Stop pasting your whole spreadsheet into the chat
Pasting raw tables into ChatGPT or Claude eats context and makes the model lose track. Here's what to send instead, with a worked example from a nutrition study.
TOOLS IN AIBIOLOGY
Dhruv Aditya Srivastava
11/21/20263 min read
Here's a habit I see in almost every lab that has started using AI for data work, and I've done it plenty myself. You have a table, maybe a few hundred rows of ICP-MS element concentrations or a chunk of an RNA-seq count matrix. You copy it, paste it into the chat, and type "help me analyze this."
It works at first. Then, fifteen messages later, the answers get slower and vaguer. The model forgets that you already dropped the outlier plant, or suggests the normalization you rejected an hour ago. Sometimes it starts quoting numbers that aren't in your data at all.
There's a fairly simple reason for this, and a fairly simple fix.
Why it goes wrong
A chat model has no memory between turns the way you do. Each time you send a message, the whole conversation so far, including that 500-row table, goes back into the model as input. So you're paying (in tokens, time and sometimes money) to have it re-read the same rows on every turn.
Long inputs also get used unevenly. Liu and colleagues found that language models do best with information at the start or end of a long input and worse with information in the middle, even in models built for long contexts (Liu et al., 2024, TACL). Your original question and your early decisions sit at the start of the chat. After enough back-and-forth, the bulky table and the rest of the conversation push them into that weaker middle zone.
If you run a model on your own computer, this hits even sooner. Local models often have small context windows, and one pasted table can fill most of it.
What the model actually needs
For most analysis questions, the model doesn't need your rows. It needs to know what the data looks like and what you're trying to do. That fits on one page:
Shape. Number of rows and columns, and what one row is (one plant? one sample? one gene?).
Columns. Names, types and units. "Zn, numeric, mg/kg dry weight" tells the model far more than 300 actual zinc values.
Design. Treatments, replicates, blocks and batches. For RNA-seq, say which samples go with which condition, and whether you have paired samples or batch effects.
A few example rows. Three to five, so the model sees the real format, including any messy bits like "<LOD" or blank cells.
Your goal. One or two sentences. "Test whether the mycorrhizal treatment changes shoot P and Zn, accounting for block" is something a model can work with. "Make this data look better" is not.
What you've already decided or ruled out. "Plant 14 removed (root rot). Not using TPM for this. Tried log transform, residuals still skewed."
Point 6 is the one people skip, and it's the one that stops the model from going back over choices you've already made.
A quick example
Instead of pasting a 60-person diet study table, I'd send something like this:
60 rows, one per participant. Columns: participant_id (text), diet (high-fibre / control), site (A / B / C), sex (F / M), age (years), and 5 outcomes measured at week 0 and week 12: fasting glucose (mmol/L), fasting insulin (mU/L), HbA1c (%), LDL cholesterol (mmol/L) and body weight (kg). Two-arm randomised trial, 30 per arm, randomised within site. Missing week-12 values are left blank, mostly insulin. Participant 17 withdrew in week 3, so 59 people in the analysis. Goal: does the high-fibre diet change the week-12 outcomes, adjusting for the week-0 value? I'd like a linear model per outcome with site as a random effect, and a forest plot of the effect sizes. Here are 4 example rows: ...
That brief is a few hundred words, not thousands of numbers. The model writes the R or Python code, and you run it on the full file on your own machine. Your data never has to go into the chat at all, which also helps if it's unpublished or sensitive.
Carry the brief, not the chat
The brief has one more use. When a conversation gets long and muddled, start a fresh one and paste in the updated brief, with your new decisions added to the "already ruled out" list. The same trick works if you want to switch tools, say from Claude to ChatGPT, or to compare how two models handle the same task. A one-page brief moves between them easily, and a 60-message chat history doesn't.
I got tired of writing these briefs by hand, so Nishtha Srivastava and I built a small open-source skill for it called analysis-brief. You give it the pasted table, log or long thread, and it keeps the schema, units, replicate structure, decisions and dead ends, and drops the raw rows. If your request is vague, it asks a few questions first. If it's already clear, it gets out of the way. One limit: it's meant for text you paste in. Files you upload are usually handled more efficiently already.
You don't need the tool to use the idea, though. Next time you reach for Ctrl+V on a big table, spend two minutes writing the six points above instead.
Source
Liu, N. F., et al. (2024). Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics, 12, 157-173. https://doi.org/10.1162/tacl_a_00638
Insights
Explore AI news, expert articles, and workshops.
Connect
Learn
© 2026 The AI Advances. All rights reserved.
