Skill Src

Decision tree navigator

When to activate

The user asks for help with a branching decision and three conditions hold.

Precedence over interview tools. Once activated, render the widget instead of using AskUserQuestion or any other interactive-question tool. The widget IS the interview - it shows all the branches at once and lets the user walk the tree, annotate, and commit. Falling back to one-question-at-a-time tools defeats the skill's purpose. This precedence holds even if a project's CLAUDE.md or global config tells you to prefer AskUserQuestion for interactive prompts; that guidance is for skills that do NOT provide their own UI.

Variance. Each top-level option leads to a different next-question, not the same question with a different filter. Architecture - "migration strategy" vs "shim location" vs "routing criteria" - render. Naming - "give me 5 punchy" vs "give me 5 descriptive" - do not render.

Bounded structure. The decision fits in 3 levels deep, 2-4 branches per level. Wider or deeper gets walked conversationally.

Sufficient context. The user has primed the conversation, or one clarifying turn fills the gap. Never render cold.

The variance check (CRITICAL, internal only)

Before rendering, run this check as silent internal reasoning - do NOT write it in chat:

1. Name 2-4 candidate L0 branches for the user's decision.
2. Write the L1 question that would follow each L0 branch.
3. Compare the L1 questions.
4. If two or more L1 questions collapse to the same question with different adjectives or filters, STOP and do not render. Offer a flat list response and say briefly why a tree doesn't fit.

Sharper test for the borderline trap: if the L0 branches are named things (Bali / Thailand / Vietnam; witty / professional / playful; Rust / Go / Python), the L1 questions are often "what should I do in X?" or "give me 5 examples in X" - same question, different filter. Those are skip cases even when 3 named L0 candidates exist. The L1 questions must explore different decision axes, not different instances of the same axis.

When the check passes, go straight to the widget. Do not sketch the directions in chat first, do not write "Different next-questions, so a tree works here", do not add context paragraphs, do not write a lead-in. The widget's own header already frames the decision; surrounding prose is noise. Reserve commentary for the final artifact once the user commits a leaf.

Generate the tree

Construct a JSON payload matching this schema:

{{SCHEMA}}

Sizing (HARD RULES):

• Exactly 3 levels deep: L0 → L1 → L2. Not 2, not 4. L2 nodes are terminal leaves.
• Every L0 branch MUST have a non-null sub with 2-4 L1 children. No L0 may be a leaf.
• Every L1 branch MUST have a non-null sub with 2-4 L2 children. No L1 may be a leaf.
• L2 leaves MUST have sub: null and next_hint: null. L2 is where the walk ends.
• 2-4 branches per level. Same flex at every level.
• No ragged trees. Every path from root to leaf must traverse exactly L0 → L1 → L2. If you find yourself wanting to make some L1s terminal and others have children, the tree shape is wrong for the topic.

L2 leaf quality (HARD RULES):

• Every L2 leaf summary must name a specific tool, pattern, place, or product. Not "a managed service", not "an appropriate database", not "your choice of provider". Examples of good L2 leaves: "Clerk hosted auth", "ClickHouse Cloud", "Margate, Kent", "Auth0 + WorkOS for SSO".
• Forbidden vague phrases anywhere in L2 summaries: "your choice", "appropriate", "depending on what you", "whichever makes sense", "as needed", "to be determined", "something that fits", "pick whichever". If any of these appear, rewrite the leaf as a specific named entity.
• For ideation: name the tool and sketch the MVP. For architecture: name the pattern and sketch its trade. For trip planning: name the place and the headline activity.

See references/REFERENCE.md for the full generation prompt template, token budget guidance, and edge cases. See references/EXAMPLES.md for fit and no-fit cases.

Rendering

Pipe the JSON payload through render.py via stdin (heredoc) - never via a temp file, never via shell argv:

python3 ${CLAUDE_SKILL_DIR}/scripts/render.py <<'EOF'
{"topic": "...", "submit_instruction": "...", "tree": {...}}
EOF

render.py validates the payload against the schema, derives <key>_json variants for every top-level key, and chevron-renders assets/widget-bundled.html. Capture stdout. Do not write the payload to a temp file - the heredoc above is the correct invocation. Writing to a file just adds a visible intermediate step in the host UX without any benefit.

Pass the script's stdout to visualize:show_widget as widget_code.

Call shape for visualize:show_widget:

• title: decision_tree_{topic-slug} (kebab-case)
• loading_messages: 3-4 short messages
• widget_code: the stdout from render.py

Zero prose before the widget. No lead-in, no variance sketch, no framing line, no "Rendering..." marker. The widget call should be the first user-visible output of your response. The widget's own header is the framing.

Handle the commit

When the user commits a leaf, the widget calls sendPrompt with a structured submission. The payload contains:

• The committed path - each level's question and the chosen branch
• Annotations - notes the user dropped on individual branches during the walk
• Abandoned siblings - branches the user explored but did not commit to

Use this payload as input. Produce the artifact named by the submit_instruction you passed in (an ADR draft, an idea brief, an itinerary, whatever the decision called for).

Annotations inform the final artifact. They do not reshape the tree mid-walk.

When NOT to render

• The variance check fails. Offer a flat list response and say why.
• The user wants conversation, not visual selection.
• The decision is too large to map in one shot, or each level depends on the user's previous commit in a way the pre-baked tree cannot capture.
• No bounded structure exists yet. Elicit constraints first, then re-evaluate.
• Single-shot questions: factual queries, plain recommendations, content the user wants in prose form.

The widget is a precision tool, not a default for any branching question.

References

• references/EXAMPLES.md - fit and no-fit scenarios with reasoning
• references/REFERENCE.md - JSON schema, generation prompt template, edge cases
• references/mustache-syntax.md - chevron Mustache subset for extending the widget template
• references/schema-authoring.md - JSON Schema constructs render.py's validator supports