I send one paper section at a time to keep costs down

Start with the question you need the paper to answer

A paper summary tells you what the authors say. A reading guide helps you decide what to read, what to question, and what to understand before moving on. That distinction matters when you pay for each AI request: asking for a complete explanation immediately can produce a long answer before you know whether the paper is relevant.

Begin by writing your own purpose in one sentence. Perhaps you want to reproduce an experiment, understand a training method, or judge whether a result supports your project. Add your background and available reading time. These details give the model a useful target without requiring the whole paper.

Extract readable text before spending tokens

Use a lawful copy of the paper and check whether its text extracts cleanly. Two-column layouts can scramble paragraphs, and equations or figure captions may disappear. Inspect a sample before sending anything. Do not assume a model can read a paper from its title or retrieve it from a citation.

Keep section headings, page numbers, and figure labels where possible. They will help you verify the guide later. Remove repeated headers and obvious extraction clutter, but preserve qualifications and limitations. If the paper contains private or unpublished material, confirm that you are allowed to send it to the service you use.

Build a reading map from the abstract and headings

For the first request, send the title, abstract, section headings, and your purpose. Ask the model to propose a reading order, identify likely prerequisites, and list questions that require checking the full text. Explicitly tell it not to infer experimental details from the abstract.

Keep the requested answer short, perhaps 400 words. A model such as Gemini 2.5 Flash can be a starting point for this limited task, but judge the result rather than assuming a model name guarantees accuracy. The output should be a provisional map, not a verdict on the paper.

Expand only the sections your purpose requires

Follow the map and send one relevant section at a time. Include your purpose again, along with only the earlier notes needed to interpret that section. This avoids repeatedly sending the entire paper or carrying an unnecessarily long conversation forward.

Ask for a plain-language explanation, the section's main claim, its supporting evidence, and two questions to answer while reading. Require the model to distinguish the authors' statements from its own interpretation. For technical passages, ask it to preserve notation and flag missing definitions rather than silently filling gaps.

If an explanation depends on a figure or equation that did not extract correctly, provide that material through a supported input method or inspect it yourself. A confident explanation of missing evidence is not a useful shortcut.

Turn explanations into checks you can perform

Once you have covered the essential sections, ask for a compact guide organized around reading actions. Each action should point to a section or page and give you something observable to check.

• Explain the research question in your own words before reading the method.
• Identify the assumptions that make the method applicable.
• Find the comparison that supports the central result and note its evaluation conditions.
• Write down one limitation and explain whether it affects your intended use.

Ask for a small glossary only for terms that block understanding. A long glossary can look productive while adding cost and little learning value.

Verify the guide and stop when it serves its purpose

Check the central claim, one important numerical result, and every source reference in the guide against the paper. Section references can be wrong, and a real number can still be attached to the wrong experiment. Mark unresolved questions instead of asking the model to make the guide sound complete.

MyRelay gives you one key and one budget across supported models, with tokens and cost visible for every request. Create a MyRelay account if you want to track this workflow's spending request by request. Use that visibility to notice unnecessary retries and oversized inputs.

Stop when you have a reading order, trustworthy checkpoints, and a short list of uncertainties. The guide has done its job when it helps you read the paper yourself, not when it replaces the paper with another long document.