Documentation first
Product intent and architecture become source material for implementation—not cleanup work after the code exists.
AI coding workflow for production software
AI doesn't needbetter code generation.It needs better engineering.
mstack init installs project planning and ownership; mstack ai setup adds the specialist agents and engineering runtime that turn AI output into software worth shipping.
npm install -g @imisbahk/mstack@latest19 agents20 skills15 AI environmentsv0.5.0 current
01 Why mstack
The model isn't
the bottleneck.
Everyone is arguing about which coding model is better.
GPT.Claude.Gemini.Codex.
But a faster model still cannot rescue the wrong product, an undefined boundary, or a decision nobody wrote down.
The bottleneck is thinking: understanding users, choosing the right scope, designing the system, and verifying what happens when reality disagrees with the happy path.
mstack captures the workflow I use to take software from an idea to production—and makes it available wherever you work with AI.
02 The system
One workflow.
Every engineering layer.
Not another project generator. A maintained set of decisions, specialists, procedures, and safeguards that live with your repository.
AI runtime
One neutral engineering pack for 15 verified coding environments, from Claude Code and Codex to Antigravity, Kimi Code, Copilot, and OpenCode.
Specialist agents
Nineteen focused roles with explicit responsibilities, boundaries, inputs, output contracts, and safe handoffs.
Engineering skills
Twenty reusable procedures covering every lifecycle phase plus architecture, debugging, security, performance, and release work.
Prompts and hooks
Nineteen prompt packs and four local, reviewable automations that reinforce the workflow without hiding it.
Repository bootstrapping
Install the workflow into a new or existing project while preserving files you already own.
Reference templates
Ten practical starting points for discovery, experiments, product, architecture, features, decisions, APIs, and production concerns.
Project validation
Detect incomplete planning, manifest drift, unsafe runtime state, and missing setup before they become release problems.
Safe reconciliation
Recorded ownership, previews, approvals, backups, and recovery make updates inspectable and reversible.
Production workflow
A repeatable path from user evidence to implementation, review, release readiness, and operation.
19specialist agents
20engineering skills
19prompt packs
4automation hooks
10reference templates
15AI environments
03 Multi-agent engineering
One assistant is not
an engineering team.
Build Like This coordinates phase-gated specialists, running independent lanes in parallel where supported and named sequential passes elsewhere while one lead owns synthesis and shared decisions.
Idea
A hypothesis, not a specification.
Discovery lanes
Product and user researchers test the problem, users, needs, and alternatives in parallel where supported, or as named sequential passes elsewhere.
Decision gate
The Product Manager integrates evidence into scope, non-goals, and a measurable outcome.
Design lanes
Architecture, data, security, UX, and operations make bounded decisions against one product definition.
Delivery lanes
Backend, frontend, database, and test specialists work concurrently where supported after contracts stabilize.
Review lanes
Code, security, accessibility, performance, and release evidence are checked independently.
Production
One release owner integrates the evidence, deploys with authority, and feeds learning back into the workflow.
Each specialist has a responsibility, strict boundaries, required inputs, an output contract, and rules that prevent overlapping edits or recursive agent sprawl.
Parallel where supported and independent. Accountable at every gate.04 Build Like This
Code starts
halfway down.
The strongest engineering workflow begins before implementation and continues after deployment. Every step reduces a different kind of uncertainty.
Idea
Start with a belief worth testing, not a stack worth using.
Users
Name the people whose behavior or outcome should change.
Needs
Separate observed problems from plausible assumptions.
Features
Choose the smallest behavior that can prove useful.
Product
Make scope, non-goals, success, and evidence legible in product.md.
Architecture
Define boundaries, contracts, data, security, and failure in architecture.md.
Backend
Implement validated contracts, domain behavior, persistence, permissions, and recovery.
Frontend
Build the complete accessible journey against stable contracts.
Deploy
Verify readiness, release progressively, observe, and recover safely.
Improve
Use outcomes and incidents to return to the earliest phase affected by new evidence.
05 Why it works
Good software is
a chain of good decisions.
Think before generating
Fast code is useful only after the problem, constraints, and intended outcome are clear.
Keep decisions visible
Documents, contracts, and ADRs give humans and agents the same system to reason from.
Use specialists
Product, architecture, security, debugging, and delivery require different modes of judgment.
Verify the whole journey
Success paths are not enough. Permissions, failure, recovery, accessibility, and operations are part of the product.
Don't ask AI to build your app.
Give it an engineering system to work inside.
06 Start here
Put better engineering
in your repository.
Requires Node.js 20.11 or newer. Run init for project planning, then ai setup for agents, skills, prompts, and runtime guidance. Preview every repository change before applying it.
npm install -g @imisbahk/mstack@latest