieval at the source, before embedding.

  pipelines · 8 min

• 02

  Giving a coding agent readable context

  Agents re-read documents dozens of times — addresses, grep-ability and honest gaps compound.

  agents · 6 min

• 03

  Obsidian as an LLM knowledge base

  Callouts, wikilinked figures and front matter — documents becoming vault citizens.

  Obsidian · 6 min

• 04

  Token budgeting 101

  The estimate lies by ~20% on technical text — how to count, and where document tokens hide.

  tokens · 6 min

• 05

  Build an LLM-readable wiki

  Archive-preset Markdown in git: humans browse it, models retrieve it, diffs audit it.

  knowledge base · 6 min

• 06

  Roadmap: a CLI, agent skills, stronger local OCR

  Same parsers, new surfaces — terminal, agents, and model-based OCR on your CPU/GPU. Plans, not promises.

  roadmap · 5 min

FAQ · Fixes for common errors

Hit an error? Find your symptom below, understand the mechanism, apply the two-minute fix — the converter's own error messages link straight here.

• 01

  Claude Projects can't read your PDF?

  Refusals, misquotes, wrong-file hunts — why PDF retrieval fails and the two-minute fix.

  workspaces · 4 min

• 02

  "I can't access this file"

  The four real causes behind every upload error, and how to tell them apart.

  uploads · 5 min

• 03

  Your notebook is too big to paste

  3.5 MB of .ipynb is mostly base64 — beating the token limit with cell addresses intact.

  .ipynb · 5 min

• 04

  Pasted a spreadsheet, got garbage

  Whitespace collapse, serial dates, merged cells — and the table format that survives.

  .xlsx .csv · 4 min

• 05

  The model invents numbers from your data

  Silent truncation and headerless values cause most of it — both fixable before the prompt.

  data honesty · 4 min

• 06

  Scanned PDFs — and the OCR fallback

  Pictures of text contain no text — the diagnosis, your options, and the opt-in local OCR.

  .pdf · 4 min

• 07

  Copying from a webpage floods the chat with junk

  Menus, banners and teasers ride along — article extraction keeps only the article.

  .html · 4 min

• 08

  Your workspace keeps missing half a document

  Three retrieval failure modes — bad chunks, duplicate copies, ambiguous filenames.

  workspaces · 5 min