agent-ux

name	agent-ux
description	How to use vscode/askQuestions and vscode/memory tools correctly. Load this skill whenever you need to ask the user structured questions or persist working state across turns.
user-invocable	false

Agent UX — askQuestions and memory

askQuestions

Use vscode/askQuestions any time you need structured input from the user before proceeding. It renders a compact form UI — do not ask questions in prose when this tool is available.

When to use it

• Before recording data that the user must confirm (project name, key decisions, stack choices)
• When presenting options for the user to choose from (Yes/No, A/B/C, multi-select)
• When you need a short freeform answer in a specific format (a date, a URL, a name)

When NOT to use it

• For simple acknowledgement ("shall I continue?") — offer a handoff button or ask inline
• When you already have enough information to proceed without ambiguity
• For long, open-ended narrative answers — ask in prose and allow the user to reply freely

Exception — recap + continue: if a "shall we move on?" decision is preceded by a chat-visible recap of what was just recorded (e.g. a list of outcomes added, features captured, metrics defined), then askQuestions is the right surface because the user is making a real confirm-or-correct decision, not a trivial acknowledgement. The recap makes the option label meaningful. This pattern is required by the initiate-project skill at every "add another / accept and continue" gate.

Constraints

Field	Limit	Notes
question	200 chars	One sentence. State the decision clearly.
header	50 chars	Short unique identifier — shown as the question label in the UI.
options[].label	~40 chars	Keep option labels brief and scannable.
Questions per call	3–4 max	Batch related questions; never dump a full form in one call.

Patterns

Binary choice with recommended default:

{
  "header": "Include auth",
  "question": "Should authentication be included in phase 1?",
  "options": [
    { "label": "Yes — include auth in phase 1", "recommended": true },
    { "label": "No — defer to phase 2" }
  ],
  "allowFreeformInput": false
}

Multi-select:

{
  "header": "Target platforms",
  "question": "Which platforms should be supported?",
  "options": [
    { "label": "Web" },
    { "label": "iOS" },
    { "label": "Android" }
  ],
  "multiSelect": true,
  "allowFreeformInput": false
}

Freeform short answer:

{
  "header": "Organisation name",
  "question": "What is the client's organisation name?"
}

---

memory

Use vscode/memory to persist working state that needs to survive across turns in this conversation.

Scope selection

Scope	Path prefix	Use for
Session	/memories/session/	In-progress notes for this conversation only. Cleared when the session ends.
Repository	/memories/repo/	Verified facts about this codebase. Survives across sessions in this workspace.
User	/memories/	Cross-workspace preferences. Rarely needed in TaskFlow agents.

What to write to repo scope

• Confirmed tech stack choices: "Uses Supabase with postgres schema in /supabase/migrations"
• Confirmed MCP servers added: "agno-docs MCP added — agents should use it for Agno API questions"
• Confirmed agent modifications: "Builder given agno/* tools — see .github/agents/builder.agent.md"

What to write to session scope

• Intermediate decisions mid-conversation: "User confirmed Xero integration but deferred Stripe"
• Partial answers while filling multi-step forms
• "Waiting for" state: "Asked about deadline — awaiting user response"

What NOT to write

• Credentials, API keys, passwords, or tokens
• Data that already lives in the TaskFlow database — use the DB as the source of truth
• Large blocks of text — keep notes concise (one line per fact)