Prompt library curator

Prompt Library Curator

This skill organises a personal collection of saved prompts into a structured, navigable markdown catalog. It assigns categories, tags, summaries, and complexity ratings to each prompt, detects duplicates, and outputs a clean prompt-library.md file ready for reuse and sharing.

Input:

• Raw prompts as pasted text, an inline list, or a .md/.txt file

Output:

• prompt-library.md — structured catalog with index table, categorised prompts, and flagged duplicates

---

Language Detection

Detect the user's language from their message:

• If Russian (or contains Cyrillic): respond in Russian, use Russian category names in output, and generate the entire catalog in Russian
• If English (or other Latin-script language): respond in English, use English category names, and generate the entire catalog in English
• If ambiguous: respond in the language of the trigger phrase used and generate catalog in that language

---

Instructions

Step 1: Accept and Validate Input

1. Determine input source:

   • Pasted text block in chat
   • Uploaded or referenced .md / .txt file
   • Inline numbered or bulleted list

2. Validate input is non-empty

   • If empty: stop and respond — "No prompts found. Paste your prompts directly or attach a .md/.txt file."

3. Check for unsupported file formats (.docx, .xlsx, .pdf, etc.)

   • If detected: stop and respond — "Unsupported format. Please paste your prompts as plain text or upload a .md or .txt file."

4. Check if input looks like a conversation log rather than a prompt collection

   • Heuristic: high ratio of questions and conversational turns with no clear prompt-style structure
   • If detected: respond — "This looks like a chat log, not a prompt collection. Paste the prompt texts you want to organise, or specify which lines are prompts."

---

Step 2: Split into Individual Prompts

1. Attempt to split the raw input into individual prompts using these separator heuristics (in priority order):

   • Horizontal rules (---)
   • Numbered list items (1., 2., etc.)
   • Blank-line-separated paragraphs
   • Double newlines

2. If no separators are detected and input is a wall of text:

   • Split conservatively at double newlines
   • Add a note at the top of the output: "⚠ Separation uncertain — review split points"

3. Count the total number of prompts identified; store as N.

4. If N = 1:

   • Proceed with a single-entry catalog
   • Add note at the bottom: "Add more prompts to build a full library"
   • Skip the index table (single entry doesn't warrant it)

---

Step 3: Analyse Each Prompt

For each prompt, determine:

1. Category — assign one from this set (or create a new label if none fits):

   • Writing, Research, Coding, Analysis, Meeting, Planning, Personal, Learning, Marketing, Other

2. Tags — assign 2–4 short keyword tags (e.g., email, summary, b2b, template)

3. One-line summary — ≤10 words describing what the prompt does (e.g., "Summarise a long PDF report")

4. Complexity — one of: Simple / Moderate / Advanced

   • Simple: short, single-step prompt
   • Moderate: multi-step or context-dependent prompt
   • Advanced: complex chain-of-thought, multi-role, or system-prompt level

5. Template flag — if prompt contains placeholder variables ({{topic}}, [PERSON], <context>, etc.):

   • Preserve variables verbatim in the prompt text
   • Add template to the prompt's tag list

---

Step 4: Detect Duplicates

1. Compare prompts for near-identical intent or wording (semantic similarity heuristic):

   • Exact duplicates: identical text after whitespace normalisation
   • Near-duplicates: same core instruction with minor wording differences

2. For each duplicate pair:

   • Keep both in their respective category sections
   • List them in the Duplicates section with a note
   • Do not delete either prompt automatically

---

Step 5: Build Index Table

1. Sort all prompts by: Category (A→Z) → Complexity (Simple → Advanced)

2. Create an index table with columns: #, Category, Summary, Tags, Complexity

3. If N > 20: add a Quick Find section above the index with anchor links to each category heading

---

Step 6: Assemble Output Catalog

Build prompt-library.md using the Output Format structure:

1. Header block (title, date, stats: total prompts, categories, duplicates flagged)

2. Quick Find section (if N > 20)

3. Index table

4. Horizontal rule separator

5. One ## Category section per category, sorted A→Z

   • Under each: named prompt heading, frontmatter line (tags + complexity), original prompt text in a code block

6. Duplicates section (always present; states "No duplicates detected." if none found)

7. If user has file system access (Cowork mode): save as prompt-library.md in workspace folder and confirm path

8. If no file system access: display full output inline in chat

---

Step 7: Handle Partially Organised Input

If the user's input already contains their own headings, categories, or tags:

• Preserve existing structure and user-defined category names
• Enhance only missing metadata (add tags/complexity where absent)
• Do not overwrite or rename user-defined categories

---

Output Format

prompt-library.md structured as follows:

# Prompt Library
**Last updated:** YYYY-MM-DD
**Total prompts:** N  |  **Categories:** X  |  **Duplicates flagged:** Y

---

## Quick Find
[Writing](#writing) · [Research](#research) · [Coding](#coding) · ...
*(Present only if N > 20)*

---

## Index

| # | Category | Summary | Tags | Complexity |
|---|----------|---------|------|------------|
| 1 | Writing | Draft a cold email | email, outreach, b2b | Simple |
| 2 | Research | Summarise a PDF report | pdf, summary, analysis | Moderate |

---

## Writing

### Draft a cold email
`tags: email, outreach, b2b | complexity: Simple`

Write a cold email to [PROSPECT] introducing [PRODUCT]. Focus on the pain point of [PROBLEM] and propose a 15-min call.

---

## Research

...

---

## Duplicates
> Prompts flagged as near-duplicates. Review and keep one.

- "Summarise this document" ≈ "Give me a summary of this file" (similarity: high)

*(If none: "No duplicates detected.")*

Field rules:

• Each prompt's original text is preserved verbatim inside a code block — no edits
• Tags and complexity are on a single inline frontmatter line between the heading and the code block
• Template variables ({{}}, [], <>) are kept as-is and the prompt is tagged template
• The Duplicates section is always present in the output

---

Negative Cases

• Empty input: Respond — "No prompts found. Paste your prompts directly or attach a .md/.txt file." Do not generate output.
• Unsupported file format (.docx, .xlsx, .pdf): Respond — "Unsupported format. Please paste your prompts as plain text or upload a .md or .txt file." Do not generate output.
• Input is a conversation log: Respond — "This looks like a chat log, not a prompt collection. Paste the prompt texts you want to organise, or specify which lines are prompts."