The Prompt Lab — Output Scaffolding
Learn the output scaffolding prompting technique with concrete before/after examples.
Output Scaffolding
The Technique
Output Scaffolding means embedding the literal structure of your desired response directly into the prompt — headers, labels, placeholders — so the model fills in a pre-built skeleton rather than inventing its own format. This works because frontier models are strong completers: give them a partial document and they’ll match its register, depth, and organization rather than defaulting to generic patterns.
The Naive Prompt
Write a competitive analysis of Notion versus Linear for a B2B SaaS startup
that's choosing a project management tool.Why It Falls Short
Without structural guidance, the model decides what to compare, in what order, and at what depth — and it usually defaults to a breezy, surface-level summary with vague pros/cons lists. You’ll get “Notion is flexible while Linear is more opinionated” observations that don’t map to your team’s actual decision criteria. You then spend time reformatting the output anyway, defeating the purpose of using the model.
The Improved Prompt
Write a competitive analysis of Notion versus Linear for a 12-person B2B SaaS
startup (engineering-heavy, no dedicated PM yet) choosing a project management tool.
Use exactly this structure:
## TL;DR Recommendation
[2-sentence verdict with a clear winner for this specific context]
## Decision Criteria
| Criterion | Notion | Linear | Winner |
|-----------|--------|--------|--------|
| Engineering workflow fit | | | |
| Docs + specs in one place | | | |
| Onboarding time for non-PMs | | | |
| Pricing at 12 seats | | | |
| API / integration depth | | | |
## The Case For Notion
[3 bullet points, each starting with a concrete scenario: "When your team..."]
## The Case For Linear
[3 bullet points, same format]
## What Would Make Us Switch the Recommendation
[2 conditions — one that would flip to Notion, one that would flip to Linear]Why It Works
The scaffold forces the model to address your criteria rather than generic ones — pricing at 12 seats, onboarding for non-PMs — so the output is immediately actionable rather than editable. The table format creates natural parallelism, meaning the model must evaluate both tools on identical axes instead of drifting toward whichever product it has more training signal on. The conditional “What Would Make Us Switch” section is particularly valuable: it’s the section a naive prompt would never generate, but it’s often the most useful part of any real analysis.
When to Use This
- Recurring deliverables: If you’re generating the same type of document repeatedly (weekly status reports, customer case studies, technical specs), define the scaffold once and reuse it — you’re essentially building a template engine powered by a language model.
- Decision documents: Anytime a human needs to act on the output rather than just read it, scaffolding prevents the model from burying the lead or omitting the specific dimension the decision hinges on.
- Taming verbose models: Claude Opus 4.6 and GPT-5.4 are both prone to thorough-but-sprawling responses on open-ended tasks. Scaffolding is one of the fastest ways to get dense, structured output without fighting the model’s default verbosity through constraint alone.
Next edition: we’ll look at Stepwise Decomposition — breaking a single complex prompt into an explicit reasoning chain the model executes in sequence, and why it outperforms asking for the same result in one shot.