Develop with a coding agent
Install release-matched skills, give a coding agent scoped project context, and keep task execution explicit.
Start after purchase
Accept your GitHub repository invitation, then copy this prompt into a shell-capable coding agent in the workspace where you want to build. It checks access, chooses between a new Starter and an existing project, and uses the acquired release's references. You complete interactive sign-in yourself; never paste credentials into the conversation.
Set up Smoothie Framework for me using my purchased GitHub access.
Start by reading https://smoothie.dev/docs/llms.txt. Follow its installation, Framework acquisition, project creation, skills and MCP guides. Check the installed CLI's help before using commands; do not invent commands or flags.
First inspect this workspace and the installed tools. Ask what I want to build and where the application should live if that is not already clear. Preserve existing files, Git changes and agent configuration.
1. Verify access
Use the existing GitHub CLI authentication. Confirm it uses the personal account that received my Framework invitation, then verify access to SmoothieOrg/framework-standard. If authentication or invitation acceptance is needed, let me complete that step. Never request or display tokens. Installing the CLI does not grant repository access.
2. Install and acquire
Install the official Smoothie CLI if missing, following the documented procedure for this platform. Download a complete Framework release into a new directory, verify its source/package pair, and record the exact release. Do not substitute historical public npm packages or overwrite a download.
3. Set up the application
For a new application, inspect the available Starters and select the smallest suitable one. Preview creation, review its plan, then apply it using the returned plan ID.
For an existing application, follow the existing-project guide and preserve its architecture; do not overlay a Starter onto it. If the documented adoption flow does not support this project's package manager or workspace layout, explain the limitation before changing it.
Install dependencies with lifecycle scripts disabled unless a reviewed dependency specifically requires otherwise.
4. Configure agent assistance
Install the release-matched Framework skills for this agent host using the CLI's plan/apply flow.
Configure the documented project-scoped Smoothie MCP connection with absolute paths, preserving other servers. Verify that a context search and reference read actually work. If this host cannot connect immediately, report the required reconnect step and use CLI context meanwhile.
5. Verify and hand over
Run the application's documented typecheck, tests and build where supported. Distinguish passed checks from skipped or blocked checks. Report the project path, exact CLI/Framework versions, checks performed, MCP status and how to run the application. Do not claim an unsupported platform or unrun check passed.
You may perform the local installation, project setup and checks described above. Do not deploy, publish, change GitHub permissions, make purchases, or make billable model calls without asking.
While building, use the acquired release's local documentation and skills for API details. Use the online documentation index to find relevant guides, and report version differences rather than guessing.Start with the documentation index and read the relevant linked pages. Agents that need a complete public documentation snapshot can download llms-full.txt. It contains public guides and reference material, not the private Framework release. Once a release is installed, its retained references take precedence for that version's APIs.
Prerequisites
A coding agent helps develop your application; it is not an Agent running inside your product. First create an application or add packages to an existing project. Its retained release supplies version-matched Framework references and instructions. Generic local context and MCP tooling can also work without Framework access.
Install Framework skills
Follow Skills: installation and agent-host setup, then choose a concern and give the agent a concrete task with acceptance criteria. Installing instructions does not configure MCP, provide credentials, or authorize execution. The Skills section owns installation, discovery, task briefs, and version/conflict handling.
Search project context
A shell-capable agent can inspect local project evidence and read release-correct references through the CLI context interface. Ask it to preserve the source identity and report missing or incomplete evidence rather than substituting the latest online API. Native agent search remains useful for your application code; Smoothie does not index everything implicitly.
Connect through MCP (optional)
MCP exposes schema-defined calls to the agent host without requiring shell command construction. Follow Connect a coding agent for one scoped Codex/Claude setup. Start with context and inspection; tool execution requires separate explicit grants. Those grants constrain Smoothie, not the host's own shell or file access.
Verify the agent's work
Define permitted effects before work starts. Require changed files, actual check results, skipped prerequisites, and unresolved risks in the handoff. Use Testing and verification for the owning check procedure and CI and troubleshooting to interpret failures. Neither a discovered task nor a successful plan is passing execution evidence. Model spending and deployment need their own authority.