Skip to main content

Custom Prompts

💡TL;DR

Custom prompts replace a supported task's built-in prompt when both the global and task switches are enabled and the custom text is nonempty. Preserve the task's output contract yourself. There is no automatic template system that restores every omitted format instruction.

Overview​

First test the default prompt on a representative note. Use focused-learning or synonym-suppression settings if those address the problem. Write a custom prompt when you need a specific extraction policy, output structure or translation style, then compare its output before using it in a batch.

How It Works​

Prompt Architecture​

Notemd selects the built-in or enabled custom task prompt, adds applicable focused-learning and language instructions, and substitutes the placeholders supplied by that task. Note content may also be sent as a separate message. Each task supplies different inputs; a placeholder usable in research is not a general variable available everywhere.

Replacing the extraction prompt can remove the CONCEPT: contract. Replacing the research prompt can omit the retrieved context. A well-formed request therefore does not prove the resulting output can be parsed or is supported by evidence.

Prompt Variables​

Use the exact uppercase names and single braces shown below:

TaskSupplied placeholders
Title generation{TITLE}, {RESEARCH_CONTEXT_SECTION}
Research summary{TOPIC}, {LANGUAGE}, {SEARCH_RESULTS_CONTEXT}
Translation{LANGUAGE}, {TEXT}
Original-text extraction{REFERENCE_CONTENT}, {USER_INPUT}

{LANGUAGE} is resolved through the task's language policy. Add-links and standalone concept extraction send their text as task content; do not invent a content placeholder for them. Unknown names are not magically filled. These are task prompt substitutions, not concept-note file templates.

Focused Learning Domain​

Enable focused learning and enter a domain such as electrical engineering. The prompt receives a Relevant Fields instruction. This guides the model's attention; it is not a hard filter or a proof that unrelated concepts will be excluded.

Configuration​

TaskEnable switchPrompt field
Add linksuseCustomPromptForAddLinkscustomPromptAddLinks
Title generationuseCustomPromptForGenerateTitlecustomPromptGenerateTitle
Research summaryuseCustomPromptForResearchSummarizecustomPromptResearchSummarize
Mermaid summaryuseCustomPromptForSummarizeToMermaidcustomPromptSummarizeToMermaid
Concept extractionuseCustomPromptForExtractConceptscustomPromptExtractConcepts
TranslationuseCustomPromptForTranslatetranslatePrompt
Original-text extractionuseCustomPromptForExtractOriginalTextcustomPromptExtractOriginalText

All of these require enableGlobalCustomPrompts. An empty custom prompt falls back to the built-in prompt. Merged original-text extraction has its own built-in format and does not share the ordinary custom extraction override. Diagram capability generation has its own typed output requirements; the Mermaid-summary override is not a universal diagram prompt.

Example​

For concept extraction from circuit notes, enable the global and extraction switches and use a focused instruction that retains the parser contract:

Extract the central electrical engineering concepts from the supplied note.
Prefer specific component names and physical mechanisms over generic nouns.
Return one concept per line in this exact format:
CONCEPT: Term Name
Do not add explanations or a surrounding code block.

Run it on a copied note. Inspect the returned entries, filenames and missing concepts, then compare with the default prompt. Preserve your original prompt before editing so you can restore it.

Tips​

  • Test one task at a time; change scope before increasing batch size.
  • Preserve source evidence in research prompts and exact excerpts in extraction prompts.
  • Keep executable identifiers and required output markers stable when localizing prompt prose.
  • Recheck behavior after changing model or provider. Prompt instructions are not validation guarantees.

Next Steps​