Skill of the Day: Give Every Agent Task a Fixed Output Path

Today's skill is simple: every agent task that produces an artifact should have a fixed output path.

Many agent tasks fail not because the model is incapable, but because the delivery target is too soft. “Write a science article” sounds clear to a human, but it gives the system no hard acceptance point. A better task says: write the draft to `/reports/daily-pipeline/drafts/science-xxx.md`; the file must include frontmatter, three-locale titles, excerpts, body content, a risk checklist, and next steps. After completion, the controller must run `ls -la` and `wc -l`.

This skill fits any artifact-producing task: documents, reports, SQL drafts, QA verdicts, cover manifests, and publishing checklists. It is less useful for open discussion or early brainstorming, where flexibility matters. But once the work moves into execution, “please complete this” should become “please write this file and make it host-verifiable.”

Checklist:

• Give each task one primary output path.
• Use an absolute path.
• Include date, task name, and track in the filename.
• Specify a minimum byte count or line count.
• Require raw host-side `ls` / `wc` output as success evidence.
• Treat the agent's natural-language conclusion as supporting context, never primary PASS evidence.

The value is reduced ambiguity. No file means not done. A tiny file means not acceptable. A real file with weak content goes to QA. Once the boundary is clear, the whole team becomes easier to operate.