Stop Copying Agent Skills Between Claude Code, Cursor, and Codex — Index In Place Instead
Any engineer running more than one AI coding tool has the problem itlackey describes in part five of his practical skills management series: three tools, three separate skills directories, none of them talking to each other. Claude Code has its skills in ~/.claude/skills/. Codex has a different path. Cursor keeps its rules somewhere else entirely. The result is rebuilding the same deploy scaffold three times, forgetting which version of the database migration helper is actually current, and spending fifteen minutes grepping across directories for the one script that handled the edge case correctly last month.
The series' central framing — that this is a discovery problem, not a tooling problem — is what makes this installment worth reading. Every solution that requires you to move your files creates a new maintenance burden: copies drift, the canonical version becomes unclear, and you're back to searching, just with more directories. The prior parts of the series covered progressive disclosure (giving agents only the skills relevant to the current task), a unified local asset tool called akm, remote context via OpenViking, and community knowledge sharing via Context Hub. Part five synthesizes those pieces specifically for the multi-tool reality that is now the default for serious AI-assisted engineering: indexing in place, making every skill file discoverable from wherever it lives without moving or syncing anything.
The key principle generalizes well beyond personal workflow. Any shared context artifact — coding standards, architectural decision records, runbooks — benefits from discovery rather than synchronization. For teams standardizing on multiple coding tools across a shared codebase, building a skill index is a team infrastructure decision, and "index in place, don't move files" is the principle that keeps it from becoming another maintenance burden six months later.