Español

Force Ambiguous Choices Into a Strict 3-Line Brief

When your AI hits an ambiguous choice, force it into three lines.

Ask an assistant about a trade-off — which library, which schema, ship now or wait — and you usually get three balanced paragraphs that end in "it depends." Technically correct, practically useless. You still have to decide, and now you have to read all of it first before you can. The model hedges because hedging is safe and complete-sounding, but the hedge pushes every ounce of decision work back onto you.

Why "it depends" is a cop-out

An ambiguous choice is exactly where you wanted the model's help. Instead it optimizes for not being wrong: it lists considerations, weighs both sides evenly, and declines to commit. That reads as thorough. It's actually the model refusing the hard part of the job — picking. A recommendation you can override in five seconds is worth more than a balanced essay you have to adjudicate yourself, because the recommendation gives you a default and a reason to argue with.

The fix: demand the decision-brief format

When a choice is ambiguous, require exactly three lines and forbid hedging:

Ambiguous choice? Respond in EXACTLY three lines, no hedging, pick one:

1. ELI10:  Explain the choice like I'm ten. What are we actually deciding?
2. STAKES: Blast radius (what it touches) and reversibility (how hard to undo).
3. REC:    One recommendation + a one-line reason. You MUST pick one.

An example, fully committed:

1. ELI10:  Do we store dates as text or as real timestamps?
2. STAKES: Touches every query that filters by date; switching later means a
           migration. Reversible now, expensive to reverse after launch.
3. REC:    Use real timestamps — text dates break sorting and cost you a
           migration the moment you need a range query.

You get the choice, the risk, and a clear pick in about ten seconds. You can still override it — but the agent has to commit to a recommendation first, which is the whole point. A named default is something you can react to; "it depends" is something you have to unpack.

Power moves

• Force the pick, allow the override. "You must recommend one, even at 55/45 confidence" beats a refusal. Add "and state your confidence" if you want the closeness surfaced without reopening the hedge.
• Make it a standing rule for agents. Put the three-line format in the system prompt so every ambiguous fork resolves the same way instead of ballooning into essays.
• Cap the reason at one line. The value is the commitment plus a single load-bearing reason. More prose reintroduces the wall of text you were escaping.
• Escalate only real ambiguity. Reserve the brief for genuine trade-offs. Trivial choices don't need it, and unclear requirements should trigger a question, not a forced guess.

Resources

• OpenAI — Prompt engineering guide
• Anthropic — Prompt engineering overview
• Architecture Decision Records (ADR) — Michael Nygard
• The Cognitive Style of PowerPoint — Edward Tufte