PM Butler

AI-assisted building collapses the front half of the software lifecycle. People go from idea to "press enter and let it build" without validating the problem, defining scope, or recording decisions. The failures show up later as scope creep, rework, and no record of why anything was decided. Existing project management tooling assumes humans doing the planning, not an agent doing the building.

Solo build, distributed as a skill, so no server and no UI.
Had to work inside Claude Code and Cursor, where the user already is.
Decisions had to outlive the chat session.
Advice had to stay advisory. The tool cannot own the user's decisions.

One general-purpose PM prompt: no phase discipline, and it drifts.
Standalone PM web app: pulls people out of their workflow and loses context.
Templates only: passive. Nothing challenges weak evidence or scope.

1. One persona per lifecycle phase. Oracle, Sculptor, Architect, Alchemist, Gatekeeper, Shipper, and Librarian, coordinated by a Head Butler. Each owns a phase with explicit questions and outputs.
2. A fixed handoff chain. The Librarian hands back to the Oracle, so lessons feed the next discovery round.
3. Severity gates. Concerns are tagged Watch, Concern, or Blocker, and a Blocker prevents a phase from closing.
4. Evidence over assertion. In discovery, "I think" does not count. Thin evidence raises a flag, and skipping discovery requires an explicit override that gets logged.
5. Decisions live in git. Project Edition scaffolds PROJECT.md plus eleven markdown files covering decisions, scope, risks, and lessons. Optional structured memory uses a folder or SQLite backend.
6. Coach, don't decide. Personas advise and the user decides. A crisis mode drops the persona flourishes for clarity.
7. Distinct voices, deliberately. The personality makes the pushback memorable and keeps people engaged with the process.

Flow: user, Head Butler routing, persona, structured output, Project Edition markdown, git, with an optional memory backend.

Limitations: it structures and challenges thinking, but it cannot verify evidence a user invents. Its value depends on users actually committing the logs.

Target: run it on a real project and track decisions logged per project. No adoption or effectiveness figures have been measured yet.

I built this in about three weeks and have had little user feedback so far. The open question is whether the gates change how people work on real projects, and that is what I plan to measure next.