Skip to main content
Beta

​
The 3 areas for prompts

Prompting in Corti happens on three surfaces in the generation block, and they form a hierarchy. Each one steers the model less forcefully than the one above it. Use all three, but put each instruction on the strongest surface that fits: placement matters more than volume.
  1. Section-level instructions — primary signal. The contentPrompt, writingStylePrompt, and miscPrompt fields. Your main lever: free-text direction over what a section contains, says and how it reads.
  2. Section-level outputSchema descriptions — secondary signal. Each node’s description is targeted, per-field guidance: narrower in scope than instructions and lower in the hierarchy, but precise where it applies.
  3. Section-level heading — ambient signal. A clear, self-explanatory heading (or a standard clinical-note section name) is always present and nudges the model even with no other instruction. Weakest, but free.
In addition to section-level prompts, the surface of instructions also exist at the Template-level where the prompt applies generally to all sections.
This cookbook captures the patterns Corti’s own Standards converge on, and is the prompting counterpart to Customize (Cookbook): Customize tells you where to put changes; this cookbook tells you what to write inside them. Schema mechanics live in Section Schemas — here we include them to be part of the prompts.

​
Section-level instructions

FieldUse forDon’t use for
contentPromptThe semantic content routed to this section — what substance belongs here vs. adjacent sections.How to render it; tone; structure; voice.
writingStylePromptThe general section’s style — voice, person, length, register, terminology.What content belongs; rendering rules.
outputSchema descriptionsThe node specific guidelines — Prompt instructions for each specific node in the outputSchemaRestating content scope or global style.
The litmus test for a prompt you’re about to write:
  • “What information goes in this section?” → contentPrompt (Include/Exclude).
  • “How should the prose read, in general in this section?” → the writingStylePrompt
  • “How should this specific node behave” → the relevant outputSchema description, at the node it applies to (top-level description for whole-output instructions; a field/item description for field-specific ones).

​
Recipes

1. contentPrompt — Include / Exclude

The “what to write about” prompt. Pair Include with Exclude when neighbouring sections have related scope.

2. writingStylePrompt — rules + examples

Tone, voice and style instructions.

3. miscPrompt — defensive guardrails

Fall-back lever for repeated misbehaviour contentPrompt and outputSchema can’t prevent — edge cases, layout, vocabulary, prohibitions.

4. outputSchema descriptions — per-node prompting

The second free-form prompting surface — scoped per node (field, item, top-level). Use for per-slot scope, formatting rules, and fallbacks.

5. Template instructions.prompt — domain priming

The single template-level prompt. Prime the model with specialty, setting, and shared rules that apply to every section.

Section Schemas

The outputSchema surface in detail — typed constraints + rendering primitives.

Corti Standards

Source material for these recipes.

Create a Section

Where both surfaces live — the instructions and outputSchema blocks.

Customize (Cookbook)

Structural customization — pair these prompts with the right persistence strategy.