Experimenting with AI prompts (Prompt Guru)

I have tinkered with a browser-based chat user interface called Guru. It is easy, runs exclusively in the browser with puter.js and is supposed to be a clean playground to play around with input requests. I wanted something simple where I could: - Try different fast styles. -Watch the AI ​​stream answers in real time. - or export discussions for a later review. What is different? The special sauce is the fast guru kernel sitting under the bonnet. Any prompt that you enter is executed by a complex optimization formula called Marm (meta-algorithmic model) before it is sent to the model. Marm is basically a structured process to make the input requests better: - compression → inflate equipment and tighten the language. - Frame → Surfaces hide intention and sharpen the questions. - Improvement → adds useful structure such as roles, formats or restrictions. - Evaluate → executing checks for clarity, accuracy and analogy. Then it goes on: - Validation grille → "Teen -Test" (can a beginner be released in a line?), "Expert test" (exactly enough for one Pro?) And "analogy test" (it is mapped on something familiar?). - Voltage tests → Put on commands under edge conditions (brevity, contradictory roles, security checks). -Scoring & repetition → If the entry request does not continue, become automatic and run through it again or mark the error. - Lesson mode → Explains changes to you with a compact EC → A ++ (explain, compare, apply) so that you learn from optimization. So every conversation is not only a answer-it is also a mini-lesson in the fast design. You can try it out here: Repo: Some functions in: Mobile-friendly layout with a single Hamburger menu. Support for several models (yes, including GPT-5). /Download sessions and export transcripts in JSON or Markdown. Settings modal for model / temperature / max token, whereby the values ​​are stored locally. Authelled from PUTER.com (or simply use a temporary account if you want to test quickly). I built it for myself as a decent space for learning and testing, but others that they could experiment with prompt engineering could also be useful. Feedback is more than welcome!