From 281450174951103bb301f3343d7cf5873847ce5f Mon Sep 17 00:00:00 2001 From: Leon van Zantvoort <303196+zantvoort@users.noreply.github.com> Date: Mon, 5 Oct 2026 00:19:24 +0200 Subject: [PATCH] docs: bring the llms.txt and llms-full.txt headers in line with the CLI setup --- website/scripts/generate-llms-full.sh | 23 ++++++++++++++++------- website/static/llms.txt | 19 ++++++++++++------- 2 files changed, 28 insertions(+), 14 deletions(-) diff --git a/website/scripts/generate-llms-full.sh b/website/scripts/generate-llms-full.sh index 3a5c497b8..fe9ebefd3 100755 --- a/website/scripts/generate-llms-full.sh +++ b/website/scripts/generate-llms-full.sh @@ -147,14 +147,23 @@ cat > "$OUTPUT" <<'HEADER' > context, no transparent lazy loading, no proxy generation, and no entity state > management, so the code is exactly what runs. The domain model gives one-line, > type-safe queries across relations, returning the whole entity graph in one -> statement. The CLI installs rules and skills for coding -> agents, plus a local MCP server that exposes only schema metadata while -> keeping database credentials away from the LLM. Built-in verification -> (validateSchema(), SqlCapture) lets the agent check its own work before -> anything is committed. +> statement. +> +> For coding agents: the Storm CLI installs rules and skills for Claude Code, +> Cursor, GitHub Copilot, Windsurf and Codex. An existing rules file (such as +> CLAUDE.md) keeps its content; Storm adds its own marked block. Optionally it +> sets up an MCP server that runs locally and is read-only: it exposes schema +> metadata (table definitions, column types, constraints), never the database +> credentials, and row data only when the user enables it. Built-in +> verification (validateSchema(), SqlCapture) lets the agent check its own work +> before anything is committed. +> +> Get started (interactive, run by the user in a terminal; needs Node.js 18+): +> `npx @storm-orm/cli init` in an existing project, or `npx @storm-orm/cli demo` +> in an empty directory. Installed globally (`npm install -g @storm-orm/cli`), +> the same commands are `storm init` and `storm demo`; `storm update` refreshes +> rules and skills without prompts. > -> Get started: `npx @storm-orm/cli init` (existing project) or -> `npx @storm-orm/cli demo` (empty directory) > Website: https://orm.st > GitHub: https://github.com/storm-orm/storm-framework > License: Apache 2.0 diff --git a/website/static/llms.txt b/website/static/llms.txt index 2c2b949f9..b8ce113c3 100644 --- a/website/static/llms.txt +++ b/website/static/llms.txt @@ -12,14 +12,19 @@ > statement. > > For coding agents: the Storm CLI installs rules and skills for Claude Code, -> Cursor, GitHub Copilot, Windsurf and Codex, plus an optional local MCP server -> that exposes schema metadata (table definitions, column types, constraints) -> while keeping database credentials away from the LLM. Built-in verification -> (validateSchema(), SqlCapture) lets the agent check its own work before -> anything is committed. +> Cursor, GitHub Copilot, Windsurf and Codex. An existing rules file (such as +> CLAUDE.md) keeps its content; Storm adds its own marked block. Optionally it +> sets up an MCP server that runs locally and is read-only: it exposes schema +> metadata (table definitions, column types, constraints), never the database +> credentials, and row data only when the user enables it. Built-in +> verification (validateSchema(), SqlCapture) lets the agent check its own work +> before anything is committed. > -> Get started: `npx @storm-orm/cli init` (existing project) or -> `npx @storm-orm/cli demo` (empty directory) +> Get started (interactive, run by the user in a terminal; needs Node.js 18+): +> `npx @storm-orm/cli init` in an existing project, or `npx @storm-orm/cli demo` +> in an empty directory. Installed globally (`npm install -g @storm-orm/cli`), +> the same commands are `storm init` and `storm demo`; `storm update` refreshes +> rules and skills without prompts. Website: https://orm.st GitHub: https://github.com/storm-orm/storm-framework