{
  "$type": "site.standard.document",
  "description": "I adopted skills quite late, but the timing turned out to be right. Skills now work in Droid, Claude Code, and Codex — so I can write them once and",
  "path": "/skills-are-the-missing-piece-in-my-ai-coding-workflow/",
  "publishedAt": "2026-02-02T04:39:00.000Z",
  "site": "at://did:plc:bryys25pc2fnagnyxqgsglhd/site.standard.publication/3mn26bjkkmh23",
  "tags": [
    "AI",
    "Tools"
  ],
  "textContent": "I adopted skills quite late, but the timing turned out to be right. Skills now work in Droid, Claude Code, and Codex — so I can write them once and use them across all 3 agents.\n\nI wrote about my shared agent setup previously. Skills fit right into that — they live in the same shared directory, softlinked so all agents pick them up.\n\nWHAT SKILLS ARE\n\nSkills are small markdown files with a name, description, and a prompt. When you tell the agent to do something that matches a skill's description, it picks up the skill and follows the instructions. Think of them as reusable slash commands that the agent can call.\n\nHere's a simple one I use for committing:\n\n---\nname: commit-succinct\ndescription: Commit code succinctly. When user says \"commit\"\n  or \"git commit\"\n---\n\nCommit the dirty changes with a short, succinct message.\n\nIn practice, the agent doesn't always pick up the right skill — or any skill at all — especially when triggered by single, common words like \"commit\". So I mention in my AGENTS.md that when I say \"commit\", it should use the commit-succinct skill. With that hint, it works reliably — I just say \"commit\" and it stages changes, writes a concise message, and commits.\n\nWHY THEY CLICK\n\nBefore skills, I'd either type the same instructions repeatedly or rely on custom commands (which work differently across agents). Skills are just markdown files with a standard format that all 3 agents understand.\n\nA skill can also call another skill. I have a \"review+fix\" skill that calls a \"review dirty\" skill in a loop — Codex reviews the changes, Droid fixes what's flagged, repeat until clean. I'll write more about that next.\n\nThis makes it easy to build up complex workflows from simple pieces. \"Review+fix, commit, deploy when good\" — and the agent knows what each of those steps means because each one is a skill.\n\nWHEN YOU DON'T NEED THEM\n\nSometimes you need to test something carefully, sometimes you don't. Skills aren't about automating everything — they're about not repeating yourself for the patterns you've already figured out. The agent still needs your judgment for the hard parts.",
  "title": "Skills Are the Missing Piece in My AI Coding Workflow"
}