Most developers know AI tools exist. Fewer know how to make them actually work in a real codebase, on a real deadline, without burning through tokens and producing code that doesn’t run. That’s exactly what this session was about.

Khachik Gasparyan — full-stack PHP developer from Primis with two years of daily Claude Code practice — walked us through his actual workflow. Not theory. Not slides with bullet points.

What we covered

The session opened with a quick landscape of the tools that matter right now — Claude Desktop, Claude Code, Cursor, GitHub Copilot — and what each one is actually good for. Khachik was direct: not all tools are equal, and choosing the right one for the right job saves more time than any prompt trick.

From there, we moved into prompting. The most common mistake engineers make isn’t using the wrong tool — it’s giving it no context. “Fix this bug” tells Claude nothing. A prompt with the file path, the expected behavior, the framework version, and the constraints? That gets results.

The live demo

The core of the session was a live build: adding a referral and bonus system to an existing PHP/Laravel repository — from prompt to working task plan — using Claude Code in the terminal. Khachik showed how he sets up CLAUDE.md (his AI’s persistent memory), how he uses TaskMaster to break large features into manageable sessions, and how handoff files let you switch contexts without losing progress.

One key insight from the demo: Claude gets dumber after 200k tokens. Not metaphorically — literally. The more context is loaded, the more reasoning degrades. The fix is simple: use /compact early, and treat each session like a focused sprint, not an open-ended conversation.

Agents — your permanent team

The second half of the session introduced AI agents: specialized Claude configurations that behave like team members. Khachik’s setup includes a backend agent, a frontend agent, a security auditor, and an architect. They live in his CLAUDE.md folder. Claude picks the right one automatically based on what you’re working on.

The architect agent is worth highlighting: before touching a single line of code, it reviews the whole repo, maps out the change plan, identifies which files will be affected, and asks clarifying questions. It’s planning mode taken seriously.

Security — the part most people skip

Khachik spent time on something most AI tutorials ignore: what happens to your code and tokens when you use these tools. The short version — turn off training data sharing in Claude’s settings, never paste API tokens or passwords into a prompt, and use .cloudignore to protect sensitive files. He’s seen tokens get leaked through innocent-looking prompts. It happens.

Key takeaways

• Don’t do everything in one prompt — divide tasks into sessions and use handoff MD files between them
• CLAUDE.md is your AI’s memory — document your stack, standards, and preferences once
• 200k tokens is the sweet spot — use /compact to reset context without losing progress
• Never paste tokens or passwords into AI — use .cloudignore and turn off training data sharing
• Always review what AI changed — git diff, manual check, reviewer agents. AI can hallucinate functions that don’t exist
• Agents are your team — set them up once, and they work for you on every session automatically