DEV Community

Manuel Bruña
Manuel Bruña

Posted on

Why I stopped copying project context into every AI tool

Why I stopped copying project context into every AI tool

A quick personal note, because this one finally clicked for me.

For months my repos collected folders: .claude/, .cursor/, .codex/, .windsurf/. Each one held roughly the same thing — roles, rules, a bit of project memory — just written slightly differently and drifting apart every week. Update one, forget the others. New teammate opens the repo in a different editor and the agent suddenly "knows" less.

The fix I landed on is simple to say: the knowledge belongs to the project, not to the tool.

That is the whole idea behind APC (Agent Project Context): one neutral, repo-owned context layer (AGENTS.md + .apc/) that any agent can read, instead of N tool-specific folders fighting to be the source of truth. Runtime stuff — raw sessions, tokens, IDE prefs — stays local and out of the repo. MCP still handles external tools. APC just answers "what should an agent know about this project?"

I wrote a full, visual deep-dive on the migration — the fragmentation cost, the project-vs-runtime boundary, the canonical .apc/ structure, model: inherit to dodge vendor lock-in, and a two-phase migration that doesn't force you to delete anything on day one.

Read the complete walkthrough here:

👉 https://dev.to/agentprojectcontext/migrating-to-apc-one-unified-context-for-your-ai-agents-4ac8

If you juggle more than one AI coding tool in the same repo, I'd love to hear what you'd put in shared context versus what you'd keep tool-specific.

Top comments (0)