The Prompt Lab — Counterfactual Anchoring
Learn the counterfactual anchoring prompting technique with concrete before/after examples.
Counterfactual Anchoring
The Technique
Counterfactual Anchoring asks the model to first articulate what would have to be true for a different conclusion to be correct before committing to its answer. This forces the model to stress-test its reasoning against an alternative reality, surfacing assumptions it would otherwise skip. The result is more calibrated, defensible output — especially valuable when the stakes of being wrong are high.
The Naive Prompt
We're a mid-sized SaaS company (~$8M ARR) considering moving our
entire data pipeline from Snowflake to DuckDB to cut costs.
Should we do it?Why It Falls Short
A model given this prompt will typically pattern-match to “here are the pros and cons of DuckDB vs. Snowflake” and produce a balanced but ultimately non-committal answer. It won’t interrogate the assumptions baked into the question — that cost is the real bottleneck, that the migration is feasible at your scale, or that DuckDB actually solves your specific problem. You get a generic technology comparison when you need a decision.
The Improved Prompt
We're a mid-sized SaaS company (~$8M ARR) considering moving our
entire data pipeline from Snowflake to DuckDB to cut costs.
Before giving your recommendation, answer this first:
What would have to be true about our situation for this migration
to be a BAD idea — even if DuckDB is technically cheaper?
List the conditions that would make staying on Snowflake correct.
Then, given those conditions, give your actual recommendation
and flag which of those "stay on Snowflake" conditions I should
verify before proceeding.Why It Works
By asking the model to construct the opposing case first, you force it to name the hidden assumptions — things like concurrent user load, multi-region replication needs, existing BI tool integrations, or engineering bandwidth for migration. The final recommendation then arrives after those landmines are on the table, so it’s contextualized rather than generic. You also get a built-in due diligence checklist, which turns a chat answer into something you can actually act on.
When to Use This
- High-stakes binary decisions where a wrong call has real costs — vendor selection, hiring choices, architectural pivots, pricing strategy changes
- Any time you suspect the question itself might be wrong — when you want the model to challenge the framing before answering it (pairs well with Claude Opus 5, which tends to engage substantively with the adversarial scaffolding rather than glossing over it)
- Pre-mortem style analysis before pitching a plan internally — use this to surface the objections your skeptical CFO or CTO will raise, so you’ve already addressed them