Make a guide · The Guide Guide
Start here
From sources to an interactive guide · by hand or automated

The Guide Guide

How to turn a stack of research papers, reports or other sources into an explainer people understand, working with an AI assistant. The method grew out of building two interactive guides on how polymer melts flow: careful notes first, an outline you approve, then one chapter at a time, each one checked.

By Michael Lusk, researcher, Center for Polymer and Materials Technologies (CPMT), Ghent University

About 7 minutes of reading · three routes · prompts fold away until you open them

  1. STEP 0Set up a Project
  2. STEP 1Gather sources
  3. STEP 2Read and take notes
  4. STEP 3Outline prompt
  5. STEP 4Approve the outline
  6. STEP 5Build each chapter
  7. STEP 6Final check
  8. STEP 7Short version
You review and decide before the work goes onOptional

The method has eight steps, and you can work through them in three ways. By hand, you run each prompt yourself in a Claude Project and check every stage. Automated, you paste one prompt into Claude Code, which runs Steps 2 to 6 on its own and stops for your approval. Run kit is the automated run with extra tools, for people at home in Claude Code. You choose after planning your guide.

How this page works

  • Type your topic once, just below. Every prompt on the page fills in with it.
  • Prompts are folded. Open one, fill anything still marked in yellow, and press Copy.
  • Pick a route after planning your guide: By hand, Automated or Run kit. The steps in the side menu change with it.
  • Tick each step when it is done. Your ticks stay in this browser, so you can come back later.
  • Dotted words, such as Project, show a short definition when you point at them or tap them.

Your topic

Example

These lines fill every prompt. The example shows how they look; click into any box and it clears for your own.

Any kind of source works, from journal papers to reports and slide decks. Files with little selectable text, such as scanned PDFs or slide decks that are mostly pictures, give weaker notes and more gaps.

Showing the example. Your text is saved in this browser as you type.

Choose what your readers will get

Try each feature, then switch on the ones you want in your guide. Your choices go straight into Plan your guide, just below, where a few more options wait.

t = Wr (1)

Reading time is the word count divided by reading speed. Open any symbol to see its name, units and how it is built.

tReading timeminutes (min)

How long a reader needs for the text alone, without the time spent on figures and tools.

Built
Divide the word count W by the reading speed r. Words ÷ (words per minute) leaves minutes.
WWord countwords

The number of words in the reading text. Folded prompts and fold-outs like this one are left out, because readers open them only when they need them.

Built
Counted live from this page: … words of reading text. Derived
rReading speedwords per minute

How fast an adult reads silently. A large review of reading studies puts non-fiction at about 238 words per minute (Brysbaert, 2019, Journal of Memory and Language). Stated

Built
A rule of thumb, not a measurement: technical text with equations reads slower, so this page uses 200. Illustrative
t ≈ – min

Air-source heat pumps deliver less heat on the coldest days.

An invented example. Click the small number.

  • Stated the source says it
  • Derived calculated from stated values
  • Read from figure taken off a graph, approximate
  • Illustrative an example, not data

In the guide only the symbol shows, like this: the unit delivered 3.8Stated units of heat for each unit of electricity.

Point at the symbol, or tap it. Readers can switch the symbols off.

A heat pump’s coefficient of performance tells you how much heat it delivers for each unit of electricity. It falls as the temperature lift grows.

An example sentence from a guide on heat pumps. Point at, or tap, the dotted words.

Outside air cold Heat pump Room warm electricity heat is moved, not made

A still picture that explains a process at a glance.

Equation 1 as a plot. A faster reader’s line is flatter; the dot is this page.

More options, such as self-check questions and a glossary, are in Plan your guide.Go to Plan your guide

Plan your guide

Example

Who the guide is for, how long it runs and what it includes. All three routes use these choices. By hand turns them into the outline prompt and style rules in Step 3; Automated and Run kit build them into the instructions for the run.

Who will read it
How long it should take to read
30 min
6,000words of main text
4chapters
5interactive figures

Rules of thumb, not limits: 200 words a minute (Equation 1 turned around, W = r × t), one chapter per 8 minutes, one interactive figure per 6 minutes. Illustrative

What the guide should include

Choose your route

All three follow the same method and use the choices above. You can switch at any time.

Step 0

Set up a Project

A Project keeps your sources, your instructions and every saved document in one place, and every new session can read them. That matters more than it sounds: a long build fills the AI’s context window, and the next session needs somewhere to pick up from.

  1. Create a new Project and name it after your topic.
  2. Upload your sources to the Project’s files.
  3. Paste Prompt 0 into the Project’s instructions, not into a chat. It then applies to every chat in the Project.

What went wrong for us

Building our first guide, on polymer melts, ran so long that the conversation filled up several times. Each time, the work carried on because the notes, the outline and a build log were already saved in the Project.

Step 1

Gather your sources

A guide can only be as good as its sources. Three to ten sources is a workable range: enough to compare, few enough to read closely. Start with one overview or review to frame the topic, then add the original studies or reports that hold the data.

Check each source before you upload it:

0 of 4 checked

Watch out for

Scanned PDFs and slide decks with little text. The AI reads scanned equations and figures poorly, and a slide that is mostly pictures gives it little to work with. A wrong digit or a misread symbol can slip into the notes without any warning.

Step 2

Read and take notes

This step makes everything after it trustworthy. The AI reads every page, figures included, and writes notes that record where each fact was found, with the exact words that support it. Later, every claim in the guide can point back to a page, a table or a figure.

  1. Run Prompt 2a once. The AI lists your sources, saves a reading plan that splits them into rounds (set by the source count in Your topic), and reads round 1.
  2. Spot-check the notes: pick three numbers and search the source for the quote next to each.
  3. Run Prompt 2b for each remaining round. Each round gets its own notes document: Part 1, Part 2, and so on.
  4. When every round is done, run Prompt 2c, the cross-check.

What went wrong for us

Nine papers were too many to read well in one go; three notes documents of three papers each worked. Several key numbers existed only in figures, so they were tagged read from figure. Reading closely also turned up errors in the papers themselves, which became correction notes in the finished guide.

Step 3

Generate the outline prompt

The outline settles who the guide is for, what it covers and how long it runs, before a single chapter is written. Your choices in Plan your guide turn into two blocks of text. The outline prompt asks the AI for a plan. The style rules fix the conventions every chapter must follow, so the last chapter matches the first.

  1. Check your choices in Plan your guide: the reader, the goal, the reading time and the features.
  2. Press Copy both and paste them into one message, in a new chat inside the Project.
  3. Keep the style rules: you paste them again at the start of every build session in Step 5.

Your two blocks

What went wrong for us

Our first outline for this very page planned nine chapters and about 12,000 words, far more than anyone wanted. Nobody had said how long it should be. The reading-time slider is there so the AI knows from the start.

Step 4 · you decide

Review and approve the outline

The outline is the cheapest place to change course. Nothing has been written yet, so moving a chapter or cutting a feature costs one message. Read the outline once against these six questions:

0 of 6 checked

Any “no” goes into Prompt 4a. Repeat until every answer is yes, then approve with Prompt 4b.

What worked for us

Because nothing had been written, cutting this page from nine chapters to one page of steps took a single message. The same change after building would have meant starting again.

Step 5 · you decide

Build chapter by chapter

Build one chapter, check it, fix it, then move on. Small steps keep each chapter within what one session does well, and mistakes are caught before later chapters build on them.

  1. Run Prompt 5a in a chat inside the Project. It already contains your style rules from Step 3.
  2. When the chapter is built, run Prompt 5b in the same chat. It returns a pass or fail table and changes nothing.
  3. Open the chapter yourself: move every slider, open every fold-out, read one section aloud.
  4. Ask for the fixes you want, then mark the chapter done below and start the next.
  5. If the chat fills up, or you come back another day, open a new chat with Prompt 5c.

Chapter tracker

0 of 4 done

Tap a chapter to move it on: not started → built → reviewed → done. Prompts 5a and 5b fill in the first chapter that is not done.

What went wrong for us

Some chapters were too big for one turn and had to be built in parts. One calculation gave wrong answers until it was tested against values printed in the papers. Heavy figures slowed the whole page until they were set to load only when scrolled into view.

Step 6 · you decide

Final check in a fresh session

When every chapter is done, get a second pair of eyes. A new session carries none of the building session’s assumptions, so it reads the guide the way a reader will.

  1. Open a new chat in the same Project, so the sources and notes are there, and attach the finished HTML file.
  2. Run Prompt 6a. It returns a numbered table of issues and changes nothing.
  3. Decide on each issue: fix or skip. The fix list below writes Prompt 6b for you.

Fix list

8 to fix · 0 to skip

Tap an issue number to switch it between fix and skip. Skipping is fine for low-severity polish or anything you disagree with.

Watch out for

A checker without the sources can only check the guide against itself. Run the audit inside the same Project, never in a chat on its own.

Step 7 · optional

A short version

Some readers want the main points in five minutes. A short version made from the finished, checked guide keeps the same sources and the same style.

What went wrong for us

Our first short version summarised the theory itself, when what was wanted was what the research had found. The prompt now asks for the angle first.

Later: add a source

A guide goes out of date when a new study appears. Upload the new source to the Project and run Prompt 8. It takes notes on the new source only, and proposes changes before it touches the guide.

Automated · A1

Before you start

Claude Code is Claude working directly with the files in a folder on your computer. On this route it runs Steps 2 to 6 for you: it reads every source, checks its notes against the sources, writes the outline and waits for your approval, then builds and audits the whole guide.

Check these first:

0 of 4 checked

Automated · A2

Choose the compute intensity

The run splits the work into roles. A coordinator plans, checks and settles edge cases; helpers read, build and check. Stronger models work more carefully and use more of your plan’s allowance, so choose what your plan can carry.

Compute intensity
RoleModel
Relative usage of your plan’s allowanceModerate

A rough comparison between settings, not a measurement. Illustrative The top model tier, above Opus, is left out on purpose: it uses far more allowance than this job needs.

Which setting?

On a Pro-type plan, start with Light. On a Max-type plan, Balanced suits most guides; choose Thorough for a guide you will publish or teach from. If the run reaches a usage limit it stops, and you can pick it up again later where it stopped; the next step says how.

With your setting, the run goes through these phases. The step on the right is the matching step of the By hand route.

    Automated · A3 · you decide

    Run it in Claude Code

    One prompt starts the whole run. It stops for your approval and waits for you.

    1. Open a terminal in your project folder, the one that holds sources, and start Claude Code.
    2. Switch to the coordinator’s model by typing /model opus.
    3. Press Shift+Tab until “accept edits on” shows under the input box. Changes to files in this folder then go ahead without a question each time.
    4. Copy Prompt A and send it. Claude Code still asks before it runs a command. When you are comfortable with one, choose the option not to be asked again, and stay nearby for the first few minutes, until the questions stop.
    5. When it stops with the outline, check it against the six questions in Step 4. Reply with your changes, or with “Approved”.
    6. It stops again when the first chapter is built. Open guide/guide.html and judge the tone, the level and the design. Reply with your changes, or with “Continue”.
    7. The rest then runs without you, apart from an occasional permission question. If the session ends early or reaches a usage limit, start a new one in the same folder and send Prompt A again. It reads its own logs and carries on where it stopped.

    Watch out for

    Everything the run knows lives in the files it writes: the plan, the notes and the logs. Keep the folder as it is until the guide is finished, or a resumed session has nothing to pick up from.

    Want stricter checks?

    The Run kit route adds a script that checks every quote against its source, a tested page template and helper roles with fixed models. It needs Python and a few more set-up steps.

    Run kit · A3 · you decide

    Run it with the kit

    The run kit holds the instructions and the tools for the run, and Prompt B starts it. The run stops for your approval and waits for you.

    1. Press Save the run kit below and unzip it into your project folder, so that run-instructions.md and the kit folder sit next to sources.
    2. Open a terminal in that folder and start Claude Code.
    3. Switch to the coordinator’s model by typing /model opus.
    4. Press Shift+Tab until “accept edits on” shows under the input box. Changes to files in this folder then go ahead without a question each time.
    5. Send Prompt B. Claude Code still asks before it runs a command. When you are comfortable with one, choose the option not to be asked again, and stay nearby for the first few minutes, until the questions stop.
    6. When it stops with the outline, check it against the six questions in Step 4. Reply with your changes, or with “Approved”.
    7. It stops again when the first chapter is built. Open guide/guide.html and judge the tone, the level and the design. Reply with your changes, or with “Continue”.
    8. The rest then runs without you, apart from an occasional permission question. If the session ends early or reaches a usage limit, start a new one in the same folder and send Prompt B again.

    The run kit

    one zip file, built from your choices
    • run-instructions.mdthe full instructions for the run, built from your choices
    • kit/guide-template.htmlthe starting page, with ready-made parts such as equation boxes, source tags and figures
    • kit/check_quotes.pya script that checks every quote in the notes against its source
    • .claude/agents/the four helper roles, each with the model you chose. The folder name starts with a dot, so your file manager may hide it
    Save it again whenever you change a setting.
    Read the kit’s files, or create them by hand

    If saving does not work in your view, create the three files yourself: copy each block into a text file with the name shown. Without the .claude folder the run still works; the coordinator then sets each helper’s model itself.

    Watch out for

    Everything the run knows lives in the files it writes: the plan, the notes and the logs. Keep the folder as it is until the guide is finished, or a resumed session has nothing to pick up from.

    Automated · A4 · you decide

    Review what comes back

    The run ends with three files, next to its notes and plan. Read them in this order.

    1Open issuesreview/unresolved-issues.mdEverything the run could not settle from the sources, each with its location, the evidence and the options. Decide these first.
    2Decision logreview/decision-log.mdEvery choice the run made on its own, with the reason, the evidence and which model decided. Rows marked for review deserve a look.
    3The guideguide/guide.htmlOne self-contained file. Open it, move every slider, open every fold-out, and read one section aloud.

    What the decision log looks like

    #PhaseDecisionOptions consideredChoice and reasonEvidenceDecided byReview?
    7VerifySMITH21 gives a different value in a table and in a figureUse the table; use the figureTable value used; the figure value is quoted as read from figure, ±10%. Tables are printed numbers.SMITH21 p. 6, Table 2; Fig. 4Coordinator (Opus)no
    12BuildChapter 3 runs 40% over its reading timeTrim the worked example; split the chapterWorked example cut to one case, so the approved outline stays as it is.Outline, chapter 3: 8 minCoordinator (Opus)yes

    Invented example rows. Illustrative

    1. Glance at review/quote-check.md. It lists any quote the script could not find in its source.
    2. Work through the open issues. Tell Claude Code what to do about each one, in the same session or in a new one started as in the step before.
    3. Check every decision-log row marked for review, and any row you disagree with.
    4. For a second opinion, start a new Claude Code session in the same folder and send Prompt C. It audits the finished guide from scratch and changes nothing.
    5. Optional: make a short version with Prompt 7.

    Later: add a source

    When a new study appears, put the file in sources, start Claude Code in the same folder and send Prompt D. It takes notes on the new source only, and stops with its proposed changes before it touches the guide.

    Watch out for

    The run checks its own work at every phase, yet it cannot know what you meant. Your outline approval and your read of the open issues are where your judgement comes in.

    Finished? Add your guide to the library

    Other researchers can then read it, and mark it once they have followed it and found it to work. No account needed.

    Submit your guide

    Found this useful?

    The Guide Guide is free to use and share. If it saved you time, you can buy me a coffee. It is entirely optional, and very much appreciated.

    Buy me a coffee

    buymeacoffee.com/theguideguide