# A thesis research base — chapter by chapter, argument by argument Recipe No. 40, Research and civic life. From The know.sh Cookbook: https://know.sh/cookbook/thesis-research - For: a PhD candidate in history, two years into the archives and starting to write up - You bring: typed archive notes from a dozen visits, a chapter plan your supervisor has seen, the notes from every supervision meeting you can find, and the AI assistant you already use - You get: a thesis shelf with a document per chapter, each opening with the chapter’s argument, and the evidence, counter-arguments, decisions and open questions filed beneath it - Time: an evening for your assistant to file the notes, a week of evenings checking them, then half an hour after each archive day or supervision - Keep it: private while you work; share one document by a read-only link A history dissertation is written from two piles. One is the archive: minute books, ledgers, letters and newspapers, noted over two years in notebooks and phone photos. The other is the argument, which lives in your head, in a chapter plan from last spring and in whatever you agreed with your supervisor in the meeting you did not write up. This recipe makes the thesis a shelf and each chapter a document. Your assistant — Claude, ChatGPT or a local model — does the filing: it reads your typed archive notes and supervision notes and files each item as a finding under the chapter it serves. You check every reference in the editor and write the one thing it must not: the overview of each chapter, which says in a paragraph what that chapter argues. The argument is always at the top of the page, and the evidence, the objections, the supervisor's decisions and the open questions are one click below it. The thesis itself is written in your word processor, by you. know.sh is the research base the chapters are drawn from, and the record of how the argument changed. ## What you will use - **Shelf**: The thesis as a shelf, *Thesis: baths and wash-houses*, with the working title and the submission date in its line. - **Research document**: One document per chapter; the overview states the chapter’s argument in a paragraph you rewrite as it changes. - **Finding**: Archival evidence as *Evidence*, counter-arguments as *Risk*, supervision outcomes as *Decision*, gaps as *Question*. - **Your AI assistant**: Files your archive and supervision notes as findings under the right chapters, checks a chapter’s claims against its evidence, and writes quizzes. - **The editor**: Where you check each archive reference, move findings between chapters, set types and **Key**, and write every chapter argument yourself. - **Revisions**: Every version of every chapter argument, marked as yours or an assistant’s; **Restore** if a rewrite went wrong. - **Campaigns**: The writing year laid out by a coding assistant such as Claude Code: a set per chapter, with archive visits that block drafts. - **Public link**: A password link to one chapter’s document for your supervisor to read before a meeting. ## Method ### 1. Have your assistant build the thesis shelf Connect your assistant to know.sh once, following the support page for Claude, ChatGPT or a local model. Give it the chapter plan and ask it to make a shelf named for the thesis, with the working title and deadline as its line — "Clean bodies, public purses: baths and wash-houses in three northern towns, 1846–1914. Submission September 2027" — and one document per chapter, numbered in the title so the shelf lists them in order. Tell it to leave every chapter overview as a single line, "Argument to be written". Ask for a second shelf, *Archives*, for visit logs and finding aids; the thesis shelf is for argument. ### 2. Give it the archive notes to file Attach your typed archive notes, a visit at a time. Ask the assistant to file each item as an *Evidence* finding in the chapter it serves, titled with what it shows and where it is from — "Baths Committee cuts second-class price to 1d, March 1878" — with the transcription or paraphrase exactly as you typed it, the full archive reference (repository, collection, box, folder, page), and a line on why it matters. Items that fit no chapter go in the *Archives* shelf's visit log. Two hundred notes become two hundred findings in an evening. It is filing, not interpretation: tell it not to summarise away your wording. ### 3. Check every reference in the editor Now go through each chapter in **Edit**. Check every archive reference against your notebook or the catalogue; references are the easiest thing to garble in transit. Drag findings into the order the chapter will use them, move misfiled ones, and delete duplicates. Where a finding is the only evidence for a claim, mark it **Key**, and highlight the decisive line so the finding's number is bold in the index. Those are the findings to reread before every supervision. ### 4. Write each chapter’s argument yourself Replace each placeholder overview with the chapter's argument, in three to six sentences of your own: the claim, the evidence it rests on, and what it adds to the existing literature. "Chapter 4 argues that the penny second-class bath, not the Act, is what made public bathing working-class in Thornleigh after 1878." If you cannot write that paragraph, you have found the chapter that needs thinking before it needs writing. Put a *Question* finding at the top of it that says what you do not yet know. ### 5. Keep the counter-arguments beside the evidence An examiner will ask what else explains your evidence. File each objection as a *Risk* finding in the chapter it threatens: "The social-control reading: baths as a tool of moral reform rather than something workers wanted." Say who makes the argument, what evidence supports it, and how you answer it, or that you cannot yet. Your assistant can file these from your reading notes on the historiography, but the answer to each is yours. A chapter with ten *Evidence* findings and no *Risk* findings has not met its critics. ### 6. File supervision as decisions, and share before meetings After each supervision, give your assistant your meeting notes and ask for a *Decision* finding in each chapter they concern, titled with the date and the decision: "12 September: drop Kellbrook, keep two towns". Questions you could not answer go in as *Question* findings. Read them in the editor; the wording of a decision should be one you and your supervisor would both recognise. Before the next meeting, press **Share** on the chapter under discussion and make a link with a password and an expiry of a week. Your supervisor reads the argument and its findings, without your highlights or notes, and replies by email or in the meeting. ### 7. Focus the chapter you are writing Press **Focus** on the chapter in progress. For seven days it rises to the top of your title page, so it is the first thing you open. When the week ends, *Revisit* says so, and you decide whether to renew it or move on. As you draft, the argument will shift. Rewrite the overview when it does. Revisions keep each version with its date, so the history of the chapter's thinking is on record; if a rewrite loses something, open the earlier version and **Restore** it. ### 8. Lay out the writing year as a campaign Planning tools need a coding assistant with the know.sh plugin; Claude Code has one (`claude plugin marketplace add knowsh-curtis/know-cli`, then `claude plugin install know-dev@know-sh`, start `claude` and sign in with `/mcp`). Ask it to read the thesis shelf and lay out a campaign: one operation set per chapter, operations for each remaining archive visit, draft, revision and supervision, and dependencies where one waits on another. It sets the plan out; it does not write chapters. Mark operations *active* and *done* as you go, and look at the graph when a deadline moves: it shows which drafts are blocked by an archive trip you have not booked. ## Prompts to try Your assistant, connected to know.sh (Claude, ChatGPT or a local model): > I am attaching my typed notes from two visits to Thornleigh Local Studies. Using know.sh, file each numbered item as an Evidence finding in the chapter document it serves on my shelf “Thesis: baths and wash-houses” (the chapter plan is in the shelf notes). Title each with what it shows and its date; keep my transcription word for word; put the full archive reference on its own line. Items that fit no chapter go in the visit log on my “Archives” shelf. Do not change any chapter overview. Your assistant, connected to know.sh (Claude, ChatGPT or a local model): > Using know.sh, read the overview of “Chapter 4 — Who bathed” on my shelf “Thesis: baths and wash-houses” and list every claim it makes. For each, name the Evidence findings that support it and any Risk findings that challenge it, and list claims with no supporting finding separately. Leave the list as proposed notes on the overview; do not edit it. A coding assistant with the know.sh plugin, such as Claude Code: > Using know.sh, read the shelf “Thesis: baths and wash-houses” and create a campaign called “Writing year to submission, September 2027”. Make one operation set per chapter. In each, add operations for archive visits still needed (from Question findings), first draft, supervisor reading and revision. Add a blocks dependency from each archive visit to the draft that needs it. Leave everything in draft status. Do not write or edit any chapter text. ## Variations - Still at the reading stage? Start with the [literature review](/cookbook/literature-review): a document per theme, a finding per book or article, then move the synthesis into the chapter overviews here. - Working with interviews as well as archives? File interview summaries on a shelf of their own, as in the [oral history collection](/cookbook/oral-history), and link them from the chapters that use them. - Preparing for the viva: ask your assistant for a quiz on each chapter drawn from its *Risk* findings, and take it in know.sh, so you practise answering your critics. - For a book from the thesis, move the chapter documents to a new shelf and keep a *Decision* finding on each for what the publisher’s reader asked you to change. ## Where it falls short - No archive photos or scans. know.sh holds text only, so keep images of documents in your photo library or a folder, and write in each finding which image it is. - There are no footnotes and no reference formatting. Keep citations in your reference manager and full archive references in finding bodies. - Your supervisor reads a link to one chapter at a time and cannot comment in it; the conversation happens in meetings and email, and you file it. - Campaigns are a plan, not a reminder system, and creating one needs an assistant with the planning permissions, such as Claude Code. Nothing will tell you a deadline is near; you have to look. - Smaller local models call tools less reliably and can misfile a long batch of notes. Give them one visit at a time and check what they file; Revisions shows every change. ## A note on authorship and your institution’s rules A research organiser, not a ghost-writer. The dissertation must be your own work: the argument, the reading of the sources, and the prose. Use your assistant to file notes, check a chapter's claims against its evidence and lay out a schedule. Do not have it draft chapters or chapter arguments. Check your university's policy on generative AI in research degrees, and follow its rules on declaring AI use in the thesis. Assistants make mistakes. Any reference or fact an assistant supplies must be checked against the original document before it goes near a chapter. Archive references in particular are easy to garble: check them against the catalogue, not a summary. Some archives also restrict giving copies or transcriptions of their holdings to outside services; check before you hand notes to an assistant. Before sharing a chapter, even with your supervisor, check whether any archive restricts quotation of its material in unpublished work. Indexed under: Thesis chapters, Chapter arguments, Archival evidence, Counter-arguments, Supervision, Writing schedule, Dissertations.