Your AI Knowledge Base — Hands On.

Session 3 was all doing, not just talking. We opened laptops, set up GitHub repositories, generated tokens, connected Claude to our repos, and watched AI read our project context for the first time. We also heard from Shweta and Samarth about their IDH dashboard — a real project built with AI.

Every new Claude conversation starts from scratch — like a brilliant stranger walking into your office. They know nothing about your project, your preferences, or your history. You have to explain everything from the beginning.

But here's the trap: the longer a conversation runs, the higher the token cost, the more hallucinations creep in, and the worse the output gets. It's like giving someone work at 9 AM vs 6 PM — same task, completely different quality.

The solution: close sessions, start fresh, carry forward only what matters via CLAUDE.md.

During the session, we walked through a real project repository (mGrant V3) to show how knowledge files are structured. Here's the architecture that makes AI remember everything.

This is the template we used in the session. Copy it, fill it in, commit it to your repo. Your AI starts learning about you the moment you do.

This is the security standard shared during the session. It goes in every repository — personal and organisation. Copy the full text below.

A skill is a cheat sheet you give your AI. Write it once, use it forever. It's an SOP in markdown format.

One of the key discoveries in this session: Claude UI is great for reading, but limited for writing.

Claude UI can access your GitHub repo via token and understand your context — it reads your CLAUDE.md, your skills, your project history. But it struggles to push changes back: creating PRs, committing files, and updating repos consistently.

Claude Code (CLI) can both read AND write. It's the full-power version — it creates branches, commits code, raises PRs, runs tests, and connects to external tools. That's what we'll set up in Session 4.

For now: use Claude UI for conversations with context. CLI comes next.

The dashboard started on Lovable (a free AI coding tool) because it was free on March 8th. From Lovable, it moved to GitHub, then to Claude for refinements. The stack: Lovable for UI scaffolding, GitHub for version control, Claude for logic and debugging.

After deployment, 30+ of the 69 indicators showed no data. The team had given exact form numbers, schemas, and logic — but the language was "too human" for AI to map correctly.

The fix took 1.5 days — creating a revised logic sheet with structured categories, form references, and descriptive examples. After that, every indicator worked. The lesson: invest in documentation, not more prompts.

A forbidden keyword in the backend was silently blocking database queries. After 6 attempts, Sonnet couldn't find it. Switching to Opus with extended thinking finally cracked it.