Alessio Franceschelli

You drop an AI agent into your real codebase and it produces garbage. Wrong style, subtle bugs, side effects everywhere. The problem isn't the model. Your codebase isn't ready for AI. Think of it like handing off a biathlon relay with no map of the course 🥇🧵👇

Full post with a concrete example of the ask tool in action: alessio.franceschelli.me/posts/ai/stop-…

Copilot can be far more cost-effective than it looks. Waste premium turns on avoidable back-and-forth? Expensive. Stretch them with the ask tool? Very different story.

The trick: tell the agent to use the ask tool, show the plan, ask if you're happy, and keep iterating until you say yes. One premium request → a proper working session with clarifying questions, plan refinement, and explicit sign-off.

The GitHub Copilot trick nobody talks about: use the ask tool to stop wasting premium requests. 🧵👇 I keep seeing engineers burn 100s of requests an hour and wonder why the tool feels expensive. The real problem? Using premium turns like free chat messages.

No more running `wsl --shutdown` manually. Check out the full step-by-step guide on how to set this up cleanly: alessio.franceschelli.me/posts/windows/…

The fix: Use a systemd *user* service to bridge `socat` and `npiperelay`. It lives only in your user session. When you close the terminal, the service stops, and WSL properly powers down. Best of both worlds.

Want to access your Windows ssh-agent keys in WSL but hate how it breaks automatic shutdown? I finally got tired of WSL draining my battery in the background and found a fix. 🧵👇

Quick checklist and concrete before/after examples on the blog. Stop the vibes, start the briefing. alessio.franceschelli.me/posts/ai/your-…

Think of it as a day-one briefing for a smart engineer. They don't need "write clean code" or "use plan mode." They need "the User model has a soft-delete flag, always filter by deleted_at IS NULL." Traps, invariants, verified commands.

🧵👇 Your AGENTS​.md (or CLAUDE​.md) is probably full of vibes. Two kinds: motivational philosophy ("Demand elegance", "No laziness") and personal workflow habits ("Always use plan mode", "Track progress in todo​.md"). Neither tells the AI anything about YOUR project.

@aicreatorpath I haven't used it much lately because of this issue. Will play with it over the weekend and report back!

Oh nice, I had the exact same freezing issue with Claude Code on WSL. Was about to switch back to native Windows. The startup hang and slash command lag were killing my workflow. Curious if the fix also helps with memory usage — I noticed Claude Code can get pretty heavy when working on larger repos with lots of context.

🧵👇 Love the Claude Code CLI but frustrated by it freezing and hanging inside WSL? You're not alone. I almost gave up on it, but I found an incredibly simple fix to make it snappy again.

The fix is in the post! Time to have a fun weekend building with Claude Code! Check out the quick blog post for more details. alessio.franceschelli.me/posts/ai/claud…

The problem: In WSL, Claude Code would hang for seconds at startup, stutter while typing, and freeze before every slash command. Borderline unusable.

@csharpfritz by portability I meant more lock in into the tool or I can just naturally reuse the same docs with plain models. Thanks for the info. I will try to dig into this over the weekend, as I am looking into agent teams tooling easy to deploy at scale

Great question! .squad is its own folder because it's not docs or specs - it's live state. Agent memories, decision logs, routing rules, ceremony configs. Docs describe a system; this folder IS the system's memory. Structure matters but it's flexible - agents read specific files (decisions .md, history .md, charters) but you can add custom files freely. Portability: copy the folder to another repo and the team remembers everything. That's the whole point - institutional memory that travels.

I juggle 4-6 projects at once. It used to be chaos. Every time I switched projects, I'd lose context. "Wait, what was I building here? Why did we choose #Aspire for this?" I'd re-explain the whole vision to my team or re-read old notes. Then I realized: my AI agents should remember what I'm building. → That's the problem Copilot #Squad solves. Each project lives in its own `.squad/` folder. When you switch from Project A to Project B, the Squad on Project B already knows: - Architectural decisions you made last month - Why you chose that specific tech stack - What agents are running (and their personalities) - Outstanding work & blockers No re-explaining. No lost context. The team remembers. Right now, I'm shipping a Minecraft server integration with Aspire. Multi-protocol streaming, live game state sync, the whole thing. That Squad file stores every decision: "Why Aspire? Replicated services. Why not Docker only? Needs local dev parity." When I jump back to another project, I don't start from zero. My AI team already has institutional memory. → This scales to 6 projects without your brain melting. The alternative? Prompting the same context into every agent every session. Fragile. Forgettable. Human. Not scalable. Squad bakes memory into the system. That's the move.

@csharpfritz why .squad and not a more standard docs or specs folder? Does the document need to have a specific structure? essentially, how portable is this?

The .squad/decisions.md file is where the team records every major call. Before I ship a breaking change, agents read that file first. No rogue AI decisions. The team stays aligned across sessions.