| name | wolfram-prompt |
| description | Author a Wolfram Prompt Repository resource (a deployable LLM Persona, Function, or Modifier prompt) as a literate-markdown document and convert it to the official definition notebook with MarkdownToNotebook. Use this whenever the user wants to create, write, draft, or publish a Wolfram Prompt Repository resource, an LLMPrompt, an LLM persona / function / modifier - especially when they would rather write markdown than hand-edit notebook cells. Also use it when asked to add chat examples, options, or metadata to such a prompt. |
Authoring a Prompt Repository resource in markdown
MarkdownToNotebook fills the official Prompt Repository definition notebook (the
one CreateNotebook["PromptResource"] opens, with its docked Deploy/Submit toolbar)
from a literate-markdown document. A Prompt resource is one of three types -
Persona (a chat character), Function (an LLM-backed callable), or
Modifier (a post-processor that reshapes another LLM's output) - and each has
the same notebook layout, with type-specific slots becoming optional. The author
writes YAML frontmatter and ## section headings; the converter chooses every
cell style. Use the Prompt template.
Model new documents on the worked example -
https://github.com/WolframInstitute/MarkdownToNotebook/blob/main/examples/AdaLovelace.md
(a Persona) - and read
https://github.com/WolframInstitute/MarkdownToNotebook/blob/main/docs/resource-notebooks.md
(the "Prompt" section) for the slot-by-slot mapping.
Read first - the canonical guidelines:
Frontmatter
Fence a key: value YAML header with --- at the very top:
---
Template: Prompt
ResourceType: Prompt
PromptType: Persona # Persona | Function | Modifier
Name: CamelCaseName # noun for Persona, verb for Function, past-tense verb for Modifier
Description: One-line summary that starts with a verb and has no trailing punctuation
ContributedBy: Author Name
Keywords: [keyword one, keyword two]
Categories: [Fictional Characters] # type-specific list, see below
Topics: [History of Computing, Mathematics]
RelatedSymbols: [LLMPrompt, LLMSynthesize]
RelatedResources: [OtherPromptName]
Links: ["[label](https://example.com)"]
---
Name follows strict per-type grammar - the repository review enforces it:
- Persona - CamelCase noun phrase (
Yoda, Wolfie, MockInterviewer, AdaLovelace)
- Function - CamelCase verb phrase (
Summarize, CodeWriter, TitleSuggest)
- Modifier - CamelCase past-tense verb (
HaikuStyled, Translated, ELI5)
Categories is checkbox-style and the allowed values depend on PromptType:
- Persona categories: Advisor Bots, Character Types, Fictional Characters,
Purpose-Based, Roles, Writers, Writing Genres
- Function categories: AI Guidance, Chats, Content Derived from Text,
Education, Entertainment, For Fun, General Text Manipulation, Linguistics,
Prompt Engineering, Special-Purpose Text Manipulation, Text Analysis,
Text Generation, Wolfram Language
- Modifier categories: Computable Output, For Fun, Output Formatting,
Personalization, Text Styling
Sections (each ## Heading fills a slot)
Required:
## Prompt - the actual prompt body, plain prose. For Function prompts, write
named template slots as `{argName}` and they become TemplateSlot placeholders
that LLMResourceFunction substitutes at call time. This is the most important
part of the resource; the Prompt Repository style guidelines
go into detail on writing it well.## Chat Examples - one or more wl cells using the chat-style invocation
(ChatEvaluate[..., LLMPrompt["MyPromptName"] ...]) that exercise the
persona in a Chat Notebook context. Mark these cells #| eval: false
unless a deployed prompt + LLM connection is available at build time.## Basic Examples (and any of the usual Scope, Applications, Properties and Relations,
Possible Issues, Neat Examples subsections) - programmatic examples that show
the prompt working through LLMSynthesize / LLMResourceFunction rather
than a chat session. Same example-section conventions as a Function Resource;
for the per-cell authoring rules (one demonstration per cell,
:-terminated caption, --- between siblings) see
docs/examples.md.
Optional, Persona-only:
## Persona Icon - one wl cell that evaluates to an image; becomes the avatar
shown in Chat Notebooks.## Cell Processing Function - one wl cell with a pure function applied to
every user input cell before the model sees it.## Cell Post Evaluation Function - one wl cell with a pure function applied
to the model's output cell after it lands.
Optional, Function-only:
## Output Interpreter - one wl cell with the function applied to the model's
textual reply to coerce it into a computable result (e.g. ToExpression, an
Interpreter).
Optional, any type:
## Usage - usage statements, one paragraph per signature (see the Function
Resource skill for the `signature` + prose convention).## Details & Options - bulleted notes on how the prompt is configured.## LLM Tools - one wl cell whose value is a list of LLMTool instances
the prompt should have access to.## LLM Configuration - one wl cell whose value is the extra
LLMConfiguration options (model, temperature, ...).
Code-cell options
A fenced wl cell carries #| option lines at the top (the Quarto cell-option
convention), one key: value per line: eval: false (show code without running),
file: path (replace the body with a local file or URL), screenshot: true
(rasterize a produced Notebook), tear: h (torn-paper screenshot capped to h
points), flag: future|excised|.... Record an example's expected result in an
<!-- => ... --> comment after the cell. Inline math is $...$. To link a
documented symbol inline, wrap an inferred ref in <code>:
<code>[Range]()</code> - the empty parens make markdown viewers render it as a
clickable link, and the <code> wrapper applies code styling. The converter
routes the empty-URL link through linkInferred to a paclet: ref; the -out.md
twin further rewrites it to the public web URL.
Examples that load the deployed resource (LLMPrompt["MyPromptName"])
cannot evaluate before the resource exists, and LLMSynthesize / ChatEvaluate
need an active LLM connection at evaluation time. Mark these cells
#| eval: false so the build does not error; once the prompt is published
and the LLM is configured, the cells run as written. (Do not use the
LLMPrompt[ResourceObject[EvaluationNotebook[]]] form - EvaluationNotebook[]
is the user's notebook at call time, not the deployed resource, so the
lookup fails both headlessly and at runtime in a chat session.)
Build & deploy
wolframscript -f build.wls
The notebook the converter writes carries the same docked Deploy/Submit toolbar a
hand-authored Prompt notebook has; click Submit to the Prompt Repository in the
toolbar, or ResourceSubmit[NotebookOpen[\"MyPrompt.nb\"]] from a session, to
publish. For a private cloud copy use DeployResource[..., "CloudPublic"].
Check
Before submission, run the docked Check button (top of every resource
definition notebook) - it lints the document against the submission
guidelines and reports hints by level. Headless, the same lint runs through
DefinitionNotebookClientCheckDefinitionNotebook[nbo]` after stamping
CellIDs and saving (the headless build does not assign CellIDs, and the
scraper needs them to locate cells):
Needs["DefinitionNotebookClient`"]
UsingFrontEnd @ Block[{nbo = NotebookOpen[File["MyResource.nb"]]},
CurrentValue[nbo, CreateCellID] = True;
SelectionMove[nbo, All, Notebook];
FrontEndTokenExecute[nbo, "Save"];
Normal @ DefinitionNotebookClient`CheckDefinitionNotebook[nbo]
]
Each row is <|"Level" -> ..., "Tag" -> ..., "Parameters" -> ...|> with
Level one of Suggestion / Warning / Error. Common tags to address
before submission: DescriptionTooLong (shorten to under 128 chars),
ExampleTextLastCharacter (end an example caption with :),
FoundUnformattedCode (wrap a stray WL symbol in `backticks` or in
an inferred link with empty parens like [Range]() (substitute the actual symbol name for Range), ThreeDotEllipsis (use … not ...),
NotASystemSymbol (link foreign function-repo names instead of formatting
them as system symbols), LargeCellBounds/CellHeight (rasterized output too
big - crop it with #| tear: h or shrink the source). The repo's
check.wls runs the same lint on every built .nb and prints a per-file
summary.
