Analyze data
ResearchCoding
Splits getting the code from interpreting the output, so the model cannot invent results it never ran.
Fill the blanks and it copies as you filled it
Step 1 [Upload the data file] - What I want to know: - About the data: - Give me one runnable block of code and don't interpret the results yet Step 2 — after I've actually run it [Paste the output] Interpret the output above in three sentences or fewer.
0 of 2 blanks filled · anything left blank copies with its brackets
This splits getting the code from interpreting the result. If you ask a model to interpret output it never actually ran, it can invent the numbers.
What this prompt does
Compared with asking the same thing without the length and format lines.estimate
Estimated output tokens
700
Asked loosely
180
This prompt
Energy (Wh)
1.215
Asked loosely
0.423
This prompt
Token counts are estimated from character counts, not measured. Energy uses the per-band figures measured by Jegham et al. 2025 as published.
What that saves
15.8 Wh · 5.54 gCO₂e · 74 mL — assuming 20 uses a day
Tap a card for how the conversion was done.
Reported backfire — Specifying output lengthSpecifying a length longer than needed backfires (a 34% increase in 11% of scenarios)
Where these figures come from
We read the prompt to see whether it carries that instruction. The percentages below are published figures for the rule itself, not a measurement of what this prompt saves.
Assigning a role
44%mean reduction · 95% CI 29–59%
Podder et al. 2026 BP#4 · 19 use cases / 61 prompts
by modelGPT-4o 34%Command R+ 50%Mistral-7B 49%
Specifying output length
38.7%mean reduction · 95% CI 29.7–47.7%
Podder et al. 2026 BP#5 · 18 use cases / 57 prompts
by modelGPT-4o 35%Command R+ 49%Mistral-7B 32%
Reduced in every scenario. No backfire reported