Affects everyone running build-proposal
It confirms your ask before researching it
The bug was real and observed. Given this request:
should we add a shared telemetry and usage-analytics layer across all 13 plugins so we can see which skills and commands actually fire in real sessions
the tool restated it as the same thing plus a goal nobody had mentioned:
… and let that data drive keep/merge/cut decisions.
It then researched against its own invention — a dispatched search agent, twenty-odd file searches, all before the human saw a word. Two rules now prevent this. Restate, do not enrich forbids adding a motive you never stated: if one seems missing, that is a clarifying question, not a blank to fill. A new checkpoint then presents the objective and waits for Looks right or Refine it, capped at two rounds.
Affects everyone finishing a proposal
Publishing is always offered, never assumed
Every finished proposal now gets one question: publish this as a shareable page, yes or no? Nothing goes out without an explicit yes.
The subtlety is what cannot turn the question off. A blanket “no further questions” covers the proposal's own decisions, not this one — publishing is the only step in the workflow you cannot undo by editing a file, so it keeps its own confirmation.
Affects everyone running write-prompt
You can name the prompt type directly
The first word of your request now sets the type, if it is one of system, task, creative, or analytical:
claude "/plan-agent:write-prompt creative a bedtime story about a lighthouse keeper"
The convention already existed, but as a private arrangement between two tools, documented for exactly one type. It is now a real feature for all of them, advertised in the command's own argument hint.
This matters because the type is not cosmetic — it picks both the questions you get asked and the techniques applied to the draft. Guess wrong and you answer a batch of wrong questions before the mismatch is visible. When no type is named and the classifier is confident, a new checkpoint asks before any of that starts.
Affects deep runs — you wait less
Research actually runs in parallel now
The instruction said to run the codebase search and the web research together. In practice the tool ran them one after the other, with 21 sequential file searches wedged in between.
The old wording prescribed one specific technique. The new wording permits either route to concurrency and instead forbids the single setting that broke it.
Affects teammates indirectly
The two biggest instruction files were cut down
An instruction file is paid in full every time its tool runs — there is no partial load — and a typical run reads maybe a quarter of it. Both of the plugin's largest files now keep only the rules in the core and push the mechanics into reference files that load on demand.