🧂 MSG

Sh*t tastes so good, it's probably loaded with MSG.

The counterpart that relies on /cook, it's a heavily opinionated coding agent workflow and harness that depends on human approvals more than autonomy.

💻 Install

msg + cook (recommended)

curl -fsSL https://raw.githubusercontent.com/ndisisnd/msg/main/install.sh | bash -s -- --with-cook

msg only

curl -fsSL https://raw.githubusercontent.com/ndisisnd/msg/main/install.sh | bash

🗂️ Skills

Run /msg to browse these interactively, or invoke any skill directly. /msg --gui opens a local, Notion-style PRD board (Kanban/table, light + dark) where you can edit PRDs, drag statuses, tick off todos, work the INTAKE.md backlog on an Intake tab (grade chips, lane drag), browse project docs, read run reports, and run Claude prompts — served on 127.0.0.1 only.

Run reports. eng --build, /pre-merge, and /post-merge each end a run by writing report-[n].md into the PRD's features/prd-[n]/reports/ folder (features/reports/ when no PRD applies) — a plain-language record of the work done (features, code changes, lines added/deleted, tests passed/failed) plus what you can expect and the exact steps to verify the feature works. Post-merge's staging report carries the human test script; its production report renders release-style with any no-rollback platform flagged IRREVERSIBLE. The board renders them under a dedicated Reports tab, grouped by PRD. Schema: .claude/skills/shared/refs/report-schema.md.

Safety floor. The safety floor is never relaxed: write powers are scoped per skill (eng commits to feature branches only; pre-merge opens exactly one feature→staging PR and never merges; post-merge is the only merger, and nothing reaches main except via its double-confirmed staging→main release), and the human gates (preview approval, staging sign-off, production double-confirm) never disappear. /post-merge --production guards main behind branch protection (green CI + human review). See ARCHITECTURE.md § Safety floor.

📐 Plan

Skill	Description
/msg --init	One-time project bootstrap — batched interview (project basics, architecture, design system, release flow), then scaffolds devkit/ (AHA.md, GLOSSARY.md, ARCHITECTURE.md, DESIGN-SYSTEM.md, OPEN-QUESTIONS.md), seeds devkit/policy.json (init:false + release flow), the root INTAKE.md backlog ledger, and root files. Idempotent.
/msg --init-staging	Adds a staging branch to a direct-flow repo (branches off prod, offers branch protection) and flips devkit/policy.json release flow to staged. The only mode that creates a staging branch.
/intake	The planning front door — captures feature ideas and bugs as graded rows in the root INTAKE.md ledger. Owns the requirements interview: fleshes out thin ideas, suggests adjacent ones, splits compound/hybrid asks and XL ideas into discrete rows, and grades each in a single-turn banded judgment (complexity C: / token-cost T: / sequencing S: — bands only, never fake-precise numbers). Feeds plan-pm.
/intake --delete	Removes a row from the backlog — the only destructive intake mode. Runs a warning pass first (a PRD it would orphan, other rows graded S:blocked-by-#n against it, a completed ship record, log history), then requires an explicit confirm. Never renumbers — the # gap stays, because renumbering would silently repoint every blocked-by reference. Deletes ledger rows only, never a PRD folder or file.
/intake --update	Edits a row already in the backlog. Lists every un-shipped row in full, then changes the one you pick — idea / goal / type on backlog rows only (in-progress rows belong to their PRD; use /plan-tune). A material change re-runs the grading rubric and the same hybrid/≥8 split gates as capture; a cosmetic one keeps the grade. Every edit lands in INTAKE.md's ## Update log. Never writes status or prd.
/plan-pm	Principal PM — the autonomous PRD writer. Consumes a graded intake row and drafts the full PRD solo (edge cases, feature/acceptance table, user flows, error handling) to features/prd-[n]/, pausing only for batched open questions and breaking/critical touches. Stamps the intake row in-progress + its prd mapping. --roadmap analyses the existing PRDs (flagging bloat/overlap, proposing approval-gated split/merge/fold/trim, reading the intake S: grades) and sequences them into phases in roadmap/roadmap.md, viewable on the /msg --gui Roadmap tab.
/plan-tune	Staff PM contract certifier — runs a fixed seven-check certification (--product: checks 1/2/3/6; --eng: 2/4/5/6/7), each tied to a named downstream consumer ("no check without a consumer"). Auto-selects the tune type, auto-fixes every Critical + Major with a terminal # | Sev | Found | Fixed table, asks once about Minors, pauses only on a product-decision finding. Each auto-fix writes a category-tagged learning to devkit/AHA.md so the next plan-pm draft self-heals.
/plan-em	Engineering Manager — auto-runs plan-tune certification before each wave (product before plan, eng before build; roster approval is the single human gate), spins up specialist agents to write engineering sections into the PRD, then synthesises the output.

🔨 Build

Skill	Description
/eng	Platform-agnostic engineering agent — --plan proposes file changes for approval and writes the per-feature todo tickets in the same pass, --build writes code from the todos (falling back to exec-table rows). --build --loop adds a plan-tune review cycle after each build pass. --build roadmap=roadmap/roadmap.md runs an autonomous product-operations orchestrator that executes a whole roadmap phase-by-phase via eng/pre-merge subagents — fixing critical+major by default, guarding production (DB/data/config pauses, branch-isolated, never pushes/merges), and reporting on an interval.
/pre-merge	The CI gate — takes a feature branch from "eng says done" to "PR open against staging". Runs a preflight-driven pipeline executor over the components[] manifest in devkit/policy.json: it prunes to the components this project actually has, topo-sorts them on their dependencies, and runs independent components as parallel waves — sync → parallel correctness + security waves → coverage → regression tail → smoke health-check → the merged preview human-review gate → opens PR feature→staging. preview is the sole human gate in pre-merge: it absorbs the retired qa (visual capture folded in), stands up an ephemeral, isolated, pokeable env, and serves one unified Approve/Reject artifact carrying the machine evidence + the significance-rated manual-test-plan checklist. smoke is its health precondition — a fired preview always gets at least a default-liveness check plus the critical-path golden flows, runs first, and short-circuits so no one is ever asked to approve a dead preview. With no manifest it refuses no_manifest and points you at /pre-merge --init. Absorbs the old /review and /test; emits a severity-graded JSON verdict. --init detects the pipeline — including whether a .github/workflows/ pipeline runs the gate on PRs — offers to install/scaffold the missing (free/OSS) pieces, and writes the manifest; --update reconciles it as the code drifts. (--doctor is a deprecated one-release alias for --init.)
/post-merge	The ship gate — the only skill that merges. --staging verifies green CI, merges the feature→staging PR, deploys staging, smoke-verifies the deploy (smoke_cmd from PLATFORMS.md), emits a human test script, and stamps staging-signoff: on approval. --production double-confirms, opens the release PR staging→main (rollback notes per platform; iOS IRREVERSIBLE), merges on green CI + human review, then deploys and smoke-verifies the live target. Branch protection (post-merge-protection.sh) is the machine enforcement — now policy-conditional (enforced/optional/skip per devkit/policy.json, so private Free repos that can't set protection aren't blocked). --init sets up protection/deploy tooling + the release flow, and guards the protection offer on a CI workflow existing (so it never requires status checks nothing produces). (--doctor is a deprecated one-release alias for --init.) Ship gates never collapse.

---

Credits to my dear JC who previously had her own harness with a bajillion agents. Great times.