PTN—11
Structured Input
Replace part of the free-text prompt with fields, templates, or forms that guarantee the request arrives complete.
Free text is maximally flexible and minimally guiding. When a task has a known shape — a translation needs a target language, a trip plan needs dates and a budget — a blank box makes people guess what to include, and the model pays for every omission with a guess of its own. Structured input gives the request a skeleton: fields for what is required, prose for everything else.
Structure is also how expression scales beyond one-off requests. A form-shaped prompt is repeatable, shareable, and auditable in ways a paragraph never is — the difference between a colleague’s clever prompt you paste and mutate, and a template with three labeled slots anyone can fill correctly on the first try. This is the mechanism behind Notion AI’s block-level actions and every serious internal prompt-template system: the expertise lives in the structure, not in each user.
The failure mode is over-structuring. Fields are a commitment to a task shape, and the moment someone’s intent falls outside that shape, structure stops helping and starts obstructing. Every structured surface needs a pressure valve — a free-text field, an escape to open conversation — and the structure itself should be the starting point the system fills or suggests, not a form the person completes from zero.
When to use
- The task has required parameters, and omitting one predictably degrades the output.
- The same request shape recurs across users or sessions and deserves a reusable template.
- Free-text requests routinely arrive incomplete, triggering avoidable clarification round-trips.
- Non-expert users need to produce expert-grade requests without learning prompt craft.
Design considerations
- 01Structure only the fields that materially change the output; leave the rest to prose.
- 02Always include a free-text field — structure scopes the request, it must not cap it.
- 03Pre-fill fields from available context (selection, page, history) so the form starts half-done.
- 04Keep templates one screen tall; a structured prompt that scrolls has become paperwork.
- 05Let power users bypass the form entirely with an equivalent free-text or command path.
- 06Show how the filled structure composes into the actual request, so the mapping stays legible.
- 07Validate at the field level before submission instead of failing the whole generation after.
Pitfalls
- ✕Form creep — a quick request becomes a twelve-field wizard, and the flexibility that made AI input appealing is gone.
- ✕Rigid taxonomies that cannot express the user’s actual intent, forcing lies in dropdowns to get past the gate.
- ✕Structure as decoration: fields that get concatenated into a prompt the model then ignores half of, with no feedback about which half.
- ✕No escape hatch — the structured path is the only path, and every out-of-shape request has to fight the form.
In the wild
- Notion AI
- Block-level AI actions present structured menus of transforms scoped to the current content, replacing prompt-writing with selection.
- Linear
- AI-assisted issue creation works within Linear’s existing structured fields — title, description, labels — so generated output lands in a schema, not a blob.
- GitHub Copilot
- Slash commands like /fix and /tests are micro-templates: a verb plus the current code context, a complete structured request in one token.
- Figma
- AI features such as visual search and prompt-to-design run through scoped, purpose-built entry points rather than one general chat box.
- ChatGPT
- Custom GPTs let builders bake instructions and required context into a template, so end users express only the variable part of the request.