End-to-end workflows

Four complete workflow descriptions for common research tasks. Each workflow is a step-by-step procedure you can follow from a standing start. The prompts are drawn from C.prompts; the underlying format is the markdown hub described in A.markdown-central.

Prerequisites for all workflows: Claude Code running in a project folder. A CLAUDE.md is not required for Workflow 4 (which creates one), but all other workflows benefit from having one. → D.tutorial.setup, D.tutorial.firstproject

Workflow 1: Reading a set of papers

Scenario: You have 5–20 PDFs to engage with. You want to come away with a synthesis, a set of notes, and a sense of what to read next — without spending a week on it.

Time: 20–40 minutes of active work; Claude works in the background while you review outputs.

Folder structure:

/myproject/
  CLAUDE.md
  /papers/         ← put PDFs here before you start
  /papers_md/      ← markdown versions will go here
  reading_notes.md ← final output

Step 1: Convert PDFs to markdown

Convert all PDF files in /papers/ to markdown.
Save each as [original-filename].md in /papers_md/.
Note any files where conversion was partial or failed.

Check that the markdown files have appeared. Open one in a text editor to confirm the text is readable and not garbled.

Step 2: Triage and prioritise

Read all markdown files in /papers_md/. My research focus is: [your specific question
or argument — refer to CLAUDE.md].
For each paper: (1) relevance to my focus (high/medium/low), (2) one-sentence summary.
Write a triage table to papers_triage.md. Sort by relevance.

Review the triage table. Adjust the priority order if Claude has misread your focus. This is cheap to do now and expensive to discover later.

Step 3: Deep read the high-relevance papers

Read the papers rated "high" in papers_triage.md.
For each one:
1. Central argument (one sentence)
2. Key evidence or method
3. Strongest objection a critic would make
4. Most useful quote for the argument that [my argument]
Write to reading_notes.md, one section per paper.

Step 4: Synthesis

Based on all the papers in /papers_md/:
1. What themes appear across more than one paper?
2. What substantive disagreements exist?
3. What gaps — questions raised but not answered — appear most often?
4. What would be the most useful thing to read next?
Add a "Synthesis" section to reading_notes.md.

Step 5: Update your CLAUDE.md

Based on these papers and the synthesis in reading_notes.md,
suggest 3–5 updates to my CLAUDE.md that would help you assist me better going forward.

Review the suggestions. Accept those that accurately reflect your current understanding; discard or rewrite those that do not.

What success looks like: A reading_notes.md with per-paper notes and a synthesis section; a triage table that helps you decide what to read next; an updated CLAUDE.md that knows what you have just read.

Workflow 2: Writing a grant section

Scenario: You are writing a grant application (GAČR, ERC, Horizon, or similar). One section needs to be drafted. You have notes, a project description, and relevant sources already in your project folder.

Prerequisites: A rich CLAUDE.md with your project description, key claims, and relevant publications. The section of the call text you are responding to saved as a file or pasted into the session.

Time: 1–2 hours including review and iteration.

Step 1: Orient Claude to the task

I am writing the [section name] section of a [GAČR / ERC / Horizon] application.
The call text for this section is in [call_section.md / or paste it here].
Read my CLAUDE.md and the call text. Tell me:
1. What are the three most important things this section needs to demonstrate?
2. What in my CLAUDE.md is most relevant to this section?
3. What is missing from my CLAUDE.md that I should add before we draft?

Answer the questions Claude raises. Add the missing context to your CLAUDE.md or directly into the conversation.

Step 2: Draft

Draft the [section name] section. Constraints:

- Max [X] characters/words (check call requirements)

- Register: formal academic, matches the tone of the call text

- Key points to hit: [list 3–5 from your notes]

- Do not add claims I haven't made in my notes or CLAUDE.md

- Where I've used a specific phrase in my existing documents, use it

Step 3: Review the draft yourself first

Read the draft completely before asking Claude to revise anything. Mark:

Step 4: Targeted iteration

Revise paragraph [N]. Problem: [specific issue — too long / too vague /
doesn't address the reviewers' likely concern about X].
Proposed direction: [your idea of what it should say instead].
Write two alternative versions, labelled A and B. Do not change the other paragraphs.

Repeat for each paragraph that needs attention. Make targeted changes, not wholesale rewrites — each revision is easier to review than a complete new draft.

Step 5: Stress-test before submitting

Play the role of a rigorous reviewer for [GAČR / ERC / Horizon].
Read this draft section.
What are the three strongest objections a reviewer might raise?
What claim would they ask me to support with a citation?
Do not soften the critique.

Address the objections before submission.

What success looks like: A draft that is substantially your argument in substantially your voice, with Claude having handled the mechanical load of first-draft generation and structural suggestions. Every factual claim verified against your own research. Reviewer objections addressed.

Workflow 3: Processing an archival corpus

Scenario: You have returned from an archive trip with a folder of scans. You want to know what you have, what is readable, and what is in it — before investing hours in close reading.

Time: Depends on corpus size; typically 30–90 minutes of active work.

Folder structure:

/archive_project/
  CLAUDE.md
  /scans/             ← original scans (PDFs or images)
  /transcriptions/    ← readable text versions go here
  corpus_overview.md
  extracted_data.csv

Step 1: Triage what is readable

I have [N] scans in /scans/.
They are probably [13th-century inquisition registers / notarial acts / court records]
from [region, date range].
For each file:
1. Can you read it? (yes / partial / no — note why if not)
2. Document type (deposition / register entry / other)
3. Any names or dates visible in the first few lines
4. Approximate completeness (full / partial / fragment)
Write to corpus_overview.md as a table.

Review the table. Separate the readable from the unreadable. Unreadable scans may need manual transcription or specialist OCR before Claude can work with them.

Step 2: Convert readable files to text

For all files marked "yes" or "partial" in corpus_overview.md:
Convert each to a text file and save in /transcriptions/ as [original-filename].txt.
For partial files, include everything that was readable and mark unreadable
sections with [illegible].

Step 3: Extract structured data

Read all files in /transcriptions/. Each is a [document type].
Extract from each:

- Document ID: [filename]

- Date: [as given in the document, or "not stated"]

- Location: [as given, or "not stated"]

- Named persons: [list with any identifying information]

- Key content: [2–3 sentence summary]
Do not infer information not in the document. Write "not stated" where absent.
Write to extracted_data.csv with one row per document.

Step 4: Find patterns and connections

Based on extracted_data.csv and the transcriptions in /transcriptions/:
1. What is the overall date range and geographic spread of this corpus?
2. Which persons appear in more than one document?
3. What recurring themes or situations appear?
4. What would be the most productive documents to read closely first?

Step 5: Deep read the priority documents

For the priority documents identified in Step 4, run the DISSINET source analysis prompt from C.prompts to extract persons, places, and relations for network data.

What success looks like: A corpus_overview.md that tells you what you have; an extracted_data.csv with structured data from every readable document; a priority reading list; persons appearing across multiple documents already flagged.

Workflow 4: Setting up a new project

Scenario: You are starting to use Claude Code for a specific research project — a paper, a chapter, a grant proposal, or a long-term corpus. You want a properly configured project folder that Claude will be able to work in effectively from session to session.

Time: 30–45 minutes.

Step 1: Create and initialise the folder

mkdir myproject
cd myproject
git init
claude

Claude launches in the working directory and reads whatever is there.

Step 2: Tell Claude about your project

I am setting up a new research project.
Before I describe it, ask me five questions that will help you understand it
well enough to be useful as an assistant.

Answer Claude's questions. Be specific about: what you are arguing or studying, what materials you have, what kind of output you need, and what you are not yet sure about.

Step 3: Draft a CLAUDE.md

Based on what I've told you, draft a CLAUDE.md for this project. Include:

- What the project is (2–3 sentences)

- My current focus and stage

- What I want Claude to help with

- What I want to retain as my own judgment (interpretive conclusions, research question)

- Any standing instructions (always save output to /output/, always flag uncertain claims, etc.)

Read the draft carefully. Edit anything that misrepresents the project. Save it as CLAUDE.md in the project folder. From now on, Claude reads this automatically at the start of every session.

Step 4: Set up the folder structure

Based on my project description and CLAUDE.md, suggest a folder structure.
Consider: where I'll put source materials, working files, output, notes,
and anything specific to this project type.

Create the folders Claude suggests, or a subset of them. You do not need every folder on day one — create what you will actually use this week.

Step 5: Create a start.md

Create start.md in this folder. It should describe:

- The project in 2–3 sentences (for my own reference)

- Current status (what stage I'm at)

- Immediate next tasks (what I plan to do in the next 1–2 sessions)

- Open questions (things I'm unsure about)

Update start.md at the beginning or end of each session to re-orient yourself and Claude.

Step 6: Run a first test task

Put one document in your source folder and run a task from C.prompts that matches your project type. Verify the output. Make one edit to CLAUDE.md based on what you learned.

What success looks like: A project folder with a CLAUDE.md that Claude reads automatically every session; a start.md you update regularly; a folder structure that matches your workflow; at least one completed task whose output is saved to disk.

At this point you are at Level 2 on the adoption spectrum — context-aware collaboration. → B.adoption-spectrum

Workflow 5: Syntopical reading — large-corpus field synthesis

Scenario: You have 30–80 PDFs — a field's key literature, a grant-funded corpus, or a systematic literature review. You want to understand the field as a whole: map the debates, identify the gaps, locate your own argument within the landscape. This is not triage (Workflow 1) and not deep-reading a short list. It is Mortimer Adler's syntopical reading — reading many books on the same question in parallel to synthesise what the field says, not just what each author says.

Why this is different from Workflow 1: Workflow 1 assumes 5–20 PDFs and one session. At 40+ PDFs the cognitive and context-management load changes qualitatively. You cannot hold the whole corpus in a single Claude session. Cross-source patterns emerge only after multiple passes. The synthesis you are building does not exist in any one document — you are constructing it. This requires a different infrastructure: intermediary documents, session continuity design, and a working theory of what questions the corpus should answer.

The two modes of large-corpus reading:

Mode Goal What you produce
Targeted search One specific question across the corpus (a concept, a method, a date range) A reference list with excerpts; a table of positions
Syntopical synthesis Understanding what the field is doing as a whole A field map: debates, camps, gaps, open questions

Targeted search is simpler: ask Claude to find where authors address your specific question, extract positions and quotes, and write a comparison. It is essentially Workflow 1 run repeatedly with a fixed question.

Syntopical synthesis is harder and is what this workflow addresses.

Folder structure:

/field-synthesis/
  CLAUDE.md                   ← your research questions, key terms, known camps
  /papers/                    ← all PDFs
  /papers_md/                 ← markdown conversions
  triage.md                   ← relevance table from initial pass
  /intermediary/              ← working synthesis documents (see below)
    themes.md                 ← emergent thematic map
    positions.md              ← author positions on key questions
    gaps.md                   ← questions raised but not answered
    session_log.md            ← what was done in each session
  synthesis.md                ← final output: your field account

The /intermediary/ folder is the infrastructure that makes multi-session work coherent. Without it, each session starts cold and you re-discover what the previous session already found.

Session 1: Convert and triage

Convert all PDFs as in Workflow 1, Step 1. Then:

Read all files in /papers_md/.
My synthesising questions are: [list 3–5 specific questions from CLAUDE.md].
These are not yes/no questions — they are questions where different authors take different positions.
For each paper:
1. Relevance (high / medium / low) with one-sentence reason
2. Which of my synthesising questions does it address?
3. Which 'camp' or theoretical tradition does it seem to belong to?
Write to triage.md as a table. Note anything unexpected.

At the end of Session 1, write the session entry in session_log.md:

Update session_log.md with a new entry for today:

- How many papers covered

- Anything that surprised you about the triage

- Which papers to deep-read first and why

- Any new questions that emerged that I should add to my synthesising questions

Sessions 2–N: Thematic passes

Rather than reading paper-by-paper, work thematically: take one of your synthesising questions and read across all papers that address it in a single session.

Read all files in /papers_md/ that triage.md marks as addressing [question X].
For each:
1. What position does this author take on [question X]?
2. What evidence or argument do they use?
3. Who do they agree or disagree with (by name)?
Read the current themes.md and positions.md before answering.
Update positions.md with a new section for [question X]:

- One row per author

- Columns: position, evidence type, agrees-with, disagrees-with

After each thematic pass:

Based on the new entries in positions.md for [question X]:
1. What are the 2–3 main positions authors take?
2. Is there a dominant view, or is the field genuinely divided?
3. What would it take to settle this question? What evidence is missing?
Update themes.md with a summary for [question X].
Update gaps.md with any questions raised but not resolved across these papers.
Update session_log.md with today's entry.

Each session advances one or two questions. The intermediary documents carry the state forward so the next session does not re-read.

Session N+1: Synthesis

When triage.md is fully covered and the major questions have been through thematic passes:

Read themes.md, positions.md, and gaps.md.
My own argument or research position is: [from CLAUDE.md].
Write a field synthesis to synthesis.md covering:
1. What questions this field is organised around
2. The main positions on each (with representative authors)
3. Where genuine consensus exists vs. where the field is divided
4. The gaps — what the field has not yet answered
5. Where my own work sits: which questions it engages, which positions it aligns with or departs from
Do not summarise individual papers. Write about the field as a whole.

Managing complexity at scale: what the intermediary documents do

The /intermediary/ folder solves the three hardest problems in large-corpus reading:

Context overflow. Claude cannot read 40 PDFs in one session. But it can read themes.md — a compact distillation of what those sessions already established. The intermediary files are how knowledge from Session 2 informs Session 7.

Re-discovery. Without session continuity, you read the same thing twice and ask the same questions again. session_log.md and the intermediary files are a shared memory between you and Claude across sessions.

Lost cross-source patterns. A connection between Paper 12 and Paper 34 does not exist in either file. It exists in positions.md once both papers have been processed. Cross-source insight requires a place to accumulate.

The intermediary documents are not outputs — they are scaffolding. They grow session by session and are eventually superseded by synthesis.md. Keep them rough; perfectionism at this stage slows you down.

CLAUDE.md design for a large corpus project

For syntopical work, CLAUDE.md needs more than a project description. Include:

## Synthesising questions
The 3–5 questions I am asking across the corpus.

## Known camps or traditions
Theoretical schools or methodological camps I expect to find.

## My own position
What I am arguing or studying. Claude should not contradict this — flag tension instead.

## Corpus status
Current state of triage.md and which questions have had thematic passes.
Update this at the start of each session.

The "Corpus status" field is especially important. Update it each session so that Claude (and you) know at a glance where the project stands.

What success looks like: A synthesis.md that describes the field as a whole — debates, positions, gaps — without being a series of summaries. positions.md and themes.md as a durable reference for writing. An updated CLAUDE.md that situates your own argument within the field landscape. A session_log.md that you could hand to a collaborator to bring them up to speed.

The difference between this workflow and reading 40 papers with a highlighter is not speed — it is structure. The intermediary documents force the field-level questions to stay in view across weeks of work, and the thematic passes ensure that cross-source comparisons actually get made rather than remaining potential.

Related