Shipping the work — explore in HTML, deliver in PDF

Once Claude has done the analysis, you need to do one of two things with it: explore it, or deliver it. Each has a format that's almost embarrassingly good at the job — and Claude produces both for free.

The whole idea in one line: a self-contained HTML file is the best thing ever invented for poking at data; Markdown → Pandoc → PDF is the best cheap way to hand someone a document that looks like it came from a design team. This page sets up both. It's a reference for the report problems — do it once, reuse it everywhere.

Two outputs, two jobs

	HTML explorer	Pandoc PDF
For	Thinking. Slicing, filtering, zooming, toggling assumptions.	Delivering. The frozen artifact for the boss, the customer, the design history file.
Audience	You and a colleague or two.	People who will print it, sign it, or file it.
Interactive	Yes — live charts, sortable tables, sliders.	No — static, paginated, permanent.
Dependencies	None. Any browser opens it.	`pandoc` + one PDF engine (below).
Shareable as	One `.html` file you can email.	One `.pdf` file you can email.

The interactive panels on the spend report and decision point problems are HTML explorers — single files, no install, live charts. That's path one. This page mostly exists to set up path two.

Path 1 — the HTML explorer (zero install)

This one needs no setup at all, so the entire skill is knowing to ask for it. The magic words are “single self-contained HTML file, no external dependencies” — that tells Claude to inline the data, the styling, and any chart code into one file that works offline, forever, on any machine.

PromptBuild me a single self-contained HTML file that loads this CSV (embed the data directly), shows a sortable table, lets me filter by category and vendor, and renders spend-by-category as a bar chart. No external libraries or CDN links — everything inline so it works offline.

Double-click the result. It opens in your browser. Nothing to install, nothing to host, no build step. When you want a different cut of the data, you ask Claude to change it and re-open the file. This is the fastest data-exploration loop there is, and it costs nothing to run or share.

Why inline matters: a file that pulls a charting library from a CDN breaks the day you open it on a plane, or in two years when that CDN URL rots. “Self-contained” means the file is the whole program. Keep one around and it keeps working.

Path 2 — Markdown → Pandoc → PDF

You write (or Claude writes) the report in Markdown — plain text with # headings and | tables. Pandoc is the universal document converter that turns it into a typeset PDF. Pandoc needs one helper to actually draw the PDF; we use Typst, a small modern typesetting engine, because it's tiny, fast, and has readable error messages.

brew install pandoc typst

Terminal (Homebrew)

No Homebrew? Install it first from brew.sh, or download the pandoc and typst installers from their release pages.

winget install --id JohnMacFarlane.Pandoc
winget install --id Typst.Typst

PowerShell

Close and reopen PowerShell afterward so both land on your PATH.

# Debian/Ubuntu — but check the version (see note)
sudo apt install pandoc
# Typst: download the release binary, or:
cargo install typst-cli

bash

Distro pandoc is often too old for the Typst engine (you need pandoc ≥ 3.1). Check with pandoc -v. If it's older, grab the current .deb from pandoc's GitHub releases instead of apt.

Render it — one command

That's the whole workflow. Edit the Markdown, re-run the command, get an updated PDF. Better still, don't memorize the flags — let Claude write the report and hand you the exact command:

PromptWrite a one-page Markdown report of these findings: a two-sentence summary, a findings table, and an embedded bar-chart image. Add a YAML title block. Then give me the exact pandoc command to render it to a nicely-typeset PDF using the Typst engine.

Making it look good

Pandoc + Typst already produces a clean document. A YAML block at the top of the Markdown adds a title page and metadata; a couple of flags add a table of contents and section numbers:

---
title: "Q1 Purchase-Order Spend"
author: "Reliability Engineering"
date: "2026-04-15"
---

# Summary

Total spend was **$67,414** across 49 line items...

![Spend by category](spend_by_category.svg)

| Category | Spend | Share |
|----------|------:|------:|
| Machined Parts | $12,908 | 30% |

report.md

pandoc report.md -o report.pdf --pdf-engine=typst --toc --number-sections

Terminal

Charts in a PDF are images. The PDF is static, so a figure is a file you reference — ![caption](chart.svg). Have Claude emit the chart as an .svg (crisp at any size) or a small Python/matplotlib script that writes one. The live, draggable charts are the HTML path's superpower; the PDF gets a sharp snapshot.

Prefer LaTeX? If you want LaTeX-grade typography or already have a TeX install, swap one tool: install TinyTeX (a minimal, self-managing LaTeX — the installer lives at yihui.org/tinytex) and change the flag to --pdf-engine=xelatex. Everything else — the Markdown, the YAML block, the workflow — is identical. Full TeX Live works too but it's ~4 GB; TinyTeX is a few hundred MB and installs missing packages on demand.

Where this runs

Both paths run on your own machine, with your own Claude — same as everything else on this site. The HTML file opens in your browser; the pandoc command runs in your terminal. Nothing is uploaded, nothing is hosted, and the artifacts are yours. Claude can even write a one-line Makefile or shell script so “rebuild the report” becomes a single command you never have to think about again.

Take it back to the datasets

Both report problems ship a downloadable dataset and end with a “Ship it” section. Generate the data, do the analysis with Claude, then practice both deliveries: