Start Here: Five Things That Make Everything Else Make Sense

A 10-minute read for self-directed readers. If you've just arrived at this site and don't know where to begin — begin here.

This is not a tutorial. It is a conceptual briefing: four ideas that are prerequisites for understanding everything else in these materials. In the workshop these are covered in 12 minutes before any demos. If you're reading without a presenter, this is their equivalent.

You don't need to have Claude Code installed. You don't need to have used Claude at all. You just need to be a researcher who does document-heavy work and wants to understand what this tool actually is before deciding whether to use it.

1. Claude works inside your folder — not inside a conversation

The most important difference between Claude Desktop and Claude Code is not the interface. It is the workspace model.

In Claude Desktop, you work in a conversation. You paste text, upload a file, get a response. When the session ends, Claude forgets everything. Next time: start over. If you want Claude to read a different file, you upload it again. If you have forty files to process, you upload forty times. This is what we call the upload dance — and it is the main friction point for researchers with large document collections.

In Claude Code, Claude has access to a folder on your computer. Not a copy you uploaded — the actual folder, on your disk. Claude can read any file in it, write new files, modify existing ones, and loop through an entire collection in a single prompt. You don't upload; you just point.

The folder is the workspace. This shift sounds simple, and it is — but its implications are substantial. Processing forty PDFs, extracting structured data across an entire corpus, updating a reading log file after every paper: these become single-step tasks rather than forty-step manual processes.

→ What this solves in practice: A.issue.upload-dance → The full comparison of Desktop / Cowork / Code: A1.desktop-vs-code

2. Markdown is the bridge between your documents and Claude

Claude works natively in text. Your research documents mostly live in PDF and DOCX. These two facts create a gap — and markdown is what closes it.

Markdown is a plain-text format (files ending in .md) that preserves structure — headings, lists, tables, emphasis — without proprietary formatting. Claude reads it perfectly. It also writes it perfectly: when Claude drafts something for you, it comes back as markdown, which opens cleanly in any text editor, looks beautiful in Obsidian or Zotero, and can be converted back to DOCX or PDF when you need it.

The practical workflow is: bring your documents in as markdown, get your results out as markdown.

A tool called markitdown converts PDF and DOCX files to .md with a single command. After conversion, a 300-page scanned archive becomes a text file Claude can quote from, summarise by section, extract structured data from, or compare against other documents in the same folder.

This is why markdown appears throughout these materials as the hub format for AI-assisted research work. It is not an aesthetic choice. It is the technical layer that makes the folder workspace model work for researchers who don't write code.

→ Conversion tools and the full markdown workflow: A.markdown-central → Markdown as project memory: A9.markdown-project-memory

3. CLAUDE.md is the memory Claude doesn't have

Claude has no persistent memory between sessions. Every time you start a new session, Claude starts from nothing. It doesn't know who you are, what your project is, what you've already done, or what conventions you've established. In Claude Desktop, this means every session begins with re-orientation.

Claude Code has a solution: a file called CLAUDE.md in your project folder. Claude reads it automatically at the start of every session, before you type a single word. Put your project context in that file — what the project is, what data you're working with, what you want Claude to help with, what you do not want Claude to decide alone — and that context is always present without you having to repeat it.

Think of it as the briefing document you'd give a capable research assistant on their first morning. Except this assistant gets the same briefing every morning and never forgets it.

Without CLAUDE.md, Claude answers your question. With CLAUDE.md, Claude answers your question as someone who already understands your project. The quality difference — especially for iterative work over days or weeks — is large.

A minimal CLAUDE.md might be three paragraphs: what the project is, what files are in the folder and what they contain, and one or two standing instructions (e.g. "always ask before modifying source files"). You can write it in thirty minutes. It pays back immediately.

→ How to write and use CLAUDE.md: A9.markdown-project-memory → The full personalisation system: A.issue.personalisation → For data-heavy projects with many specification files: A.issue.research-data-memory

4. On the questions you have before you start

Before trying anything new, most researchers arrive with a cluster of questions that feel like prerequisites:

Is this ethical? Where does my data actually go? Am I giving up authorship? Will this make me intellectually dependent?

These are real questions. They have real answers, and this site has dedicated documents for each of them. But there is something worth saying before you go looking for those answers.

Fail quickly.

The cost of a failed first experiment with Claude is near-zero: a few minutes, no data at risk, nothing published, nothing committed. The cost of not starting is invisible but real — you spend time worrying in the abstract rather than working with evidence about your own practice.

Try something small and low-stakes first. Use a paper you've already read. Write a CLAUDE.md for a project you're not precious about. Ask Claude to summarise something you could verify in five minutes. Your questions will sharpen considerably after one real experiment. The ethical questions become more specific. The privacy question becomes more answerable. The ownership question becomes concrete rather than hypothetical.

This is the foundational principle of research iteration with AI: fail early, cheaply, and with your eyes open — not after weeks of preparation. — B.lifecycle.0.iteration

If you want to engage with the "soft" questions now, go ahead — jump straight to:

If you'd rather try something first and come back to these later — that's also fine. They will still be here, and your questions will be better ones.

5. The adoption spectrum: where are you, and what does that mean?

Not everyone uses AI the same way. The difference between an occasional querier and someone running file-level workflows is not just practice — it is a structural shift in how AI is integrated into the work.

The adoption spectrum describes five positions:

Level 0 — Occasional querier. You open a web chat when you have a question. You might get useful answers; you also probably get hallucinations you don't notice because verification almost never happens.

Level 1 — Regular conversation partner. You use AI habitually. You've developed prompting instincts. You get real value. The risk here is subtler: sycophancy and confirmation accumulation — AI that affirms the framings you bring to it.

Level 2 — Context-aware collaborator. This is the key jump. Claude knows who you are and what you're working on. You have a project folder, a CLAUDE.md, a running context. AI is no longer answering questions from scratch — it is working inside your project. This is the right target for most researchers.

Level 3 — Workflow integrator. You have designed repeatable AI-assisted workflows, not just ad hoc sessions. Claude works on files, not just text. You've built verification steps into the process. Risk: false systematicity — the impression of rigour that repeatable workflows create, even when the AI outputs don't deserve it.

Level 4 — Agent orchestrator. Multi-step autonomous processes. You direct rather than do. This level genuinely requires programming skill and careful oversight architecture. You don't have to reach it — Level 2 is the right endpoint for most humanities researchers.

Knowing your current level is useful because it tells you what the next concrete step looks like — not just a vague aspiration to "use AI more." The transition from Level 1 to Level 2 is a thirty-minute task: write one CLAUDE.md. The transition from 2 to 3 is a file-level workflow experiment. Each step has a specific first experiment.

→ Full spectrum with risk distribution: B.adoption-spectrum → The transition guides (what to do next at each level): C.leveling-packages → Your level pages: Level 0 · Level 1 · Level 2 · Level 3 · Level 4

Where to go from here

If you've never used Claude Code: Start with installation: D.tutorial.setup. Then do the first project tutorial: D.tutorial.firstproject. Both are written for non-programmers on Windows.

If you want to understand the tool landscape first: Read the three-tier comparison: A1.desktop-vs-code. Then find your level: B.adoption-spectrum.

If you want to see what you can actually do: The task reference by level: C.tasks. The prompt gallery (copy-paste-and-adapt): C.prompts. The workflow templates: C.workflows.

If you want to understand the risks before committing: The calibration document (honest performance assessment): C.calibration. The risks map: C.dangers. The ethics decisions (data, disclosure, authorship): C.ethics-compact.

If you want the one-page reference to keep open: C.cheatsheet — shortcuts, slash commands, privacy quick-check, when-stuck protocol.

Related