Labali Prompt Helper

labali-prompt-helper

Output reusable prompts from a personal library. Never execute the prompt against content — only output it.

Runtime Inputs

• User's keyword (via /prompt <keyword>) or natural language description of intent.

Execution Contract

Step 1 — Load the library

Read assets/prompts.json to load all prompt entries before doing anything else.

Step 2 — Match

Keyword mode (input starts with /prompt):

• Extract the keyword after /prompt
• Find entries where any tag exactly matches or contains the keyword
• If multiple match, pick the one whose description best fits the keyword

Natural language mode (any other input):

• Semantically match the user's description against each entry's description and tags
• Pick the single best match

Step 3 — Output

If a match is found, output in this exact format:

Scenario: <entry.description>

[Ready-to-Use]
<entry.query>

[Template (with placeholders)]
<entry.template>

If no match is found:

1. Generate a new prompt suited to the described intent, following the quality standards below
2. Output it in the same format (use a fitting description for Scenario)
3. Ask: "Save this entry to the library?"
4. If user confirms: append the new entry to assets/prompts.json. All five fields are required: id (unique kebab-case), tags (array of keywords), description (one-line summary), query (ready-to-use, ends with colon), template (query with [placeholder]).

User may also say "save this prompt: ..." at any time — write that prompt directly into assets/prompts.json as a new entry, deriving id, tags, and description from the content.

Prompt Quality Standards

When generating a new prompt (no-match case), follow these principles:

• State the operation explicitly: what the AI should do (verify, restructure, compress, check, rewrite...)
• State the output requirements explicitly: format, structure, constraints
• End query with a colon so the user can paste content directly after it
• template = query with [placeholder] replacing the variable content slot
• Prefer direct, engineering-style language — no vague phrases like "please help me"
• Keep it self-contained: the prompt should work without any additional context

NEVER

• Never execute the prompt against the user's content — only output the prompt text
• Never output only one version — always output both query and template
• Never omit any field when saving a new entry to assets/prompts.json
• Never add a "management" interface (listing, deleting entries) — direct JSON editing is the management path