Code Tour

Code Tour Skill

You are creating a CodeTour — a persona-targeted, step-by-step walkthrough of a codebase that links directly to files and line numbers. CodeTour files live in .tours/ and work with the VS Code CodeTour extension.

Two scripts are bundled in scripts/:

• scripts/validate_tour.py — run after writing any tour. Checks JSON validity, file/directory existence, line numbers within bounds, pattern matches, nextTour cross-references, and narrative arc. Run it: python ~/.agents/skills/code-tour/scripts/validate_tour.py .tours/<name>.tour --repo-root .
• scripts/generate_from_docs.py — when the user asks to generate from README/docs, run this first to extract a skeleton, then fill it in. Run it: python ~/.agents/skills/code-tour/scripts/generate_from_docs.py --persona new-joiner --output .tours/skeleton.tour

Two reference files are bundled:

• references/codetour-schema.json — the authoritative JSON schema. Read it to verify any field name or type. Every field you use must conform to it.
• references/examples.md — 8 real-world CodeTour tours from production repos with annotated techniques. Read it when you want to see how a specific feature (commands, selection, view, pattern, isPrimary, multi-tour series) is used in practice.

Real-world .tour files on GitHub

These are confirmed production .tour files. Fetch one when you need a working example of a specific step type, tour-level field, or narrative structure — don't write from memory when the real thing is one fetch away.

Find more with the GitHub code search: https://github.com/search?q=path%3A**%2F*.tour+&type=code

By step type / technique demonstrated

Raw content tip: Prefix raw.githubusercontent.com and drop /blob/ for raw JSON access.

A great tour is not just annotated files. It is a narrative — a story told to a specific person about what matters, why it matters, and what to do next. Your goal is to write the tour that the right person would wish existed when they first opened this repo.

CRITICAL: Only create .tour JSON files. Never create, modify, or scaffold any other files.

Step 1: Discover the repo

Before asking the user anything, explore the codebase:

• List the root directory, read the README, and check key config files (package.json, pyproject.toml, go.mod, Cargo.toml, composer.json, etc.)
• Identify the language(s), framework(s), and what the project does
• Map the folder structure 1–2 levels deep
• Find entry points: main files, index files, app bootstrapping
• Note which files actually exist — every path you write in the tour must be real

If the repo is sparse or empty, say so and work with what exists.

If the user says "generate from README" or "use the docs": run the skeleton generator first, then fill in every [TODO: ...] by reading the actual files:

python skills/code-tour/scripts/generate_from_docs.py \
  --persona new-joiner \
  --output .tours/skeleton.tour

Entry points by language/framework

Don't read everything — start here, then follow imports.

Repo type variants — adjust focus accordingly

The same persona asks for different things depending on what kind of repo this is:

For monorepos: identify the 2–3 packages most relevant to the persona's goal. Don't try to tour everything — open the tour with a step that explains how to navigate the workspace, then stay focused.

Large repo strategy

For repos with 100+ files: don't try to read everything.

1. Read entry points and the README first
2. Build a mental model of the top 5–7 modules
3. For the requested persona, identify the 2–3 modules that matter most and read those deeply
4. For modules you're not covering, mention them in the intro step as "out of scope for this tour"
5. Use directory steps for areas you mapped but didn't read — they orient without requiring full knowledge

A focused 10-step tour of the right files beats a scattered 25-step tour of everything.

Step 2: Read the intent — infer everything you can, ask only what you can't

One message from the user should be enough. Read their request and infer persona, depth, and focus before asking anything.

Intent map

Infer silently: persona, depth, focus area, whether to add uri/ref, isPrimary.

Ask only if you genuinely can't infer:

• "bug tour" but no bug described → ask for the bug description
• "feature tour" but no feature named → ask which feature
• "specific files" explicitly requested → honor them as required stops

Never ask about nextTour, commands, when, or stepMarker unless the user mentioned them.

PR tour recipe

For PR tours: set "ref" to the branch, open with a uri step for the PR, cover changed files first, then unchanged-but-critical files, close with a reviewer checklist.

User-provided customization — always honor these

Step 3: Read the actual files — no exceptions

Every file path and line number in the tour must be verified by reading the file. A tour pointing to the wrong file or a non-existent line is worse than no tour.

For every planned step:

1. Read the file
2. Find the exact line of the code you want to highlight
3. Understand it well enough to explain it to the target persona

If a user-requested file doesn't exist, say so — don't silently substitute another.

Step 4: Write the tour

Save to .tours/<persona>-<focus>.tour. Read references/codetour-schema.json for the authoritative field list. Every field you use must appear in that schema.

Tour root

{
  "$schema": "https://aka.ms/codetour-schema",
  "title": "Descriptive Title — Persona / Goal",
  "description": "One sentence: who this is for and what they'll understand after.",
  "ref": "main",
  "isPrimary": false,
  "nextTour": "Title of follow-up tour",
  "steps": []
}

Omit any field that doesn't apply to this tour.

when — conditional display. A JavaScript expression evaluated at runtime. Only show this tour if the condition is true. Useful for persona-specific auto-launching, or hiding advanced tours until a simpler one is complete.

{ "when": "workspaceFolders[0].name === 'api'" }

stepMarker — embed step anchors directly in source code comments. When set, CodeTour looks for // <stepMarker> comments in files and uses them as step positions instead of (or alongside) line numbers. Useful for tours on actively changing code where line numbers shift constantly. Example: set "stepMarker": "CT" and put // CT in the source file. Don't suggest this unless the user asks — it requires editing source files, which is unusual.

Step types — full reference

All step types: content (intro/closing, max 2), directory, file+line (workhorse), selection (code block), pattern (regex match), uri (external link), view (focus VS Code panel), commands (run VS Code commands).

Path rule: "file" and "directory" must be relative to repo root. No absolute paths, no leading ./.

When to use each step type

Step count calibration

Match steps to depth and persona. These are targets, not hard limits.

Scale with repo size too. A 3-file CLI doesn't get 15 steps. A 200-file monolith shouldn't be squeezed into 5.

Writing excellent descriptions — the SMIG formula

Every description should answer four questions in order. You don't need four paragraphs — but every description needs all four elements, even briefly.

S — Situation: What is the reader looking at? One sentence grounding them in context. M — Mechanism: How does this code work? What pattern, rule, or design is in play? I — Implication: Why does this matter for this persona's goal specifically? G — Gotcha: What would a smart person get wrong here? What's non-obvious, fragile, or surprising?

Descriptions should tell the reader something they couldn't learn by reading the file themselves. Name the pattern, explain the design decision, flag failure modes, and cross-reference related context.