| name | generate-behavior-questions |
| description | Generate frame-specific clarifying questions about ambiguous UI behaviors from a Jira epic and its linked Figma designs, Confluence pages, and Google Docs. Uses iterative content loading, parallel Figma frame analysis, and cross-content synthesis to produce targeted behavior questions organized by Figma frame. |
Generate Questions
Generate frame-specific clarifying questions for a feature by gathering all context (Jira, Confluence, Google Docs, Figma designs) and analyzing them comprehensively.
When to Use
Use when the user wants to:
- Review a feature's designs and generate clarifying questions
- Identify gaps, ambiguities, or contradictions in a feature spec
- Prepare for a feature review by listing what needs clarification
Typical trigger: "Generate questions for PROJ-123" or "What questions do we have about this feature?"
Required Input
- Jira issue key (e.g.,
PROJ-123) — the epic or story to analyze
Procedure
Phase 1: Fetch the Starting Issue
Call MCP tool extract-linked-resources with the Jira issue URL (e.g., https://myco.atlassian.net/browse/PROJ-123).
This returns the issue content as markdown with YAML frontmatter — save it directly to .temp/cascade/context/jira-PROJ-123.md.
The frontmatter contains discoveredLinks grouped by type (figma, confluence, jira, googleDocs) with relationship info (parent, blocks, etc.). Parse these to build your initial .temp/cascade/context/to-load.md.
If hasMoreComments: true, call again with commentsStartAt to get additional comment pages.
Phase 2: Iterative Content Loading
For each non-Figma URL in to-load.md (prioritize parent and blocks relationships first):
- Call
extract-linked-resources with the URL
- Save the returned markdown to
.temp/cascade/context/{type}-{identifier}.md
- Parse
discoveredLinks from frontmatter — add any new URLs to to-load.md
- Repeat until no unloaded non-Figma URLs remain
Important: Figma URLs are collected but NOT loaded here — they go in the ## Figma section of to-load.md for Phase 3.
Phase 3: Figma Batch Load
For each Figma URL collected:
- Call MCP tool
figma-batch-load with the Figma file URL
- This returns a download URL for a zip file
- Download and extract the zip:
curl -o .temp/cascade/figma/{fileKey}/batch.zip "{downloadUrl}"
cd .temp/cascade/figma/{fileKey} && unzip -o batch.zip
- The extracted zip contains:
.temp/cascade/figma/{fileKey}/
├── manifest.json ← frame list with metadata
├── prompts/
│ └── frame-analysis.md ← analysis prompt for frames
└── frames/
├── {nodeId}-{name}/
│ ├── image.png ← frame screenshot
│ ├── structure.xml ← semantic component tree
│ └── context.md ← annotations/connections
└── ...
Phase 4: Parallel Frame Analysis
For each frame listed in manifest.json:
- Launch a subagent using the
analyze-figma-frame sub-skill
- Pass only the frame directory path (e.g.,
.temp/cascade/figma/{fileKey}/frames/{dirName}/) — the subagent reads the files itself
- Subagents write
analysis.md to their frame directory
Do NOT read context.md, structure.xml, or image.png yourself. The subagent handles all file reading. You only pass the path.
Run all frame subagents in parallel — they are independent of each other.
Wait for all subagents to complete before proceeding.
Phase 5: Scope Analysis
Use sub-skill analyze-feature-scope to produce the scope analysis. This:
- Combines all frame analyses with epic context, reference docs, and per-frame annotations
- Categorizes every feature: ☐ In-Scope, ✅ Already Done, ⏬ Low Priority, ❌ Out-of-Scope, ❓ Questions, 💬 Answered
- Groups features by user workflow (not by screen)
- Returns a self-healing recommendation based on ❓ count
This produces .temp/cascade/scope-analysis.md.
Self-healing check: If the analyze-feature-scope sub-skill recommends CLARIFY (>5 unanswered ❓), present the scope analysis to the user and ask them to answer questions before generating behavior questions. Only proceed to Phase 6 if the user confirms or if ≤5 ❓ remain.
Phase 6: Generate Questions
Using the scope analysis and all frame analyses, generate frame-specific clarifying questions.
Critical Filtering Rules
- Cross-Screen Awareness: If ANY screen shows a behavior (component style, position, interaction pattern), that behavior is DEFINED — do NOT ask about it
- Scope Markers: Only ask about ☐ (in-scope) features. Skip ✅ (already done) and ❌ (out-of-scope)
- Context First: If any context source (Jira comments, Confluence, Google Docs, Figma annotations) answers a question, don't ask it
- No Duplicates: Don't repeat questions already present in Figma comments (check per-frame
context.md files)
- Check answered questions: If a question was previously asked (as a Figma comment) and has a reply, mark it 💬 — don't re-ask
Question Assignment Rules
- Every question must be assigned to the MOST RELEVANT Figma frame
- NO general/cross-cutting category — pick the closest screen
- Frames with no questions are omitted from output
Output Format
Present questions to the user in this format:
# Behavior Questions
## [Frame: Screen Name (nodeId: 123:456)](https://www.figma.com/design/{fileKey}?node-id=123-456)
1. ❓ What is the expected behavior when the user clicks the "Submit" button with invalid data?
2. ❓ Should the filter panel persist its state across page navigation?
## [Frame: Another Screen (nodeId: 789:012)](https://www.figma.com/design/{fileKey}?node-id=789-012)
1. ❓ Is the data table sortable by all columns or only specific ones?
Important formatting:
- Each frame heading is a markdown link to the Figma frame URL
- Node IDs use hyphens in URLs (
123:456 → node-id=123-456) but colons in heading text
- Only include frames that have questions
- Number questions within each frame section
Phase 7: Present Questions and Ask Next Action
Present the generated questions to the user.
After presenting, ask the user what they'd like to do next. Present these options clearly (use structured question UI if available — e.g., askQuestions in Copilot, AskUserQuestion in Claude Code):
- Post questions to Figma — Post as comments pinned to each frame (use
post-design-questions-to-figma sub-skill)
- Post questions to Jira — Post as a single structured comment on the issue (use
post-design-questions-to-jira sub-skill)
- Answer questions here, post Q&A to Figma — Walk through each question interactively, post answered Q&A pairs as Figma comments (use
answer-design-questions-post-to-figma sub-skill)
- Answer questions here, post Q&A to Jira — Walk through each question interactively, build a growing Q&A comment on Jira (use
answer-design-questions-post-to-jira sub-skill)
Do NOT offer "modify questions" as an option — if the user wants to modify, they'll say so naturally.
Important Notes
- Subagent depth: Frame analysis subagents are leaf tasks — they do NOT launch further subagents
- Cache efficiency: If
.temp/cascade/figma/{fileKey}/ already has valid data, figma-batch-load may return cached data — don't re-download unnecessarily
- Error resilience: If a single frame analysis fails, report the error and continue with remaining frames
- Fresh comments: To get the latest Figma comments (e.g., replies to previously posted questions), re-run
figma-batch-load — it always fetches fresh annotation data per frame
