Enis
282 posts

Enis
@enisdev
developer • web dev enthusiast • svelte lover ❤️




Switching from Claude Code to Codex should take minutes. But for most people, it means starting from scratch (here's why)... Most engineers think they're locking into a harness. But they're actually locked into their context. So every time your assistant learns something, gains new skills, or automates another workflow, it becomes harder to switch. This is why, while building my personal AI assistant for our upcoming Manning book, I've been converging toward one architecture: Make the harness disposable. Keep everything else. Here's what that looks like: 1/ Unified memory Everything the agent knows lives in one place. Instead of scattering memory across multiple systems, I prefer building it on top of a single database that supports: Documents Full-text search Semantic search Knowledge graph traversal For me, that's @MongoDB. One database is enough for the majority of AI assistants. 2/ Serving layer The agent should talk directly to an interface. Typically, that's an MCP server. The database stores the memory. The MCP server owns the business logic. It decides: How memory is searched How memory is updated Which tools the agent can use When conversations become long-term memory This separation is what makes the system portable. 3/ Harness Claude Code. Codex. OpenCode. Gemini CLI. These should be the most replaceable part of your stack. If your context layer is portable, changing harnesses should be little more than a configuration change. The new assistant should immediately know: Who you are What you're working on What you've learned What matters to you Because your memory moved with you. To sum up: The model isn't the moat. The harness isn't the moat. Your context layer is. Design your architecture accordingly. P.S. If you're interested in building an AI assistant that remembers your research, notes, conversations, and preferences across any harness, I break down the full architecture in Decoding AI Magazine Read it here: decodingai.com/p/the-context-…


/grill-me and /grill-with-docs are changing No more "one question at a time" - now it asks questions in rounds Faster, less token spend, but still keeps dependencies between questions clear. Shipping in v1.2




@thsottiaux tl;dr version: - set up CLIProxyAPI with Claude and Codex auth - Connect to Claude Code - Make "claudex" alias that sets some env vars Took like 2 prompts (I already had the proxy set up tbf)






If you set gpt-5.6-sol to "ultra", all the subagents it spawns will also be set to ultra. IMO this is a fumble. Causes massive token burn for no good reason. At the very least, I should be able to hard-set the subagent effort level to "medium". Claude Code is far ahead here

$2,000 WEBSITE. $0 BUILD COST. ONE FOUNDER. No web agency. No freelance developers. No expensive software. Just Claude doing the heavy lifting. The design was finished in hours. The code was generated in a day. Every revision took minutes. The client paid $2,000. The build cost was $0. The value wasn't in writing every line of code. It was knowing how to direct AI to build it.

















