Improving default tagging AI with custom rules

[johnfromberkeley]

I wish we could edit the default prompt for AI tagging. I've created a lot of custom rules to try to improve the prompt.

One of my biggest issues is that I'm getting constant duplicates. For example, San Francisco and San Francisco. Here is my current stack. Each of these is a custom prompt under five hundred characters.

It looks like I'd have to fork the entire project to create my own version with my own prompt for tagging. It would be so much better if we could edit the default prompt anyway. I'm especially curious if anyone has figured out the duplication issue I'm facing.

Section 1: Role and Philosopy

Include BOTH descriptive tags AND base concept tags. Descriptive tags capture specific themes, topics, & ideas. Base concept tags classify the page into fundamental categories. Both are necessary.  ALWAYS tag the SPECIFIC SUBJECT of the page. If about John Steinbeck, tag "john-steinbeck" in addition to "authors". If chocolate cake, tag "chocolate-cake" in addition to "recipes". If Berkeley, tag "berkeley" & "california". Don’t stop at the base category. Capture the specific entity or topic.

Section 2: Plural vs. Singular

When a tag represents a BUCKET OF THINGS (a collection you navigate into), use the plural form: videos, tools, apps, lists, recipes, books, resources, authors.

When a tag represents a CONCEPT or ABSTRACT IDEA (a discipline or topic of study), use the singular form: video (formats), programming (the discipline), music (theory).

Proper nouns, adjectives, and verbs used as tags follow their own natural forms without forced pluralization.

Section 3.1: Special Vocabulary A

Special Vocabulary A

Apply these specific tags when criteria match in addition to other relevant tags. Always supplemental, not replacements.

- discovery: Pages for exploration and finding new interests—not immediately actionable, but valuable for discovery.
- lists: Pages containing a list of items (products, tools, resources, people, recommendations).
- recommendations: Pages suggesting or recommending specific items, services, or actions.

Section 3.2: Special Vocabulary B

Special Vocabulary B

Apply these specific tags when criteria match inn addition to other relevant tags. Always supplemental, not replacements.

- reference: Pages useful as lookup tools, documentation, specifications, or utilities.
- resources: Pages compiling useful materials, links, or assets for ongoing use.
- guides: Pages providing step-by-step instruction or educational content.

Section 3.3: Special Vocabulary C

Special Vocabulary C

Apply these specific tags when criteria match inn addition to other relevant tags. Always supplemental, not replacements.

- tips: Pages offering practical advice, tricks, or shortcuts.
- unread: Mark pages as "unread" if title or content suggests you have not yet processed this page.

Section 4: Descriptive Tags

Generate tags capturing specific themes, topics, entities, and subjects. Use lowercase with hyphens: machine-learning, web-development, john-steinbeck, chocolate-cake.

Tag both broad categories AND narrow specifics. Include the specific subject even when a base concept tag already applies.

Avoid redundancy—do not include synonyms or near-duplicates as primary tags.

Section 5: Rules

SPECIAL RULES

- INCLUDE AS MANY RELEVANT TAGS AS THE PAGE JUSTIFIES. There is NO tag limit. A well-tagged page may have twenty, thirty, or more tags. Skipping tags to stay under an imaginary ceiling is WRONG.
- Do NOT skip base concept tags. These are essential.
- Do NOT skip specific subject tags. Capturing "recipes" without "chocolate-cake" is incomplete.
- Tags must be all lowercase. NO EXCEPTIONS.

Section 5.1 - Tag Normalization A

Tag Normalization A

Prevent duplicate and variant tags. Before suggesting a tag:

1. Format check: Reject any tag not in all-lowercase with hyphens (e.g., "San Francisco", "BayArea", "bayarea" → "san-francisco", "bay-area").

2. ResolutionWhen a variant exists, suggest the correctly-formatted lowercase version. If "san-francisco" exists, reject "San Francisco" and "sf". 

Do not suggest tags that duplicate existing content through formatting variants.

Section 5.2 - Tag Normalization B

Tag Normalization B

Variant detection: Check for existing tags that represent the same entity using different forms:
   - Full name vs. abbreviation ("san-francisco" vs. "sf")
   - With/without spaces ("chocolate-cake" vs. "chocolatecake")
   - With/without hyphens ("machine-learning" vs. "machinelearning")
   - Capitalization differences ("AI" vs. "ai")
   - Misspellings or typos

Section 5.3 - Tag Normalization C

Tag Normalization C

Duplicate enforcement: Scan your output before returning. If "guides", "Guides", and "GUIDES" would all appear, output only "guides". If "note-taking" and "notetaking" both exist in the vocabulary, pick the canonical form "note-taking" and use it.

The final array should contain zero duplicates and zero variant formats of the same concept.

Section 6.1: Examples A

Examples A

Web page: Wikipedia article about Toni Morrison, discussing her novels, literary awards, and influence on American literature.

Tags returned: "authors", "writers", "american-literature", "novelists", "biography", "toni-morrison", "literary-awards", "books"

Web page: YouTube video titled "10 Best Open Source Photo Editors in 2024"

Tags returned: "videos", "open-source", "photo-editing", "lists", "software", "photography", "recommendations", "tools", "photo-editors"

Section 6.2: Examples B

Examples B

Web page: Recipe for chocolate layer cake with espresso buttercream

Tags returned: "recipes", "baking", "cakes", "chocolate-cake", "desserts", "espresso", "buttercream", "cooking"