What's in this build
A lean, self-hosted prompt engineering assistant. Single FastAPI service no database, no heavy infrastructure. Every feature listed below is live and callable from this UI or via the API.
Your raw prompt hits POST /intent. The LLM reads it and returns a structured signal: domain, goal, complexity, confidence score, and a list of recommended enhancement techniques. This governs everything downstream.
The active memory layer reads your stored context from GET /context/:userId. Prior session facts, preferences, and domain knowledge are injected into the enhancement prompt so the output is personalised to you.
POST /enhance receives your prompt, the intent signal, and your memory context. The system prompt in core/prompts/enhance_system.md instructs the LLM to apply 14 of 16 named techniques (chain-of-thought, persona injection, structured output, etc.) and return a fully annotated JSON result.
The raw LLM JSON is validated by core/output_validator.py against the shared contract in core/contracts.py. Each sentence in the enhanced prompt is tagged with its technique, rendered as colour-coded segments in the UI.
POST /refine takes the previous enhanced prompt and a follow-up instruction. The LLM re-applies the same contract and returns a diff-able updated version iterate until it's right.
After a successful enhance or refine, the active memory layer extracts durable facts from the interaction and upserts them into storage/data/ via core/active_memory.py. The Brain view shows the resulting knowledge graph.
Phase B live
Expected changes now live: PostgreSQL-backed memory, semantic retrieval during Enhance, scheduler-ready background jobs, JSON context migration, and end-to-end Docker deployment on Lightsail.
PostgreSQL storage
Docker Compose now runs a private Postgres container beside the app. /ready verifies the DB backend, and legacy JSON contexts were copied into user_contexts.
Semantic retrieval
Enhance ranks previous context against the current prompt and injects the most relevant 3-5 memories instead of blindly using the newest entries.
Scheduler foundation
A disabled-by-default scheduler can synthesize memory, generate proactive suggestions, and prune stale context when recurring jobs are enabled.
Live verification
Production was rebuilt on Lightsail. Postgres is healthy, public /ready reports PostgreSQL, context write/read persisted, and public Enhance SSE returned a done event.
Enhance
Technique-annotated prompt enhancement with intent-aware technique selection and segment-level explanations.
Refine
Iterative prompt refinement. Feed a follow-up instruction and get an updated enhanced prompt that preserves prior improvements.
Memory / Brain
Persistent user context stored as JSON. Visualised as a live knowledge graph. Supports incognito mode to skip read/write entirely.
Profile
Auto-ingest unstructured text or JSON from other AI tools to populate your persona and preferences into the knowledge graph.
CoThinker (voice)
Conversational voice loop. Speak your prompt idea; CoThinker transcribes via Groq Whisper, asks clarifying questions, and finalises an enhanced prompt when 8+ of 10 fields are confirmed.
Agentic Work
Streaming chat endpoint with a configurable system prompt. Tokens stream over SSE so responses feel instant. Designed for web search and browsing agentic workflows.
Neuro Mode
Scores your prompt across cognitive clarity, specificity, reasoning depth, and output alignment. Returns a rubric-based JSON report with an overall neuro score.
Connectors
Natural-language queries over your Google Workspace Gmail, Google Calendar, and Google Drive via the Groq Responses API.
MCP Server
Exposes Enhance and Refine as Model Context Protocol tools. Connect Claude Desktop, Cursor, or any MCP client directly to this instance.
Diagnostics
One-click health checks for every subsystem: intent, enhance, memory read/write, incognito, refine, and storage. Live status badges.