What a model can sort, and what stays yours
Research notes look messy because thinking is messy. You read a paper, copy three lines, add a thought in the margin, save a link to check later, and the file becomes a mix of facts, guesses and questions. A model can help you organize that pile, but only when you give it something to organize around.
Start by writing down what the notes are for. A literature review, a product decision, a talk outline and a study guide each need different fields and different levels of detail. If you stay silent about the purpose, the model will guess, and it will guess differently on every pass.
Think of the model as a careful assistant at a sorting table. It can group similar items, suggest labels, spot duplicates, turn scattered bullets into an outline and flag claims that look unsupported. It cannot invent sources you never wrote down, and it should not decide which claim is true. Traceability and judgment stay with you.
One habit pays off immediately: keep your raw notes untouched. Copy them into a working file and let the model restructure the copy. Your original wording, quotes and page numbers stay safe, and you can always compare the tidy version against the raw one.
Give every note the same five fields
The fastest way to keep notes consistent is to give every entry the same shape. A template removes a hundred small decisions, and it gives the model clear slots to fill. Five fields are enough for most beginner projects:
- Source: where the note came from, with author, title, date and page or section.
- Claim: the single idea the note supports, written in one sentence.
- Evidence: the quote, number or example behind the claim, kept verbatim.
- Interpretation: what you think it means, clearly marked as your own reading.
- Open question: what is missing, uncertain or worth checking next.
The boundary between evidence and interpretation matters more than any label. Models are good at paraphrase, and paraphrase is where subtle meaning gets lost. If a quote matters, paste it exactly as written and let the model work around it instead of rewriting it.
Use plain language inside the fields too. Short sentences beat clever jargon. When a field is empty, write unknown rather than leaving a gap, because an empty field invites the model to fill it with something that sounds reasonable but was never measured or sourced.
Once the template is set, save it as a reusable text block and paste it at the top of every new note file. From then on, every session starts in the same place.
Run the first pass in four moves
With a template in place, your first run becomes a short sequence of moves. Do them in order, and pause between moves to check the output before continuing.
- Dump: paste your raw notes into the working file, unchanged and in whatever order they arrived.
- Sort: ask the model to group the notes by theme and to name each group with a plain phrase.
- Tag: ask for the five fields on each note, and to mark anything it cannot fill as unknown.
- File: ask for a compact outline you can paste back into your own document, using the group names as headings.
Write the request as a checklist instead of a conversation. Something like: group these notes, name the groups, fill the five fields, keep quotes verbatim, mark unknowns, and output an outline. A short, explicit request is far easier to verify than a long, chatty one.
Work in batches. A hundred notes in one go produce a wall of text that is hard to check, while twenty notes at a time let you read every field. When a batch looks wrong, fix the instructions and run that batch again instead of editing the output by hand and hoping the next batch behaves.
Keep model output separate from your own writing. Mark generated sections clearly in your file so that later you know which words were yours and which were assembled from your notes.
The checklist before you trust the file
Sorting is quick; trusting the sorted file takes a few minutes of checking. Run this checklist before the green light, every time:
- Every claim points to a source you can name and find again.
- Every quote is verbatim, with a page or section number beside it.
- Nothing appears that was not in your raw notes.
- Uncertain items are flagged as open questions instead of smoothed into confident sentences.
- Group names describe the content instead of sounding impressive.
- The same term is used for the same idea everywhere in the file.
- The outline matches the purpose you wrote down in the first step.
The third check is the one people skip. Models compress, and compression can quietly add a link between two notes that you never made. Read the outline out of order and ask, for each sentence, whether you can point to the note behind it.
When a check fails, change the instruction rather than the output. If quotes drift, add a rule that quoted text is never rewritten. If unknowns get filled in, add a rule that uncertain fields stay unknown. Over a few runs, your instruction becomes a small rulebook that fits the way you actually work.
Keep the template alive for the next sprint
A structure that works once is worth keeping. Save your template, your sorting request and your checklist as a single starter file, then open it at the beginning of every new project. You get the same clean starting point without rebuilding it from memory.
Match the model to the move. A small, quick model is handy for tagging a large batch of short notes, while a stronger reasoning model earns its keep on synthesis, cross-group comparison and writing the final outline. A pair such as Gemini 2.5 Flash for the sorting pass and Claude Sonnet 5.5 for the careful read-through works well for a newcomer.
SwiftRelay fits this workflow nicely because it is pay-as-you-go and billed per token, so a short organizing pass costs what it costs and nothing more. You get Claude, GPT, Gemini, DeepSeek, Qwen, Grok and other models through one API at https://api.fastapirelay.com, and it works with Claude Code, Codex, Gemini CLI and any OpenAI-compatible SDK. New accounts also receive a small gift credit to help you start.
Make the sign-up your pit stop before the next lap: create a SwiftRelay account, paste your template, run a batch of ten notes, and see how the checklist feels. Then fix the template on the second batch and keep the better version.
After two or three runs, you stop thinking about the format and start thinking about the research again. That is the real goal. The model handles the sorting, you keep the judgment, and every new project begins one step closer to a file you can actually use.