16 At Work and in a Start-up
Quarto, reproducible publishing, coding agents, Markdown, Python, EPUB, data analysis, k-means clustering
Everything so far has used public data on a computer you control. Work is rarely like that. The data are confidential, the tools must be approved, the network sits behind a proxy, and the documents must look like the organisation’s. None of this changes the method; it changes what you may send to an agent, which agent, how its traffic leaves the building, and which templates the outputs wear. This chapter is the short, practical version, and it ends with the case the book has followed throughout: a small company with investors.
- Say what leaves your computer when an agent works, and what each kind of plan does with it.
- Work on confidential data without sending it to an unapproved tool.
- Make the tools work behind a proxy, and know what a locked-down machine allows.
- Fit outputs to an organisation’s templates and brand, and decide what a start-up should and should not build this way.
16.1 What leaves your computer
An agent’s model runs on the vendor’s servers (Section 2.1). So whatever the agent reads, a file, a command’s output, an error message with a customer’s name in it, goes there with your prompt. Quarto, Python and Git send nothing anywhere; the agent is the one part of the method that does. What happens to it then depends on the plan, not the product:
| Plan | What the vendor says |
|---|---|
| Claude, consumer | On Anthropic’s Free, Pro, Max and Ultra plans, Claude Code included, conversations are used to train new models when the user’s privacy setting allows it, and are kept for five years when it does, thirty days when it does not. |
| Claude, commercial | Under Anthropic’s commercial terms (Team, Enterprise, the API), code and prompts are not used for training unless the customer opts in, and are kept thirty days by default; zero data retention is available to qualified Enterprise accounts. |
| OpenAI API | Data sent to OpenAI’s API is not used for training unless you opt in; abuse-monitoring logs are kept up to thirty days, and zero data retention needs OpenAI’s approval. |
| ChatGPT business | ChatGPT Business, Enterprise and Edu workspace data is not used to train OpenAI’s models by default. |
| Codex | How you sign in to Codex, with a ChatGPT plan or an API key, decides which data policy applies. |
| Antigravity | Antigravity’s terms say Google uses interactions to improve its products, its employees and contractors may review them, and users can opt out in settings; they ask users not to submit confidential information. |
Two facts follow. A personal plan and a company plan of the same agent are different arrangements, so ask which one you are on before the first confidential file. And the copy on your own machine counts too: Claude Code keeps session transcripts in plain text under ~/.claude/projects/ for thirty days by default, which matters on a shared or managed computer.
16.2 Confidential data
Customer records, payroll, unpublished results, patient or student data, a start-up’s revenue and its investors’ names are not yours to send anywhere you like. Four habits keep the method safe with them:
- Read the policy first. Your organisation’s AI and data policy says which tools and plans may see which data. If it is silent, ask whoever owns the data, in writing.
- Build on a stand-in. Make a synthetic file with the same columns and types as the real one, labelled SYNTHETIC in its name and README, as the feedback sheets of Section 14.5 do. Let the agent build and test every document against it.
- Render the real thing yourself. Swap in the real file, without the agent, and render. The agent never reads it, never sees its values in an error, and never puts them in a transcript.
- Keep it out of the history and off public hosts. List the real file in
.gitignore, keep the repository private, and publish where your organisation approves: a free GitHub account publishes Pages from public repositories only; private repositories need a paid plan, and a site only your organisation can see needs GitHub Enterprise Cloud (checked 2026-11-18).
“Do not paste it” is not enough. An agent can read a file you never mentioned, if a command it runs prints it, or a test it writes loads it. Keep real confidential files out of the agent’s folder altogether until the agent’s work is done, and add a rule to AGENTS.md that names the file and forbids reading it, as a reminder that is not a guarantee.
16.3 Approved tools and locked-down machines
Many organisations approve one agent, on a company plan, and manage its settings centrally. Claude Code, for one, reads a managed settings file that an administrator installs and that no user setting can override; it can forbid commands, such as pushing or downloading, outright. Ask what your organisation has set before you wonder why a command is refused.
A machine where you cannot install software at all still has a route: the browser-based environment of Section 4.8, where your organisation allows GitHub, or a development server your IT team provides. The method needs only a terminal, Quarto, uv, Git and an approved agent; where any of them is missing, say so to IT in those words, because “I need Quarto 1.10 and uv” is a request they can act on.
16.4 Behind a proxy
Company networks often send traffic through a proxy (Chapter 4), and sometimes inspect encrypted traffic with the company’s own certificate. Both break downloads in ways that look like the tools’ fault. The settings that fix them are environment variables, set once in your shell:
| Tool | Proxy | Company certificate |
|---|---|---|
| uv (Python and packages) | HTTPS_PROXY, HTTP_PROXY, NO_PROXY |
UV_SYSTEM_CERTS=1 to use the system’s store, or SSL_CERT_FILE naming a bundle |
| Claude Code | HTTPS_PROXY, HTTP_PROXY, NO_PROXY; no SOCKS proxies |
the system’s store, or NODE_EXTRA_CA_CERTS naming a bundle |
Quarto’s own documentation says nothing about proxies (checked 2026-11-18); Quarto itself needs the network only to install, and the Python it runs comes through uv. If a download still fails, ask IT to allow the hosts the tools reach, which Claude Code’s network page lists, rather than turning certificate checks off.
16.5 House templates and brand rules
An organisation’s Word and PowerPoint templates become reference documents (Section 11.6, Section 12.3): ask for the .dotx or .potx the company uses, save it as a .docx or .pptx, and check that its PowerPoint layouts carry the names Quarto looks for. Its brand guidelines become _brand.yml (Section 12.2): the colours, the fonts the company licenses, and the logo with its alt text. A legal footer or a confidentiality notice becomes a page footer or a Typst partial (Section 13.5), written once. And an accessibility standard, which many organisations have adopted, becomes a check that runs on every page (Section 12.5).
16.6 In a start-up
A start-up has the same constraints with fewer people, and one more reader: the investor. The method pays where numbers change and readers check them: the monthly investor update (Chapter 14), the benchmark or evaluation white paper (Chapter 13), per-customer reports, and a documentation site. It pays less where design or co-editing dominates, and those can stay in the tools made for them (Section 1.5).
Investors’ data needs the care of any confidential data: revenue, the cap table and customers’ names belong in a private repository, built by the agent on synthetic stand-ins and rendered for real by you, and sent as a PDF or through a data room, not published. The habit that matters most is the cheapest: every number in an update is computed and every sentence about it guarded, so the wrong figure that would cost a founder credibility stops the render instead of reaching an inbox.
Summary
- The agent is the one part of the method that sends anything off your computer, and the plan, not the product, decides what the vendor does with it.
- Build on a synthetic stand-in with the same columns, and render the real data yourself.
- Keep confidential files out of the agent’s folder, out of Git and off public hosts.
- Behind a proxy, set the proxy and certificate variables for uv and the agent; do not turn checks off.
- An organisation’s templates, brand, footers and accessibility rules each have a home in the project: reference documents,
_brand.yml, partials and checks.
Further reading
- Anthropic, Claude Code data usage, network configuration and managed settings (checked 2026-11-18).
- OpenAI, Your data (checked 2026-11-18).
- Astral, uv environment variables and certificates (checked 2026-11-18).
- GitHub, plans (checked 2026-11-18).