next-prompt

name	next-prompt
description	Generate and execute the next agent prompt from connected personal data. Use when: the user says "what should I work on", "vana next", "what's next", or when the agent has no pending task. Also triggers after completing a task when autopilot is enabled.

Next Prompt

Pick the single most valuable thing the user should know or have done, based on their connected data and stated focus, and do it.

Config

Read ~/.vana/next-prompt.md for the user's guidance. This file is the user's standing directive — like a system prompt that steers what you focus on and how you behave. It is not a task list or backlog. Do not write tasks, action items, or deliverables into it.

If the file doesn't exist, create it with the default template below and tell the user to edit it — but don't block on that. Work with whatever you have.

# Next Prompt

## Focus

What I care about right now. Use this as context for deciding what's worth my attention — not as a checklist.

## Permissions

Do without asking:

- Research and analysis
- Writing summaries, memos, or drafts to local files
- Code exploration and reading

Ask first:

- Sending messages or emails to anyone
- Creating PRs, issues, or public posts
- Anything involving money, credentials, or shared infrastructure
- Deleting data or force-pushing

## Standing instructions

- (your preferences for how the agent works)

Flow

1. Gather context and check freshness

vana status --json

This tells you what sources are connected, their sync state, and lastCollectedAt timestamps.

Check freshness: if any source's lastCollectedAt is more than 24 hours old, suggest recollection before generating prompts:

• For one source: vana collect <source>
• For all stale sources: vana collect (re-collects all connected sources)

Only suggest recollection, don't block on it. Work with whatever data is available.

2. Read connected data

For each connected source, read the result file:

ls ~/.vana/results/

Read each JSON file. Look for timestamped data from the last 24 hours. Common timestamp fields:

• created_at, updated_at, timestamp, date
• create_time (ChatGPT conversations)
• startedAt, endedAt (Oura, activity data)

If timestamps aren't available, treat all data as current.

3. Read guidance

cat ~/.vana/next-prompt.md

4. Decide and act

Pick one action. The best action is:

• Time-sensitive over evergreen. A message aging toward an SLA beats a code cleanup.
• Aligned with the user's stated focus in the guidance file.
• Informed by what the data actually shows, not what you assume.
• Specific and completable in one agent turn.

Classify the action's risk:

Risk level	Examples	What to do
Safe	Research, analysis, local file drafts, code reading	Just do it. Tell the user what you did when done.
Visible	Sending messages, creating PRs, posting publicly	Describe exactly what you'd do in one sentence, then wait for a thumbs up.
Destructive	Deleting data, force-pushing, closing issues, spending money	Explain the action and consequences, then wait for explicit approval.

The user's Permissions section overrides these defaults.

Do not present a numbered list of options. Pick the best one. If you're wrong, the user will redirect you. That's faster than making them choose every time.

Reading sources

Each result file in ~/.vana/results/ is a different connected source. Don't assume a fixed set — new sources will appear over time. Read the JSON structure to understand what each source contains and look for what's actionable: recent activity, pending items, time-sensitive deadlines, unanswered messages, recurring patterns.

Rules

1. Never fabricate data. Only reference what's actually in the result files.
2. Respect Permissions from the guidance file.
3. If no data is connected, say so and suggest connecting sources. Do NOT run vana connect yourself — that is a separate skill (connect-data) and most sources require a headed browser you cannot access.
4. Work with whatever data IS available. Do not block on missing sources.
5. One action at a time. Do it well. Move on.
6. Don't repeat suggestions the user has already dismissed.