Generating Experiment Reports with Claude (Shuhan)

Claude can go beyond writing code — it can design and run experiments, monitor results, and generate visual reports from raw data.

The Pipeline

  1. Run experiments — Claude manages execution, monitors logs, and loops on pass/fail checks
  2. Collect results — Experimental outputs are written to structured markdown files
  3. Generate reports — Raw data is transformed into SVG visualizations

You can define a custom skill so that Claude generates SVG reports consistently every time.

Workflow in Detail

1. Read Raw Data

Start by pointing Claude at your experimental output.

2. Structure the Report

This is the step that requires the most human input — Claude should not decide what goes into the report on its own.

  • Discuss with Claude: how many sections/panels do you want?
  • Verify: ask Claude to summarize the data and propose panels
  • Iterate on the structure through conversation

3. Build the Visuals

  • Spin up a separate visual agent to render and review the SVG output
    • Prompt pattern: “Use a subagent to verify the SVG looks correct”
    • Provide an example of the desired output for reference
  • Individual visual components are typically one-shot — Claude handles them well
  • Layout and composition requires human guidance

4. Audit for Correctness

  • Use 3 subagents with clean context to independently verify results
  • Check coherence against the original paper implementation and raw data
  • Always audit the final output yourself

Key Takeaways

  • Claude can’t decide what you want in the report — the report structure needs significant human direction
  • Building skills takes iteration — keep discussing and refining with Claude
  • Verification strategies:
    • Use Deep Research to cross-check against papers
    • Have Claude read and explain the code it wrote
    • Ask the agent to explain what it built and why

Q&A

How do you create custom skills?

  • Through iterative refinement — keep discussing with Claude until the skill does what you need reliably.

How do you ensure report correctness?

  • Cross-reference with papers via Deep Research
  • Have Claude inspect its own code
  • Ask the agent to explain and justify its output

How do you fork from a current chat?

  • Use /fork or /btw to branch off a conversation
  • You can also read Claude session files directly

Working with Math in PDFs (Philipp)

Claude can parse and reason about mathematical content from PDFs. Here’s a reliable workflow:

  1. Convert — Ask Claude to create a markdown version of the PDF
  2. Summarize — Have Claude summarize the PDF content in markdown
  3. Verify understanding — Ask questions about the math to test comprehension
  4. Leverage notation — Let Claude handle all the notation formatting and manipulation

This approach works well for reading papers, understanding derivations, and working with mathematical content programmatically.