How to use the What-If Oracle agent skill for research planning

What the skill makes the agent do

Once the skill is active, the agent works through a research or project decision in four phases: it frames one operational question, explores a bounded set of scenarios, analyses and challenges each one, and then compares candidate responses. The scenario lenses are favorable (best case), reference (baseline), adverse (worst case), wild card (a concrete disruption outside the main assumptions), contrarian (an alternative to the dominant hypothesis), and second order (downstream effects and feedback).

For each scenario the agent records the assumptions, the supporting and conflicting evidence, the mechanism as a conditional narrative, the observable trigger, and the planned response. The skill is explicit that these are narratives, not forecasts. It does not estimate causal effects or calibrated probabilities, and it tells the agent not to invent them.

• Frames the decision, comparator, perturbation, horizon, outcome and baseline
• Explores several distinct scenarios, usually four to six
• Records an evidence ledger that labels each claim as observed, model-derived or assumed
• Defines triggers with a measure, threshold, review time, owner and action

Install the skill from the download ZIP

Download the ZIP for the what-if-oracle skill from this page. The skill folder contains SKILL.md and a references folder with scenario-templates.md. Unzip it so that the what-if-oracle folder sits in one of two places.

For a personal setup available across your projects, put it in your user skills folder, ~/.claude/skills/. For a single project, put it in that project's .claude/skills/ folder so it applies only there and can be shared with collaborators who use the same repository.

• Personal: ~/.claude/skills/what-if-oracle/
• Project only: .claude/skills/what-if-oracle/ inside your project folder
• Claude apps: open Settings, go to Capabilities, then Skills, and upload the skill there

Trigger the skill with a natural request

You do not need a special command. Describe a research decision and ask the agent to test it under uncertainty. The skill's description covers research planning, experimental contingencies and scientific project decisions, so requests in that territory are the ones it is built for.

Give the agent the decision, what would change, and what success looks like. The more concrete the question, the more useful the scenarios. If you want a specific format, say so, for example a scenario table with triggers and owners.

• Stress-test my cell culture study plan for reagent shortages
• What if sample attrition is higher than planned? Compare scenarios and suggest contingencies
• Help me write the risk section of my proposal using best, worst and wild-card scenarios
• Which of these two protocols stays feasible if the new instrument cannot be booked?

Example: a yield question

Suppose you plan a batch of 100 samples and expect 80 usable ones, but suspect the usable fraction could fall to 60 percent. A good prompt states the decision, the baseline, the range, the deadline and your capacity. The skill's template for this case asks the agent to check the experimental unit (participants, independent biological samples, technical replicates or batches) and to note whether missingness differs between groups.

The output should compare favorable, reference, adverse and second-order scenarios. A reagent problem, for instance, might lower yield and raise rerun costs at the same time. Both fractions remain planning assumptions until your own data support them.

What it needs, and safety notes

The source lists no required packages, API keys, credentials or network access. The skill runs no scripts and uses no network connection, so it works as plain instructions the agent follows. The qualitative workflow needs nothing beyond the agent itself.

Treat the output as exploratory. The skill says scenario prose cannot replace a power calculation, a causal estimate or a numerical simulation, and it is not meant for general business, personal or software architecture decisions. Check any external fact that matters, and do not present a generated scenario as an observation or an expert assessment.

• Requires: nothing listed in the source
• Runs scripts: no
• Uses network: no
• Use for power calculations, causal effects or simulations: no, use a proper design or model instead