Phase 3 of Midas — turn the validated opportunity into a go/no-go business case. Define the value proposition, MVP scope vs explicit non-goals, the business/monetization model, and MEASURABLE success metrics (which later phase-8 audits grade against), then capture an explicit go/no-go with HUMAN sign-off into…
Phase 4 — pin stack and architecture from the business case, Context7-verify versions, write {product}/architecture.md and one ADR per decision. Use after business case gate passes (stage businesscase → techarchitecture), before rules or code.
Phase 8 — per-sprint conformance and scope audit. Diff the code against every frozen rule (pass/fail with evidence), reconcile scope vs the business case, resolve drift, freeze {runs}/audits/audit-NN.md, update state, and select the next sprint or declare MVP complete. Use after a sprint's work lands (stage stays…
Phase 1 of Midas — the gap loop. Generate and rank blocking questions, ask them in batches, fold answers into {product}/idea.md, track {product}/open-questions.md, and loop until zero blockers remain. Use after idea-intake to pin down user, problem, metric, and non-goals.
Phase 5 keystone — freeze architecture into checkable rules, design system, playbooks, and enforcement tooling; re-render adapters. Use once after techarchitecture passes, before sprint work.
Phase 0 of Midas — capture the raw product idea verbatim, normalize it into {product}/idea.md with a one-line pitch and mode, and initialize/advance paths.state. Use to start a new product or record its founding idea.
Phase 2 of Midas — validate the (now-clear) idea against the real market. Derive research questions, fan out web searches (optional host deep-research if installed), adversarially verify every claim with citations, and synthesize a competitor matrix + differentiation thesis + top risks into {product}/market.md. Use…
Adopt Midas into an EXISTING (brownfield) project. Inventory the codebase, reverse-engineer the de-facto architecture and rules from the real code, backfill product context, and establish a baseline audit — writing into any pre-existing AGENTS.md/host adapter/source only after a dry-run diff you confirm. Use instead…
Propagation alignment pass — maps what changed to downstream surfaces (adapters, bundles, version stamps, docs, tests), runs the cheapest verify ladder, and reports gaps before merge or sprint close. Use after editing skills, rules, conventions, installer, VERSION, or any generated tree source. Complements…
Export or import Midas project knowledge as portable JSON — state, product lifecycle docs, stack rules, playbooks, frozen evidence, MCP/enforcement config (no secrets), optional tests. Use to seed a new project, share a subset between repos, or backup selective memory. Complements git; does not replace it. Runs…
Turn a recurring request or correction into a project rule ( /), playbook ({product}/playbooks/), or convention ({product}/conventions.md). Use when the user asks to capture a pattern, or after they confirm a propose-to-capture (2–3 repeats). Never write silently.
Product-authentic UI redesign — audit, three art directions, human pick, spec, then optional one-slice implement. Use when the user asks to improve, redesign, or refactor visuals/landing/UI; never jump straight to JSX. Distinct from /define-conventions (Phase 5 freeze) and /midas-verify (proof).
The sync engine and health check — re-derives the generated tool adapters from /conventions.md + rules, diffs them against disk, reports drift, and offers to re-render. Run after editing conventions/rules or when adapters look stale.
Ad-hoc investigation outside the 9-phase pipeline — notes under {runs}/explore/ /. Use for debugging or scoping that is not a phase gate or sprint task; close with --end.
Interactive intent→command guide — one AskQuestion, then What/Exact command/What happens/When NOT/Next for that option only. Use on /midas-help or when the user asks which Midas command to run. Distinct from /midas-status (pipeline PC).
Product-repo hygiene orchestrator — dead flows, orphans, ledger/doc drift, and optional lean delete-list. Path-passes midas-sweep (scope product) + midas-lean-review. Not adapter/doctor sync. Use when the repo is dirty, before close on large diffs, or after brownfield adopt.
Onboarding entry — diagnose install/setup/version, tip install or pinned update, or run one-time adaptive intake (optional --monorepo). Use when starting a project, setupcomplete is false, or engine is behind.
Root-cause investigation before bug fixes — Iron Law + 3 strikes; freeze {runs}/investigate/inv-NN.md. Use when debugging failures, after failed self-fixes, or when asked to investigate. Complements /midas-explore (open-ended) and verification.md inner loop.
Engine-only precommit quality bar — scores architecture, security, agentic design, tests, reliability, docs, simplicity, DX, code quality, maintainability, change propagation, and methodology; requires overall >= 80 before commit. Use before committing on midas-harness. Not for product installs.
Read-only context pack for resuming work — reads paths.state and assembles 15 priority paths plus a 30-line brief (where am I, what matters, open gaps). Scout-tier; no writes. Use after a break, new session, or when midas-status suggests recall. Distinct from midas-status (PC only) and midas-sweep (hygiene).
Read-only install/orientation check — thin guide to the deterministic diagnose CLI. Detects missing install, setup pending, version behind, or wrong cwd, and prints the single next CLI or slash command. Use when npx update failed, you are unsure between install/init/adopt/update, or before /midas-status on a confused…
Read-only sprint retrospective — freeze learnings under {runs}/retros/retro-NN.md without advancing stage. Use after a sprint closes or when the team wants a gstack-/reflect-style note. Complements /close-sprint (conformance) and /midas-progress (STM).
Engine-only sandbox — dry-runs the real, unmodified harness/skills/ against a small nested example product (sandbox/example-product/) on Cursor's composer-2.5 model (never composer-2.5-fast), with a traced decision-flow log and reviewable findings under sandbox/findings/. Use before committing a skill/rule change on…