Skip to content
Last updated


Summary from Dmytro (using free ChatGPT mac application):

  • Changing promts with the magic words does affect the result. But I can't say they are always correct. E.G. first question about rbac:
Search PhraseResult Quality
no magic wordsjust not usable general info
with the magic wordslooks like result is better, but actual configuration suggestions are not correct
search with AI on our sitemost correct answer
just mentioned Redoclynot usable result
point to site + additional contextlooks correct
  • When some product name mensioned, like API monitoring using Redocly results seem to be closer to the actual source, but not examples. Even after using magic words the final examples are not correct, e.g. redocly.yaml suggestion for Respect Monitoring:

    name: API Monitoring
    
    on:
      schedule:
        - cron: '0 6 * * *'  # Runs daily at 6:00 UTC
      push:
        paths:
          - 'openapi.yaml'
    
    jobs:
      lint-api:
        runs-on: ubuntu-latest
        steps:
          - uses: actions/checkout@v3
          - name: Set up Node.js
            uses: actions/setup-node@v3
            with:
              node-version: '20'
          - run: npm install
          - run: redocly lint openapi.yaml
  • In general, I am not convinced of the positive effects of llms.txt. Results seem pretty similar whether you mention the product name or provide a URL to the docs.

What made you smile?

  • It was fun to play with different requests. But After first two queries it looks like a repetitive actions with similar outputs.

What did you find confusing?

  • When I use magic words there is some illusion (llm displays Redocly logo when generating answer) that the answer is expected to be correct, but that is not always true.
  • LLM.txt docs look raw and don't have confirmation that other big players are actually going to support it.

Summary from Greg

LLM Testing Query Matrix

Test CaseBase QueryMagic Words QueryDirect Docs QueryNotes
The GatekeeperHow do I protect this debrief.md page so that it requires someone with the "pearl" role to access it?Use https://redocly.com/llms.txt and any subsequent links from that index. How do I protect this debrief.md page so that it requires someone with the "pearl" role to access it?Using https://redocly.com/docs - How do I protect this debrief.md page so that it requires someone with the "pearl" role to access it?Test RBAC functionality
The StorytellerHow do I create a code walkthrough?Use https://redocly.com/llms.txt and any subsequent links from that index. How do I create a code walkthrough?Using https://redocly.com/docs - How do I create a code walkthrough?Test documentation features
The IllusionistHow do I only show the mock server in the try it console (not a real server)?Use https://redocly.com/llms.txt and any subsequent links from that index. How do I only show the mock server in the try it console (not a real server)?Using https://redocly.com/docs - How do I only show the mock server in the try it console (not a real server)?Test API console config
The WatchmanHow do I run API monitoring using Redocly on a regular schedule?Use https://redocly.com/llms.txt and any subsequent links from that index. How do I run API monitoring using Redocly on a regular schedule?Using https://redocly.com/docs - How do I run API monitoring using Redocly on a regular schedule?Test monitoring setup
The ArtistHow do I make headings purple colored in Redocly?Use https://redocly.com/llms.txt and any subsequent links from that index. How do I make headings purple colored in Redocly?Using https://redocly.com/docs - How do I make headings purple colored in Redocly?Test styling customization

Results Tracking

Rating Scale:

  • 1 = Completely wrong/unusable
  • 2 = Partially correct but missing key details
  • 3 = Generally correct but needs refinement
  • 4 = Very good with minor issues
  • 5 = Perfect/highly accurate

Grok-3 Beta Results

Test CaseBase Result QualityMagic Words Result QualityDirect Docs Result QualityCode Examples AccuracyBest Approach
The Gatekeeper134IncorrectDirect
The Storyteller131IncorrectMagic
The Illusionist223IncorrectDirect
The Watchman121IncorrectMagic
The Artist324IncorrectDirect

links:

Base Queries.md

Magic Words Queries.md

Direct Docs Queries.md

Claude 4 Sonnet Results

Test CaseBase Result QualityMagic Words Result QualityDirect Docs Result QualityCode Examples AccuracyBest Approach
The Gatekeeper424CorrectBase
The Storyteller522CorrectBase
The Illusionist422CorrectBase
The Watchman224IncorrectDirect
The Artist544CorrectBase

links:

Base Queries.md

Magic Words Queries.md

Direct Docs Queries.md

DeepSeek-V3 Results

Test CaseBase Result QualityMagic Words Result QualityDirect Docs Result QualityCode Examples AccuracyBest Approach
The Gatekeeper122Incorrect-
The Storyteller111Incorrect-
The Illusionist111Incorrect-
The Watchman232CorrectMagic
The Artist123CorrectBase

links:

Base Queries.md

Magic Words Queries.md

Direct Docs Queries.md

Gemini 2.5 PRO Results

Test CaseBase Result QualityMagic Words Result QualityDirect Docs Result QualityCode Examples AccuracyBest Approach
The Gatekeeper111Incorrect-
The Storyteller111Incorrect-
The Illusionist311Incorrect-
The Watchman113IncorrectDirect
The Artist353CorrectMagic

links:

Base Queries.md

Magic Words Queries 2.md

Direct Docs Queries 2 2.md

Mistral Nemo Results

Test CaseBase Result QualityMagic Words Result QualityDirect Docs Result QualityCode Examples AccuracyBest Approach
The Gatekeeper111Incorrect-
The Storyteller111Incorrect-
The Illusionist112IncorrectDirect
The Watchman111Incorrect-
The Artist222Incorrect-

links: Base Queries.md

Magic Words Queries 2 2.md

Direct Docs Queries 2 2.md

Sonar Pro Results

Test CaseBase Result QualityMagic Words Result QualityDirect Docs Result QualityCode Examples AccuracyBest Approach
The Gatekeeper4x1Incorrect-
The Storyteller1x2IncorrectDirect
The Illusionist2x3-Direct
The Watchman3x1-Base
The Artist3x3Incorrect-

Sonar answer that he don't have access to llms.txt

links:

Direct Docs Queries 3 2.md

Base Queries.md

Magic Words Queries 2.md

issues:

  1. There is an issue with the editor: Dragging and dropping multiple files randomly renames them.

  2. I think the llm.tx reduces the amount of hallucination, but each LLM seems to handle it differently. Some times this approach also reduce size of text waht LLM outputs so It not really ansewer on question.

  3. Sonar pro don't see llm.txt fille at frst run i check it again later and it works this could be issue with our App (downtime)