mel-discipline — Claude Code skill: five gates before you ship a brief

What mel-discipline does

The skill lays out five gates, taken in order. First the agent states the task in one sentence, the decision it feeds, who will read the output, and a verification plan, plus the likeliest way the document could mislead that reader. Then it gathers evidence: supplied materials and its own sources come before any web search, with a search-effort ladder running from one query to ten.

Source handling is strict. Nothing enters the deliverable unless the agent opened it during the session; a citation recalled from memory is dropped or tagged as an unverified URL. Claims about what a source says require opening the document itself, not a news piece about it. At the third gate the agent names the strongest objection to its own recommendation and either answers it or concedes.

Gate four is a checklist over the finished draft: trace every citation, sweep the prose for em dashes, confirm sentence one is the verdict, flag data gaps, recompute the numbers, and check length and register. Each check has to be performed, not reported. Gate five is the honest write-up of what was verified, what was not, and which checks were done by eye. Without a tool to run, some mechanical checks stay manual, and the skill says to admit that.