deliverable

pmstack lets product managers build an evaluation lifecycle inside Claude without starting from an ML engineering dashboard. Its commands turn observed failures into task suites, graders, pass bars, and runnable YAML, then score a real target, review failed transcripts, and watch for regressions after model changes. PRDs, customer-signal synthesis, briefs, and launch checks are included, but the differentiator is making AI evals directly usable by PMs.

Founder Mode turns a set of founder-led operating ideas into 12 Claude Code commands for audits, org health, hiring, delegation, reviews, and decision checks. Each command produces a scored diagnosis, pressure-test questions, and concrete next steps, helping founders decide where deeper involvement is useful and where delegation remains appropriate rather than treating one management style as universally correct.

Shiplog watches Git commits to discover features, update progress, and retain architecture decisions and coding patterns so a new Claude Code session can continue without a full re-explanation. After each commit it can run background agents for specs, PRD status, security review, and memory. Both freeform and planned tracking modes are available, all records stay local, and the project must use Git.

pmcontext structures Claude Code work into Quick, Standard, and Full workflows with explicit approval gates before implementation advances. It stores decisions, risks, session receipts, and execution state in the user’s own Supabase project so work can resume across projects and machines. Standard and Full workflows depend on the superpowers plugin; context7 MCP is additionally required when a plan touches external libraries.