When a real test is too rare, large, or irreversible, run a controlled counterfactual: isolate one variable, fix conditions, trace the mechanistic chain, and bound what the result implies.
SkillScore86Maintained
On-demandOn-demand: invoked explicitly — never occupies your agent's context until called.
◆Research pack
Last verified 8/14/2026
What it changes
Passed 19 of 22 test cases with the skill, and 14 of 22 without it.
Without skill64%
With skill86%
Checks it passed that the model alone missed
Declines the analysis due to multiple unisolated variables
Directs stopping the thought experiment to run the actual test
Rejects the simultaneous failure scenario due to multiple variables
and 4 more in the full report
7 fixed12 passed either way1 still failing2 broke
Pass-rate lift
+23 pp
Tokens
+1%
Agent turns
0%
Response time
-36%
Model gemini-3.6-flashMeasured 8/14/2026v119 of 22 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 7 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 7
Prompt
We have a staging cluster identical to production and a load testing suite ready to run a 10-minute traffic spike to measure database connection pool exhaustion. Provide a step-by-step thought experiment analyzing connection behavior under load.
What was checked, without and with the skill
Checked
Without skill
With skill
Rejects performing a thought experiment in favor of executing the available load test on staging
Missed
Passed
case-02 · gemini-3.6-flash · 8/14/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md47 linessingle file
Thought Experiment
When empiricism is out of reach, run a disciplined counterfactual: one isolated change, fixed conditions, step-by-step mechanism, and a hard bound on implications.
When to Use
You need behavior under failure, scale, or policy you cannot cheaply trigger or measure (region outage, 100x load, one-way architecture).
A decision is expensive or irreversible and a mental trace can surface break points before commit.
Edge cases are too costly to stage, but a mechanistic chain can still expose missing controls.
When NOT to Use
A cheap real test exists (load test, flag, query, spike) → run the test; do not substitute imagination.
Adversarial security attack-path work → use red-team structure, not free-form scenarios.
You already know the mechanism and only need a decision under known facts → decide; do not dramatize.
Vague "what if everything" brainstorming without a single isolated variable → tighten or stop.
Procedure
State the question and isolation. Name exactly one primary variable or counterfactual change. Freeze all other conditions as the control world. Reject multi-variable "and also" scenarios.
Fix initial conditions. Specify system state, load, configuration, actors, and what is not changed. Write values concrete enough that another agent could replay the setup.
Trace the mechanism step by step. From t0, record what fails, queues, retries, or adapts next—and why—using known components and policies only. No hand-wavy "then everything collapses"; each step needs a causal link.
Extract invariants and break points. Note what still holds (invariants) and the first step where the system violates a requirement (capacity, correctness, safety, UX). Mark assumptions that, if false, void the chain.
Bound implications. Map insights only to actions or checks justified by the chain (limits, guards, monitoring, redesign). Label speculative leaps beyond the isolation as out of bound.
Name a discriminating real check, then stop. For the weakest link, state the cheapest observation or experiment that would confirm or kill it. Stop after one controlled chain with bounded implications; if a link is cheaply testable now, exit to that test instead of further imagination.
Output
Emit a thought-experiment record:
question: what behavior or decision is under test
isolated_variable: single change vs control world
initial_conditions: frozen state and non-changes
consequence_chain: ordered mechanistic steps
invariants: what still holds
break_points: first requirement failures and critical assumptions
implication_bound: actions/checks justified by the chain only
discriminating_check: cheapest real observation to confirm or kill the weak link