# π Your first project
**From an empty GitHub repository to a shipped pull request β in one guided loop.**
> [!TIP]
> **This is the one guaranteed path.** You will build a tiny app, run **every**
> command in the core loop exactly once, and β the part most tutorials skip β
> understand *why* each step exists. This tutorial uses **Claude Code**; MSD
> works in 15+ runtimes β see [Install on your runtime](../how-to/install-on-your-runtime.md)
> for the flag and command syntax for yours.
---
## π Table of contents
1. [The one idea that makes MSD click](#-the-one-idea-that-makes-msd-click)
2. [What you'll build](#-what-youll-build)
3. [Prerequisites](#-prerequisites)
4. [Step 1 β Install MSD Core](#step-1--install-msd-core-into-your-runtime)
5. [Step 2 β Open Claude Code](#step-2--open-claude-code)
6. [Step 3 β Create the project](#step-3--create-the-project)
7. [Step 4 β Discuss Phase 1](#step-4--clear-context-then-discuss-phase-1)
8. [Step 5 β Plan Phase 1](#step-5--plan-phase-1)
9. [Step 6 β Execute Phase 1](#step-6--execute-phase-1)
10. [Step 7 β Verify the work](#step-7--verify-the-work)
11. [Step 8 β Ship it](#step-8--ship-it)
12. [Glossary](#-mini-glossary) Β· [Troubleshooting](#-troubleshooting) Β· [What next](#-what-next)
---
## π‘ The one idea that makes MSD click
MSD Core does **not** "write your whole app in one shot." It runs a **repeating
five-step loop**, and it does the heavy thinking in **fresh, throwaway
sub-agents** so your main chat window never fills up with clutter β the quality
killer MSD calls [context rot](../explanation/context-engineering.md).
You drive that loop **one phase at a time**:
```mermaid
flowchart LR
D([π¬ Discuss]) --> P([π Plan])
P --> E([βοΈ Execute])
E --> V([β
Verify])
V --> S([π Ship])
S -. next phase .-> D
classDef step fill:#1f2430,stroke:#f0883e,stroke-width:2px,color:#e6edf3;
class D,P,E,V,S step;
```
| Step | Command | In one sentence | Typical time |
|:----:|---------|-----------------|:------------:|
| π¬ **Discuss** | `/msd-discuss-phase` | MSD asks *how* to build it and writes your answers down. | 2β4 min |
| π **Plan** | `/msd-plan-phase` | MSD splits the work into small, checkable task plans. | 1β5 min |
| βοΈ **Execute** | `/msd-execute-phase` | Fresh agents write the code and commit each task. | 2β6 min |
| β
**Verify** | `/msd-verify-work` | MSD walks you through "does it actually work?" | 1β3 min |
| π **Ship** | `/msd-ship` | A pull request is opened for you. | <1 min |
> [!NOTE]
> **Keep that table handy.** Whenever you feel lost, ask yourself one question:
> *"which step of the loop am I on?"* That's the entire mental model.
π§ Why fresh sub-agents? (the 30-second version)
A single long chat slowly degrades: the more it holds, the more the model
juggles, and quality quietly drops. MSD sidesteps this by spawning a **clean
200k-token worker** for each heavy job (research, execution) and throwing it away
after. Your main session stays lean; the shared `.planning/` files carry memory
between them.
```mermaid
flowchart TB
subgraph main [Your main session Β· stays lean]
you([You + MSD])
end
subgraph workers [Fresh sub-agents Β· clean context each time]
r1[Researcher]
r2[Researcher]
ex[Executor A]
ey[Executor B]
end
you -- spawns --> r1 & r2 & ex & ey
r1 & r2 & ex & ey -- write --> plan[(.planning/ files)]
plan -- memory --> you
classDef m fill:#132a1a,stroke:#3fb950,color:#e6edf3;
classDef w fill:#1f2430,stroke:#58a6ff,color:#e6edf3;
classDef f fill:#2d2410,stroke:#f0883e,color:#e6edf3;
class you m; class r1,r2,ex,ey w; class plan f;
```
---
## π― What you'll build
A small **Node.js command-line to-do app**:
```bash
todo add "buy milk" # β add an item
todo list # π see open items
todo done 1 # β
complete item 1
```
Items live in a local `todos.json`. It uses **only the Node.js standard library**
β nothing to install, nothing to configure β so you focus entirely on the MSD
loop, not a toolchain.
> [!TIP]
> Small on purpose. Once the loop is muscle memory, the *exact same* eight steps
> scale to a real multi-phase product.
---
## β
Prerequisites
| You need | Check with | "Good" looks like |
|----------|------------|-------------------|
| **Node.js 22+** | `node --version` | `v22.x.x` or higher |
| **npm 10+** | `npm --version` | `10.x.x` or higher |
| **Claude Code** | `claude --version` | installed and opens |
| **Git** | `git --version` | installed |
| **GitHub CLI** | `gh --version` and `gh auth status` | installed and authenticated |
| **An empty GitHub repository cloned locally** | `git remote get-url origin` | your repository's GitHub URL |
| **Internet** | β | available for installation and GitHub operations |
If you do not already have an empty repository, create and clone one now. If
`gh auth status` says you are not logged in, run `gh auth login` first.
```bash
gh repo create msd-todo-tutorial --private --clone
cd msd-todo-tutorial
git remote get-url origin
```
```mermaid
flowchart LR
A[node --version β₯ 22?] -->|no| A1[Install/upgrade Node] --> A
A -->|yes| B[npm --version β₯ 10?]
B -->|no| B1[Install/upgrade npm] --> B
B -->|yes| C[Claude Code installed?]
C -->|no| C1[Install Claude Code] --> C
C -->|yes| D[gh authenticated?]
D -->|no| D1[Run gh auth login] --> D
D -->|yes| E[origin points to GitHub?]
E -->|no| E1[Create or clone a repository] --> E
E -->|yes| R([β
Ready for Step 1])
classDef ok fill:#132a1a,stroke:#3fb950,color:#e6edf3;
class R ok;
```
---
## Step 1 β Install MSD Core into your runtime
From a terminal **in your project directory**, run the installer:
```bash
npx @golem15/msd-core@latest --claude --local
```
This tutorial uses `--local`, so MSD is installed only in this project. On a
different runtime? See [Install on your runtime](../how-to/install-on-your-runtime.md)
for its flag. You'll see a summary of what was installed, e.g.:
```text
β Installed 71 commands to commands/ (msd-.md flat form)
β Installed agents
```
Then **restart Claude Code** so it picks up the new commands and agents.
π‘ What just happened?
A `.claude/` directory in your project now holds MSD's **commands** and
**agents**. You never edit these by hand β the installer owns them and keeps them
in Claude Code's native format.
> [!WARNING]
> **Don't copy files from `agents/` or `commands/` directly** β that bypasses the
> installer's transformations and produces schema errors or missing commands.
> Always use the installer.
---
## Step 2 β Open Claude Code
For this disposable tutorial project, open (or restart) Claude Code in your
project directory with:
```bash
claude --dangerously-skip-permissions
```
You'll land at a prompt ready for input.
> [!CAUTION]
> **The permissions flag is optional.** It skips per-file confirmation while
> MSD's sub-agents read and write files. Use it only for this throwaway tutorial
> in an empty folder. To keep confirmations enabled, start with `claude` instead.
> For real work, read the [security model](../explanation/security-model.md) first.
π‘ What just happened?
Claude Code opened in the project where Step 1 installed MSD. It loaded the
project-local `.claude/` commands and agents, so `/msd-*` commands are now
available at the prompt.
---
## Step 3 β Create the project
At Claude Code's prompt:
```text
/msd-new-project
```
The first question is always **"What do you want to build?"** Paste this:
```text
A Node.js CLI tool for managing to-do items. Users run `todo add "buy milk"`,
`todo list`, and `todo done 1`. Items are saved to a local todos.json file.
No external dependencies β Node built-ins only.
```
Then answer the **clarifying questions**, choose **No** when asked *"Research before planning each phase? (adds tokens/time)"* (skip it for this small build), take the
**recommended defaults** for workflow settings, and wait for the **roadmapper**
(~1 min). Type **Approve** on the proposed roadmap:
```text
Proposed Roadmap
1 phase | 4 requirements mapped | All v1 requirements covered β
| # | Phase | Goal | Requirements |
|---|----------|----------------------------------------|-----------------|
| 1 | Core CLI | add / list / done commands, todos.json | CLI-01 β¦ CLI-04 |
```
π‘ What just got created in .planning/?
```mermaid
flowchart TD
root[.planning/]
root --> PROJECT[PROJECT.md
your description + requirements]
root --> REQ[REQUIREMENTS.md
a REQ-ID per capability]
root --> ROAD[ROADMAP.md
Phase 1 Β· status: pending]
root --> STATE[STATE.md
session memory Β· where you are]
root --> CFG[config.json
your workflow settings]
classDef f fill:#1f2430,stroke:#f0883e,color:#e6edf3;
class root,PROJECT,REQ,ROAD,STATE,CFG f;
```
These files are MSD's **shared memory** β they survive `/clear`, survive closing
your laptop, and let a fresh sub-agent pick up exactly where the last left off.
π **Do this now:** open `.planning/ROADMAP.md`. Phase 1 has a **Goal**,
**Requirements**, and **Success Criteria** β the observable behaviors execution
must deliver. This file is your map for the rest of the tutorial.
---
## Step 4 β Clear context, then discuss Phase 1
MSD is built around **fresh contexts**. Clear the window before each phase:
```text
/clear
```
Then open the discussion:
```text
/msd-discuss-phase 1
```
MSD asks about your **implementation preferences** β *how* to build, not just
*what*:
```text
> How should done items be stored β mark them in place or move them?
Mark them in place with a "done" flag.
> Should `todo list` show completed items by default?
No, hide them unless --all is passed.
> What if todos.json doesn't exist yet?
Create it silently on first add.
```
It writes `.planning/phases/01-core-cli/01-CONTEXT.md`.
π **Do this now:** open that file β find `## Implementation Decisions`. Those are
your words, captured. The planner reads this next, so every decision here flows
into the task plans.
> [!NOTE]
> **Why discuss before planning?** Decide the small stuff up front and the plan is
> right the first time β instead of you correcting a wrong plan, choice by choice.
π‘ What just happened?
`/clear` discarded the old chat context, then `/msd-discuss-phase 1` captured
your implementation choices in `01-CONTEXT.md`. The next planner receives those
decisions without needing the earlier conversation.
---
## Step 5 β Plan Phase 1
```text
/msd-plan-phase 1
```
```mermaid
sequenceDiagram
participant You
participant MSD
participant PL as Planner
participant PC as Plan-checker
You->>MSD: /msd-plan-phase 1
MSD->>You: Research before planning Phase 1: Core CLI?
You->>MSD: Skip research
MSD->>PL: 01-CONTEXT.md
PL-->>MSD: atomic task plans
MSD->>PC: verify each plan hits the goal
PC-->>You: plans saved β
```
MSD asks **"Research before planning Phase 1: Core CLI?"** β choose **Skip research** (same
as Step 3; this build is small and well-understood). A **planner** then turns
`01-CONTEXT.md` into **atomic task plans**, and a **plan-checker** verifies each
before saving.
π‘ What just got created?
```text
.planning/phases/01-core-cli/
01-01-PLAN.md β Task: todos.json read/write helpers
01-02-PLAN.md β Task: add / list / done commands
```
π **Do this now:** open `01-01-PLAN.md`. Inside the `` block: a name, the
files it touches, action steps, a `` command, and a "done" condition.
That `` isn't decoration β the executor runs it after writing code.
---
## Step 6 β Execute Phase 1
```text
/msd-execute-phase 1
```
MSD groups plans into **waves** (independent plans run in parallel), spawns a
**fresh 200k-context executor per plan**, and commits each task atomically:
```text
Wave 1 (parallel):
[Executor A] β 01-01-PLAN.md (read/write helpers) β committed
[Executor B] β 01-02-PLAN.md (CLI commands) β committed
[Verifier] Checking codebase against phase goals...
CLI-01 todo add β CLI-03 todo done β
CLI-02 todo list β CLI-04 --all flag β
Status: PASS
```
**Run your app** β your first visible result:
```bash
node todo.js add "buy milk"
node todo.js add "write tests"
node todo.js list # β both items
node todo.js done 1
node todo.js list # β only "write tests"
```
π Item 1 disappears from the default list after `done`. It works.
π‘ What just got created?
```text
.planning/phases/01-core-cli/
01-01-SUMMARY.md β what Executor A built + committed
01-02-SUMMARY.md β what Executor B built + committed
01-VERIFICATION.md β requirement coverage: PASS
```
---
## Step 7 β Verify the work
```text
/msd-verify-work 1
```
MSD reads the phase's `SUMMARY.md` files and turns their user-visible
deliverables into checkpoints. It presents one checkpoint at a time; the first
one looks like this (the test wording depends on what was built):
```text
### CHECKPOINT: Verification Required
**Test 1: Add a to-do**
Running `node todo.js add "buy milk"` creates a pending item without errors.
---
**Type `pass` or describe what's wrong.**
```
Type `pass` when reality matches, or describe what differs. MSD records the
answer in `01-UAT.md` and then presents the next checkpoint.
If a check **fails**, MSD diagnoses the root cause and writes a fix plan β re-run
`/msd-execute-phase 1`, then `/msd-verify-work 1` again. (Result:
`.planning/phases/01-core-cli/01-UAT.md`.)
> [!NOTE]
> **Why a separate verify step?** "The code was written" and "the code works" are
> different claims. Verify proves the second one *before* you open a PR.
π‘ What just happened?
MSD turned the phase summaries into user-visible checks, recorded your answers
in `01-UAT.md`, and routed any failure back through a concrete fix plan. Shipping
only starts after those checks match reality.
---
## Step 8 β Ship it
```text
/msd-ship 1
```
MSD creates a pull request with a generated body (Summary Β· Changes Β·
Requirements Addressed Β· Verification Β· Key Decisions):
```text
Pull request created: https://github.com/your-org/your-repo/pull/1
Title: Phase 01: core-cli
```
That's the **full loop** β idea β opened PR β for one phase, start to finish. π
```mermaid
flowchart LR
idea([π‘ idea]) --> loop
subgraph loop [one phase]
direction LR
d[Discuss]-->p[Plan]-->e[Execute]-->v[Verify]-->s[Ship]
end
loop --> pr([β
Pull request])
classDef a fill:#1f2430,stroke:#f0883e,color:#e6edf3;
classDef ok fill:#132a1a,stroke:#3fb950,color:#e6edf3;
class d,p,e,v,s a; class pr,idea ok;
```
π‘ What just happened?
`/msd-ship 1` assembled the completed phase's requirements, decisions, and
verification evidence into a pull request. The PR is open for review; nothing
has been merged into the default branch yet.
---
## π Doing more than one phase
For a multi-phase project, repeat **Steps 4β8** for each phase. Not sure what's
next? Let MSD detect it:
```text
/msd-progress --next
```
---
## π Mini-glossary
| Term | Meaning in MSD |
|------|----------------|
| **Runtime** | Your AI coding tool. This tutorial uses Claude Code. |
| **Phase** | One slice of the roadmap you take through the whole loop. |
| **The loop** | Discuss β Plan β Execute β Verify β Ship. |
| **Sub-agent** | A fresh, throwaway worker MSD spawns for research or execution. |
| **Context rot** | Quality decay as the main window fills up; fresh sub-agents prevent it. |
| **`.planning/`** | MSD's shared memory: PROJECT, REQUIREMENTS, ROADMAP, STATE, per-phase files. |
| **Requirement (REQ-ID)** | A single v1 capability the roadmap must cover, e.g. `CLI-01`. |
| **Success criteria** | Observable behaviors a phase must deliver, checked in Verify. |
| **Wave** | A batch of independent task plans executed in parallel. |
---
## π Troubleshooting
| Symptom | Likely cause | Fix |
|---------|--------------|-----|
| A MSD command isn't recognized | Claude Code not restarted after install | Restart Claude Code so it loads the new `/msd-*` commands. |
| `Spawning researchersβ¦` looks stuck | Research runs 1β5 min | Wait β don't interrupt. If truly hung, `/clear` and re-run the step. |
| Verify keeps failing | Real bug in the code | Let MSD write the fix plan β `/msd-execute-phase 1` β re-verify. |
| Lost track of where you are | β | Open `.planning/STATE.md`, or run `/msd-progress --next`. |
| Wrong install directory | Alternate/prerelease Claude Code config dir | Set `CLAUDE_CONFIG_DIR` to match β see [Install on your runtime](../how-to/install-on-your-runtime.md). |
---
## π What next
- [Install on your runtime](../how-to/install-on-your-runtime.md) β exact steps for every supported runtime
- [The phase loop](../explanation/the-phase-loop.md) β why it's shaped this way
- [Context engineering](../explanation/context-engineering.md) β the theory behind fresh sub-agents
- [Configure model profiles](../how-to/configure-model-profiles.md) β quality / balanced / budget tiers
- [Onboarding an existing codebase](onboarding-an-existing-codebase.md) β bring MSD to a brownfield repo
> [!TIP]
> **You now know the whole loop.** Everything else in MSD is a refinement of these
> eight steps. Welcome aboard. π