mstack-help

Explain mstack workflow routing, playbooks, and principles without starting the work.

SKILL.md

GitHub
<!-- Generated by mstack from Misbah's Build Like This workflow. Regenerate instead of editing. -->

mstack-help

Answer how to use mstack workflow resources. Do not start the engineering work unless the user clearly asked you to execute.

What to return

  1. A short answer
  2. A copy-pasteable /mstack ... prompt when useful
  3. The playbook or principle ids to trust for details

Map

  • Setup / install / regenerate → tell them to run mstack ai setup (and mstack validate --strict after). Workflow skills install through the core runtime.
  • Start work → /mstack with goal + finish condition + evidence + constraints. Sticky Custom Mode where supported.
  • Read-only question → playbook-investigation
  • Bug / regression → playbook-bug-fix (reproduce first)
  • Slow / metric → playbook-perf-hillclimb
  • New behavior → playbook-feature
  • Structure only → playbook-refactor
  • Spike / decide → playbook-prototype
  • Open PR / ship → playbook-shipping
  • CI / review threads → playbook-babysit-pr
  • Unattended long run → playbook-autonomous-run
  • Parallel lanes → playbook-orchestrate
  • Named principle → load that principle skill; do not paraphrase another project's wording

Playbooks

  • playbook-bug-fix — Bug fix: Reproduce first, confirm mechanism, ship the smallest causal fix with proof.
  • playbook-investigation — Investigation: Read-only investigation that answers with cited evidence and labeled inference.
  • playbook-perf-hillclimb — Perf / hillclimb: Improve a measured metric against a recorded baseline without guessing.
  • playbook-feature — Feature: Deliver a bounded vertical slice from outcome through verified behavior.
  • playbook-refactor — Refactor: Improve structure while preserving externally observable behavior.
  • playbook-prototype — Prototype: Run a cheap experiment that settles a decision, then throw away or promote deliberately.
  • playbook-shipping — Shipping / open PR: Open or land a verified change with an honest PR and proof.
  • playbook-babysit-pr — Babysit PR / CI: Drive a PR to merge-ready by clearing conflicts, threads, and CI in order.
  • playbook-autonomous-run — Long autonomous run: Define a checkable exit predicate and drive to it without stalling on reversible work.
  • playbook-orchestrate — Orchestrate parallel subagents: Fan out independent lanes, serialize shared surfaces, integrate with acceptance owned by the lead.

Principles

  • prove-it-works — Verify against the real artifact before claiming done.
  • fix-root-causes — Trace symptoms to the violated invariant; ship the smallest causal fix.
  • guard-the-context-window — Keep the lead thread lean; push bulk work into bounded subagents.
  • never-block-on-the-human — Proceed on reversible work; ask only for irreversible or preference calls.
  • encode-lessons-in-structure — Turn repeated verbal guidance into checks, types, docs, or scripts.
  • thinking-before-typing — Define outcome, evidence, and boundaries before generating implementation.
  • user-before-feature — Start from the user's current difficulty, not an exciting capability.
  • contracts-before-clients — Stabilize backend behavior before frontend integration.
  • smallest-complete-slice — Ship the narrowest vertical slice that proves useful behavior.
  • validate-every-boundary — Treat external input as untrusted and validate once at the edge.
  • evidence-over-optimism — Optimize and extract complexity only with measured need.
  • reversible-by-default — Prefer decisions that are easy to undo, migrate, or roll forward.
  • make-failure-observable — Failures must be visible, bounded, and recoverable.
  • name-the-trade-off — Every consequential choice needs a reason a stranger can audit.

If the message is actually a work request ("fix this bug with mstack"), switch to mstack and say once that sticky Custom Mode keeps the router on.