Developers got Claude Code in 2024 and their day changed shape. ![Four-panel comic: a paragraph-long work order — read three investor meeting notes, extract commitments, update the pipeline database — unfolds into the finished pages and a follow-up draft.](/static/blog/capy-claude-code-documents-workorder.webp) *Work order, not a prompt. One paragraph in, three pages read, pipeline database updated, follow-up draft waiting.* The old loop was: open an AI chat tab, describe the problem in a few paragraphs, read the reply, copy the code back into the editor, run it, hit an error, go back to the chat tab, paste the error, repeat. The new loop is: type a paragraph-long instruction describing a whole multi-step task, hit enter, come back when the agent reports done. Read five files, edit three of them, update the tests, run them, fix any failures, summarize what changed. That is not a chat — that is an agent doing the work. Non-developers never got that loop. The AI experience for everyone else is still a chat tab and a copy-paste shuttle between that tab and whatever document tool the work actually lives in. The step-change skipped the 95% of knowledge workers whose day is prose, PDFs, meetings, and spreadsheets — not source code. This post is about what Claude Code for documents actually means, why it's the right mental model for AI that works on notes instead of code, and why Docapybara is our attempt to ship it properly. ## The three things Claude Code changed that non-coders never got Cursor and Copilot came first but stayed close to a single file at a time — the AI helps with the line you're writing, the function you're editing. Claude Code widened the aperture. Three things stand out: 1. **Paragraph-long instructions, not chat turns.** You don't ping-pong with it. You describe a full multi-step task in one go: "read these five files, figure out why the flaky test keeps timing out, propose a fix, apply it, rerun the test, and report." The agent plans the sequence itself. 2. **Tools, not just text.** It can open files, edit them, run commands, read output, follow up with more edits. It has hands. 3. **Project-wide context, not per-session state.** A file called `CLAUDE.md` at the root of your repo tells the agent what the project is and how to work on it. Every new conversation starts with that context loaded automatically. Put together, those three things are what turned AI from "a thing you chat with" into "a thing you delegate to." And non-coders got zero of it. The AI tools marketed to the rest of us are still chat tabs. Maybe with a side panel. Maybe with a plugin ecosystem you have to stitch together on a Saturday. But the loop is still: describe, read, copy, paste, repeat. Claude Code for documents means lifting that same three-part shape — paragraph instructions, real tools, persistent vault-wide context — and putting it on top of notes, PDFs, and meeting recordings instead of source files. ## What a Claude Code-style instruction looks like when your project is a vault, not a repo Let's make this concrete. Here are five instructions a knowledge worker might type into Docapybara today. Each of them would require five or six ChatGPT turns and a lot of copy-paste to accomplish otherwise. - *"Read the three investor meeting notes from this week, pull out every follow-up I committed to in them, and add one row per follow-up to my Investor Pipeline database with the investor name, the commitment, and the due date I mentioned."* - *"Open the 47-page industry report I uploaded yesterday, summarize the three strategic shifts it argues for, draft a one-page briefing for tomorrow's client meeting, and drop each of the report's supporting citations as a footnote on the matching claim."* - *"Transcribe the call recording on today's date, label the speakers, pull out every objection the prospect raised, draft a follow-up email addressing the top three, and add the remaining objections as rows on the objections database."* - *"Go through my last month of morning-standup notes, find every project I mentioned more than twice that doesn't have a page of its own yet, create a stub page for each, and tag them with the team that owns them."* - *"Take the fundraising update draft I started on Monday, rewrite it to match the tone of the three updates I sent last quarter, pull the new MRR numbers from the October KPIs page, and add a section on hiring that pulls from the team page."* None of those are prompts. They are work orders. The agent plans the steps — which pages to open, which tools to use in which order, what to write back where — and executes the whole sequence. You don't watch it work. You come back when it's done and review the diff. That is the shape Claude Code gave developers. That is the shape Docapybara gives you. ## Why the shape needs a markdown-native vault underneath This is the load-bearing architectural decision, and it's the same bet Claude Code makes about code. Plain text is fast. Claude Code is fast on codebases because code is plain text — the agent can grep, diff, edit, and save without walking an object graph. No block tree, no JSON abstraction, no format to preserve across edits. Just files. Most notes apps made the opposite bet. Your page isn't a text file — it's a tree of JSON blocks, each with an ID, a type, and metadata. It renders beautifully in the editor. But when an agent has to do vault-wide refactors at scale, that block tree is a tax on every single operation. The agent reads slow, writes slow, and frequently corrupts formatting on round-trip. You feel it as: "the AI feature works on one page but chokes when I ask it to touch fifty." Docapybara is markdown-native. Every page is a plain markdown file, same shape Obsidian uses. The agent reads, edits, and moves content at text-file speed. Inline databases live inside the markdown via a directive — they don't require the agent to drop into a different data model mid-task. PDFs are converted to markdown on upload so the agent sees them as searchable text rather than opaque blobs. The whole vault looks like a project to the agent, not a tangle of formats. Without that, the Claude Code loop doesn't feel the same. With it, the loop just works. ## The integrated agent, 27 tools, zero plugin setup The other thing Claude Code gets right is that the agent ships with tools already wired in. You do not install a "read file" plugin, an "edit file" plugin, a "run command" plugin, and hope they talk to each other. The day you run Claude Code, you have the full toolbox. Docapybara is built the same way. The agent has 27 tools from day one — search the vault, open a page, create a page, edit a page, move pages, query an inline database, add a database row, extract from a PDF, transcribe an audio file, rewrite a section, draft from a template, and so on. Every one of those tools was wired up against the others on purpose. You sign up, you drop in a document, you talk to the agent in plain English. There is no setup Saturday. That is the difference between "a notes app with an AI chat button" and "a Claude-Code-style agent living inside a notes app." The first lets you have a conversation next to your work. The second does the work. ## Who this is for Docapybara is not a developer tool. You can run Claude Code itself on a folder of markdown notes if you want — some people do — but it is a CLI, it expects a git repo, it assumes you are comfortable managing a `CLAUDE.md` context file and a terminal session. If you are not a developer, that is friction you should not have to pay to get the productivity step-change. Who we built this for: - **Writers** drafting long-form work from their own notes, transcripts, and research. - **Founders** running fundraising, hiring, product, and content out of one brain and one vault. - **Marketers** turning customer research and call transcripts into campaign briefs without re-reading twenty pages. - **Consultants and analysts** whose week is reading PDFs and producing docs — and who need an agent that can actually read at scale. - **Operators and ops leads** whose work is SOPs, meeting notes, incident reviews, and "write up what we decided." - **Researchers** working through stacks of source material with an agent that reads, quotes, and cites rather than summarizing from a blank prompt. If your work is reading, writing, organizing, and deciding — and you are not a developer — the Claude Code loop was built for your kind of work. It just got delivered through the wrong front door. We rebuilt the front door. ## Four concrete Claude-Code-style workflows in Docapybara ### Prep for a client meeting from a pile of uploaded PDFs Drop the client's annual report, three industry pieces, and the notes from your previous engagement into a page. Tell the agent: "Compare what has changed at this client since we last worked with them, draft a one-page meeting prep with five talking points, and cite each claim back to the source document." Ninety seconds. You get a brief with citations pointing at the actual passages. The agent read everything in the vault, not just the page you had open. ### Turn a month of meeting notes into a quarterly update You have twelve meeting notes pages from the last quarter. Ask the agent: "Read every meeting note page from the last three months, extract every commitment I made and whether it's marked complete, group them by project, and draft a quarterly update I can edit." The agent walks the vault, reads each page, tracks status across pages, and produces the update grounded in your own notes — not a generic template. ### Draft a sales follow-up from yesterday's call recording Drop the audio file onto a page. Docapybara transcribes with speaker labels so you can see what you said versus what the prospect said. Tell the agent: "Draft a follow-up email referencing the three commitments the prospect made, add a table of their open questions with the answer I want to send, and put both in a new page tagged with the deal name." One sentence of instruction, one finished page, plus the email draft. ### Refactor your personal knowledge base You have 800 notes accumulated over two years. Most are not tagged consistently. Ask the agent: "Read every note in the vault, suggest a tag taxonomy of no more than fifteen tags based on what is actually there, and apply the tags to each page. Create an index page listing the tags with one line of description each. Leave a note on any page you were not sure how to tag." An afternoon of cleanup done while you get lunch. Every one of those is a paragraph-long instruction and a multi-step result. The Claude Code loop. On documents. ## Related reading Docapybara's broader positioning — why the shape of a single-user markdown vault with an integrated agent is the right answer for the non-developer knowledge worker — is covered in a few sibling posts: - The companion take on the same shape through the Cursor lens is **[Cursor (and Claude Code) for Documents](/blog/claude-code-for-documents/)**. Same product, earlier framing. - For the broader "I am not a developer and I want the Cursor/Claude Code step-change" pitch, see **[AI for knowledge workers who aren't developers](/blog/ai-for-knowledge-workers/)**. - For the full map of workflows by job type — sales, marketing, fundraising, research, ops — see the **[AI for work](/blog/ai-for-work/)** hub. - If you're specifically evaluating Docapybara against Notion, we wrote an honest **[Notion alternative comparison](/blog/vs-notion/)** covering the three shapes of alternative. ## FAQ **Is there a claude code for writers?** Yes — that is the entire pitch for Docapybara. The Claude Code loop (paragraph-long instruction, multi-step execution, full access to a project's context, tools that actually act on files) transfers cleanly to writing. You describe what you want done with your vault, the agent executes across your pages, PDFs, and transcripts, and you come back to a finished draft. Docapybara strips the CLI, the git repo, and the `CLAUDE.md` overhead so the loop is approachable from a normal notes app, not a terminal. **Is there a claude code for non-coders?** That is Docapybara's sentence-one positioning. Claude Code was built for developers and assumes a code-shaped workflow — a terminal, a git repo, a codebase. Docapybara ships the same agentic loop on a notes-shaped workflow — a vault of markdown pages, uploaded PDFs, call transcripts, and inline databases. Same mental model, different surface. If you can describe a multi-step task in plain English, you can use it. **What's the difference between "Claude Code for documents" and just using Claude Code on a folder of markdown?** Technically you can point Claude Code at any folder of markdown, and a handful of power users do. The difference is everything around the agent: PDF ingestion, inline databases, meeting transcription with speaker labels, a real editor, auth, cloud hosting so you can reach your vault from anywhere, 27 tools wired to the specific shape of notes work. You also don't manage a `CLAUDE.md`, a terminal session, or git. Docapybara is what you'd build if you took the Claude Code shape and designed the rest of the product around knowledge work instead of code. **Is Claude Code for documents just cursor for documents with a new name?** The Cursor framing came first and we still use it — see the sibling post. But Cursor and Claude Code are not the same shape. Cursor is an editor with AI bolted in; the agentic surface is a sidebar. Claude Code is agent-first; the whole interface is built around "tell me a paragraph, I'll do the multi-step work." For documents, Claude Code is the closer metaphor — you rarely want to be in a file-by-file inline-edit mode with notes; you usually want the agent to go read fifty pages and produce something new. So if you had to pick one label, Claude Code for documents is the more accurate one. **Do I need Anthropic API keys or any technical setup?** No. Docapybara is a cloud-hosted product. You sign up with an email or Google account and the agent is ready. No API keys, no model selection, no token-per-minute knobs, no terminal. The Claude Code experience, without the Claude Code ceremony. **How is this different from Obsidian with the claudian plugin?** Claudian is a community-built plugin that wires Claude Code into Obsidian — impressive, and a strong signal of demand. The gap it leaves is the integration surface around the agent: no integrated PDF-to-markdown pipeline, no built-in meeting transcription, no inline databases, no cloud sync, no account system, and the agent still runs in a terminal that happens to be pointed at your vault. Docapybara is the productized version of that same instinct — the Claude Code loop on a notes-shaped product, not stitched on top of one. **Is this local-first or self-hostable?** No — Docapybara is cloud-hosted. Your vault lives on our infrastructure, accessible from any browser. If local-first is a hard requirement for you, Obsidian plus claudian is closer to that shape. We made the cloud choice so the agent, the PDF pipeline, and the transcription stack all work out of the box without you managing a local toolchain. It's a deliberate trade: setup-free against fully local. --- Developers got a paragraph-long instruction loop that does the multi-step work and reports back. That loop is too useful to leave in a terminal. Claude Code for documents is Docapybara's attempt to bring it everywhere else — to the writers, founders, marketers, consultants, and operators who think in prose and deserve the same leverage. [Try Docapybara free](https://docapybara.com). Drop in your last five PDFs, a recent meeting recording, and a half-finished draft. Tell the agent what you want done with them, in one paragraph. See what "claude code for documents" actually feels like when it's pointed at your work.