When one mental model leaves a material blind spot on a multi-domain or high-stakes problem, sequence complementary models with named roles and a conflict rule.
SkillScore91Maintained
On-demandOn-demand: invoked explicitly — never occupies your agent's context until called.
◆Research pack
Last verified 8/14/2026
What it changes
Passed 21 of 23 test cases with the skill, and 7 of 23 without it.
Without skill30%
With skill91%
Checks it passed that the model alone missed
The pattern field specifies nested.
The pattern field is set to parallel.
The pattern field is set to sequential.
and 18 more in the full report
14 fixed7 passed either way2 still failing0 broke
Pass-rate lift
+61 pp
Tokens
-1%
Agent turns
0%
Response time
-33%
Model gemini-3.6-flashMeasured 8/14/2026v121 of 23 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 14 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 14
Prompt
We began a multi-model analysis on choosing an off-the-shelf email dispatch vendor. Step 1 analyzed compliance using a vendor security assessment model, which fully answered all security and regulatory compliance requirements and revealed zero unaccounted risks. Provide the structured output report with keys problem, gap, pattern, models, conflict_rule, convergence, divergence, recommendation, and stop_reason.
What was checked, without and with the skill
Checked
Without skill
With skill
The stop_reason field is set to single_model_suffices.
Missed
Passed
The response recommends making the decision using the single sufficient model.
Missed
Passed
case-23 · 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.md49 linessingle file
Model Combination
Core rule: Combine only when each model answers a different named question. Cap at three, name the conflict rule before applying, then synthesize once.
When to Use
One model already applied (or clearly primary) still leaves a material blind spot that another mechanism covers.
Problem spans domains (e.g. risk + choice + system structure) and stakes justify multi-lens work.
You need independent checks, not confirmation of the same conclusion.
You can name a distinct role per model before running them.
When NOT to Use
A single catalog skill fully answers the unknown — apply that skill alone.
Routine, local, or fully reversible work where multi-lens cost exceeds upside.
You cannot state what unique question each extra model answers (checkbox / model soup).
Near-duplicate mechanisms (two diagnosis skills that ask the same causal question).
Time budget cannot support genuine synthesis — prefer one honest model over contradictory partials.
Procedure
State the unknown and the gap. Write the decision question. If one model already covers it, stop and use that model alone. Otherwise name the specific blind spot (e.g. "failure modes unexamined", "displaced alternative unknown").
Pick 2–3 models with distinct roles. For each, record: model id, role (narrow / decide / stress / cost / …), and the unique question it answers. Drop any model that only rephrases another. Prefer sequential pipeline (narrow → stress → decide) over parallel unless independent concurrent checks are required.
Lock the relation and conflict rule before applying. Choose pattern: sequential, parallel, nested (macro→meso→micro), or adversarial (for/against). Predeclare the tiebreaker (e.g. reversibility class, evidence strength, ruin constraint, primary decision owner). Incompatible worldviews run sequential or adversarial — never blended.
Apply each model fully for its role only. Capture one key insight per model plus what only that model revealed. Do not re-run a model that adds no new insight.
Synthesize once. Record convergence, divergence, how the conflict rule resolves divergence, and a single combined recommendation with residual uncertainty. Stop when the recommendation is decision-ready or when further models would only reconfirm.
Before committing to a plan or launch, assume it already failed and reason backward through concrete causes — convert failure paths into mitigations, gates, and stop checks.