Editor support

A VS Code extension that highlights what a prompt references, in TypeScript and in skill files.

The Foundry Prompts extension for VS Code colours each thing a prompt references: skills, agents, files, commands, templates, arguments, harness variables, and shell. It reads each harness’s own notation, so /code-review in a Claude Code prompt, $code-review in a Codex prompt, and /skill:code-review in a Pi prompt are all marked as the same skill. The rest of the prompt stays a plain string.

In TypeScript and JavaScript

The extension highlights prompts inside the prompts tags, as tagged templates and as plain-string calls:

prompts.claude`/code-review @src/auth`;
prompts.codex("use $review");

Any object can come before the dialect: mine.claude from assemble highlights the same way as prompts.claude. A ${} substitution inside a prompt is still ordinary TypeScript, and a prompt written inside a substitution is highlighted too. Other template literals, strings, and comments are left alone.

In a template literal, notation that only works in a template is written with backslash escapes: \${CLAUDE_SKILL_DIR} and !\`git status\`. The extension highlights them as written.

In skill, agent, and rule files

VS Code opens SKILL.md, .claude/agents/*.md, .claude/rules/**/*.md, and *.prompt.md files with their own language ids. The extension reads the body of each one as a Claude Code template, so arguments, ${CLAUDE_SKILL_DIR}, and inline shell are highlighted alongside skills, files, and agents.

Frontmatter, code spans, fenced code, and link targets stay plain. Ordinary markdown, such as a README, is not highlighted at all.

What each harness reads

A command only counts when it opens the message. Later in the text, /review is a skill in Claude Code and plain text in Codex. Codex escapes a literal dollar sign as $$.

Colours

The extension does not ship a theme. Each kind of reference gets a TextMate scope that stock themes already colour, and the extension’s tests check that every kind stands out from the text around it in Dark+, Light+, and the GitHub themes that VS Code’s 2026 themes take their palette from.

To restyle one kind, target its scope in editor.tokenColorCustomizations. “Developer: Inspect Editor Tokens and Scopes” shows the scope under the cursor.

Install

The extension is not on a marketplace yet. Build and install it from a Foundry checkout:

cd tooling/vscode-extension
bun run build            # writes dist/foundry-prompts.vsix
bun run editor:install   # installs it into VS Code, replacing any earlier copy

Reload open windows afterwards. To remove it, run code --uninstall-extension foundry.foundry-prompts.

Other VS Code builds take the same file through their own CLI, such as cursor --install-extension dist/foundry-prompts.vsix --force. An editor built on an older VS Code may open skill files as plain markdown; there, only the TypeScript half works.

Limits

• Highlighting only. A skill or agent looks the same whether or not it exists. Linting checks them, through the ESLint, Oxc, or Biome extension.
• Custom dialect keys. mine.house from assemble is not highlighted, because nothing in the source says which harness it is.
• Command files. VS Code opens .claude/commands/*.md as plain markdown, and highlighting there would colour every README.
• One line at a time. A grammar matches line by line, so in a few edge cases it reads a prompt differently from the parser. For example, a command on the line after the opening backtick is highlighted as a skill. The extension’s README lists each case.

Working on the extension

The extension lives in tooling/vscode-extension. Run the “Prompts extension” launch configuration (F5) from the repository root to open a window with only this extension loaded, on its sample files. src/grammar.ts builds the grammars; after changing it, run bun run grammar and reload that window.

The tests run every grammar through a TextMate engine and compare what it highlights with what the @foundry/prompts parser reads, for each harness.