← Back to Guides · CLAUDE CODE
Best Claude Code Commands for Shipping a SaaS MVP in 2026
Short answer: a Claude Code custom command is a markdown file in .claude/commands/ — type / in a Claude Code session and the filename becomes the command name. Below are 12 commands worth having for a solo SaaS founder, grouped by job, each with what it actually does, plus a full example file for /ship and the mistakes that make a command catalog useless.
Claude Code's own docs call these "slash commands," and the on-disk format is identical to Cursor's .cursor/commands/ convention — same markdown-file-as-prompt model, different folder. If you're also running Cursor, the 20-command Cursor list covers the same ground for that folder; this one is the Claude Code-primary version with Claude Code's own frontmatter options (argument-hint, allowed-tools, model).
What a Claude Code custom command actually is
.claude/commands/
ship.md -> /ship
build-api.md -> /build-api
release.md -> /release
The file body is the prompt Claude runs when you type /name. Unlike a bare Cursor command, Claude Code commands support YAML frontmatter that constrains how the command runs:
---
argument-hint: [feature description]
allowed-tools: Read, Grep, Glob, Edit, Write, Bash(npm test:*)
model: opus
---
argument-hint— shown in the/autocomplete so you remember what to type after the command name.allowed-tools— scopes which tools the command can invoke without a permission prompt. A review command that only needsRead, Grep, Globshouldn't also carryEdit, Write.model— pins the command to a specific model (e.g.opusfor a multi-file feature build, a cheaper model for a mechanical pass) instead of inheriting whatever the session is set to.
None of this is required — a command with no frontmatter at all still works, the body is just the prompt. Frontmatter becomes worth adding once a command does something risky enough (writing files, running shell commands) that you want to scope it.
| Location | Scope | Use it for |
|---|---|---|
.claude/commands/ in the repo | This project, committed to git | Commands that reference your stack, folders, or conventions |
~/.claude/commands/ | Every project on your machine | Personal habits that don't depend on the repo |
Commit anything you depend on in CI or a headless Claude Code run to the repo's .claude/commands/ — a run outside your own machine won't see ~/.claude/commands/.
12 best commands for a solo SaaS founder
These are the AgenticKit catalog commands most reached for, each backed by a scoped agent pipeline rather than a single generic prompt.
Build
/ship — plan through a working, tested slice for one feature end to end: spec, schema migration, API, UI, tests, docs, in one turn.
/build-api — a single route with a typed contract and validation, when the data model already exists.
/build-page — a UI view against an API that's already built, instead of relitigating the backend.
/stripe-billing — wire up Stripe checkout, webhooks, and subscription state without duplicating logic Stripe already owns.
/auth-system — session handling, protected routes, and the auth tables, built to your stack's conventions instead of a generic tutorial's.
Verify
/audit — a pre-release sweep of a folder of routes: security, missing tests, dead code. Not a daily linter — run it before a release, not on every commit.
/qa-pass — exercise the actual user flows for a feature before it ships, not just the unit tests around it.
/e2e-suite — Playwright coverage for a path that touches money or auth, scoped to that path rather than every page in the app.
Operate
/release — a changelog generated from git history since the last tag, so shipping something doesn't mean no record of it exists.
/pr-description — a PR description written from the actual diff and commit messages, nothing invented.
Grow
/seo-fix — fix the specific on-page SEO issues an audit already found (title, canonical, headings, schema) instead of a vague "improve SEO" prompt.
/whats-new — turn a real shipped diff into changelog, email, and social copy, checked against the diff so it doesn't claim features that aren't there.
Full example: the /ship command file
This is the actual shape of .claude/commands/ship.md in the AgenticKit catalog — frontmatter plus a prompt that names the pipeline explicitly instead of leaving it to the model to improvise:
---
argument-hint: [feature description]
allowed-tools: Read, Grep, Glob, Skill, TodoWrite, Task, Edit, Write, Bash(npm test:*), Bash(npm run build:*), Bash(git status:*), Bash(git diff:*)
model: opus
---
Ship this feature end to end: $ARGUMENTS
1. tech-lead parses the request against CLAUDE.md conventions and writes a short technical spec.
2. backend-architect drafts any schema or migration changes needed.
3. api-engineer implements typed API endpoints with validation.
4. react-specialist builds the UI components and wires them to the API.
5. test-automator writes unit and e2e tests for the new path.
6. technical-writer updates docs and the changelog.
Do not skip the spec step even for a small feature. Do not touch files outside
the feature's scope. Stop and report back if a step reveals the request is
actually two features, not one.
The $ARGUMENTS placeholder is what you type after /ship — e.g. /ship add CSV export for the invoices table — and the "do not" lines exist because without them the model happily widens scope mid-task.
How commands differ from skills and subagents
A command only runs when you type it. A skill (SKILL.md) is loaded by the model on its own when its description matches what you're doing — no typing required, and it can bundle reference files a single command file can't. A subagent is a separate context the model delegates to with its own scoped tool access, usually invoked from inside a command or skill rather than directly by you.
The three-way distinction, and where hooks fit in as the fourth, event-triggered mechanism, is covered in more depth in Rules vs Skills vs Commands vs Hooks (written for Cursor's rule format but the commands/skills/hooks split applies identically to Claude Code). If the question is specifically about running agents in parallel versus delegating to one at a time, Claude Code Agent Teams vs. Subagents has the decision table for that.
Arguments and context: making a command take input
Most of the commands above take an argument — the thing after the command name. $ARGUMENTS in the file body is replaced with whatever you typed, so /build-api add a refund endpoint and /build-api add a webhook receiver run the same pipeline against two different specs without you editing the file. If a command needs more structure than one freeform string — say, a target file plus a description — reference named positional args ($1, $2) the same way a shell script would, and document the expected order in argument-hint so the autocomplete reminds you.
A command can also read files directly rather than relying on what you paste in: a line like "read the schema at prisma/schema.prisma before proposing a migration" inside the prompt body works, because the prompt is just instructions the model follows with its normal tool access — scoped by whatever allowed-tools permits.
Common mistakes: commands that should be skills instead
- A command you keep forgetting to type. If
/code-reviewonly runs when you remember to type it before merging, and you forget half the time, that's a sign the review should be a skill the model loads automatically on a matching task, not a command waiting for you. - One command, five jobs. A
/shipthat also runs the security audit, writes the changelog, and drafts a tweet will do all four worse than four separate commands would. Split by job:/ship,/audit,/release,/whats-new. - No "do not" line. "Write tests" invites the model to also refactor the function under test while it's in there. "Write tests, do not change production code" closes that door — every command above should have at least one explicit boundary like it.
- Grow commands with no brand voice doc behind them.
/whats-newand/seo-fixassume positioning and tone are already written down somewhere. If they aren't, the command either stalls asking, or invents a voice nobody approved. - Relying on
~/.claude/commands/for something CI needs. That folder lives on your machine only. A headless Claude Code run in CI won't see it — commit anything load-bearing to.claude/commands/in the repo.
FAQ
What is a Claude Code custom command? A markdown file in .claude/commands/ (or ~/.claude/commands/ for global) whose filename becomes the /name you type. The file body is the prompt; optional YAML frontmatter can set argument-hint, allowed-tools, and model.
Do Claude Code commands need frontmatter? No — a bare prompt file works. Frontmatter is worth adding once a command does something that should be scoped, like writing files or running shell commands.
How many commands should I start with? The 5-6 you'd actually type this week beat 12 installed and unused: /ship, /audit, /release, plus whichever of /build-api, /stripe-billing, or /auth-system matches what you're building right now.
What's the difference between a command and a skill in Claude Code? A command runs only when you type it. A skill is loaded by the model automatically when its description matches the task — no typing needed. If you keep forgetting to invoke something, make it a skill.
Can I use the same command file in Cursor and Claude Code? The format (frontmatter + markdown prompt) is similar but the folders differ: .claude/commands/ for Claude Code, .cursor/commands/ for Cursor. AgenticKit installs matching copies into both.
Get all of them in one install
AgenticKit installs the full command catalog — including every one named above — with the agents and skills behind them, into .claude/ and .cursor/. Browse the full commands catalog.
npx agentickit-cursor init --target claude
Written by Rakshit Yadav. Published 2026-10-31, last updated 2026-10-31.
Related Tools & Agents
Get the whole command set
The full catalog lands in .claude/ and .cursor/ with one install.
Browse commands →