Design experiments and quasi-experiments before analysis.
SkillScore86Maintained
Automatic
◆Research pack
Last verified 8/13/2026
What it changes
Passed 19 of 22 test cases with the skill, and 16 of 22 without it.
Without skill73%
With skill86%
Checks it passed that the model alone missed
Declines to debug software execution or GPU memory stack traces
Constructs a reduced-scale controlled trial on a small compound sample to isolate assay conditions
Audits software execution and data input pipelines to rule out implementation flaws prior to modifying core loss function math
3 fixed16 passed either way3 still failing0 broke
Pass-rate lift
+14 pp
Tokens
-1%
Agent turns
0%
Response time
-29%
Model gemini-3.6-flashMeasured 8/13/2026v119 of 22 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 3 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 3
Prompt
My PyTorch training script failed with `RuntimeError: CUDA out of memory. Tried to allocate 2.00 GiB (GPU 0; 11.17 GiB total capacity)`. Here is my training loop code. Please debug the memory leak and fix the script.
What was checked, without and with the skill
Checked
Without skill
With skill
Declines to debug software execution or GPU memory stack traces
Missed
Passed
case-22 · gemini-3.6-flash · 8/13/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md37 linessingle file
Designing Experiments
Helps choose and specify a research design before data analysis starts. This skill owns study-design decisions: what is treated, what is compared, what outcome is measured, which assumptions are required, which validation or recovery experiment should follow a failed scientific experiment, and which design is defensible.
It does not fit causal models, estimate treatment effects, interpret fitted model output from existing data, or debug software/build failures.
Decision Framework
Control Group?
Yes: Go to Step 2.
No: Consider Interrupted Time Series (ITS).
Unit Structure?
Single Treated Unit:
With multiple controls: Synthetic Control (SC).
No controls: ITS.
Multiple Treated Units:
With control group: Difference-in-Differences (DiD).
Time Structure?
Panel Data (Multiple units over time): Required for DiD and SC.
Time Series (Single unit over time): Required for ITS.
Method Quick Reference
Difference-in-Differences (DiD): Compares trend changes between treated and control groups. Assumes Parallel Trends.
Interrupted Time Series (ITS): Analyzes trend/level change for a single unit after intervention. Assumes Trend Continuity.
Synthetic Control (SC): Constructs a synthetic counterfactual from weighted control units. Assumes Convex Hull (treated unit within range of controls).
Failed Experiment Recovery
When a scientific experiment or optimization plan produces weak or contradictory results, use the same design surface to:
Separate implementation or measurement errors from design-assumption failures.
Identify which assumption should be tested next.
Define a minimal validation experiment before abandoning the approach.
State the decision rule for continuing, revising, or stopping the line of work.