PromptTune

Prompt advice is scattered across guides and papers, and iteration usually happens in a scratch file with nothing retained. The design answer was to make the improvement itself retrieval-grounded - the optimiser cites a real corpus rather than the model's own habits - and to parse the output deterministically from tag-delimited fields, so the prompt, its rationale and its checklist are separable data rather than prose to re-read.

1. 01Implemented a RAG layer ingesting Markdown guides, arXiv papers and PDFs into a 384-dimension Pinecone serverless index, with recursive 1,000-character chunking, 200-character overlap and batched upserts using all-MiniLM-L6-v2 embeddings.
2. 02Designed a structured meta-prompting engine applying a five-step Task–Context–References–Evaluate–Iterate framework, with tag-delimited LLM output parsed deterministically into optimised prompt, rationale and checklist fields.
3. 03Built a persona system with eight seeded role templates and user-owned custom personas, composing persona instructions into system context for both optimisation and chat, with per-user ownership scoping and an active-persona default.
4. 04Implemented HS256 JWT authentication with PBKDF2-SHA256 password hashing over a six-table SQLAlchemy/PostgreSQL schema covering users, profiles, personas, prompts, chat sessions and analytics, with full CRUD REST APIs for a searchable prompt library.
5. 05Added a windowed in-memory chat session store retaining the last 20 messages per session, enabling an optimise → test → refine loop against the generated system prompt.

• RAG pipeline design
• Vector search and embedding models
• Meta-prompt engineering
• Structured LLM output parsing
• Relational schema design
• JWT authentication and password hashing
• Document ingestion and chunking strategy
• REST API development
• Persona-based prompt testing
• Frontend state management with React Context