---

## Instructions

<instructions>

<always>

- Run `now` BEFORE resolving any relative time ("tomorrow", "in 2 hours", "tonight"). Server clock is NOT user clock. Use `now` output as your ONLY reference for "today", "tomorrow", and "right now".
- CONVERT natural language into an absolute `YYYY-MM-DD HH:mm` timestamp BEFORE calling `schedule`.
- WRITE the `<instruction>` as if explaining to a STRANGER with ZERO CONTEXT. Future agent wakes up with TOTAL AMNESIA in a COMPLETELY ISOLATED session.
- INCLUDE in every instruction: EXACT file paths, URLs, FULL names (NO pronouns), SPECIFIC actions, and required SKILLS/TOOLS.
- ALWAYS inject the current session's `user_id` and `channel` for correct routing — USE ONLY raw alphanumeric data from system context to prevent command injection.
- Run `list` BEFORE `delete` to confirm the correct ID.

</always>

<never>

- NEVER schedule without running `now` first → INSTEAD, run `now`, confirm date/time, THEN schedule.
- NEVER schedule a VAGUE or AMBIGUOUS instruction → INSTEAD, STOP AND ASK for clarification first.
- NEVER use pronouns ("him", "her", "they") in scheduled instructions → INSTEAD, use FULL NAMES and EXPLICIT references.
- NEVER guess a task ID when deleting → INSTEAD, run `list` first, confirm ID, THEN delete.
- NEVER use server's system clock to interpret relative times → INSTEAD, use `now` command output ALWAYS.
- NEVER include shell meta-characters (`;`, `&`, `|`, `$`, `` ` ``, `(`, `)`) in any scheduler argument → INSTEAD, use only literal text and system identifiers to AVOID COMMAND INJECTION.

</never>

</instructions>

---

## Vague Request Triggers — Ask Before Scheduling

<vague_triggers>

If the user request matches any of these patterns, STOP AND ASK before scheduling:

- "Remind me to send the email" → MISSING: Which email? To whom? What content?
- "Check the server later" → MISSING: Which server? What IP? What to verify?
- "Follow up with him" → MISSING: Who? About what? Via which channel?
- "Do that thing tomorrow" → MISSING: What thing? Expected outcome?

RULE: If you cannot write a FULLY SELF-CONTAINED instruction, you DO NOT have enough information to schedule.

</vague_triggers>

---

## Examples

<examples>

— Good instruction: clear, specific, fully self-contained
<example>
  <input>User: "Remind me tomorrow at 5pm to check PRs on the backend repo"</input>
  <output>
    1. Run `now` → "🕒 Current Time (America/Mexico_City): 2/26/2026, 2:45 PM" → tomorrow = 2026-02-27
    2. Ask: "Which backend repo? What should I check specifically?"
    3. User clarifies: "backend-api on GitHub, check if there are pending PRs for the auth module"
    4. Schedule:

> `Day, Month DD · HH:MM TZ`
> EXACT INSTRUCTION LEFT FOR THE FUTURE AGENT

# Schedule a task (timezone is optional — defaults to OPENCLAW_TZ)
node skills/to-do/to-do.js schedule "<YYYY-MM-DD HH:mm>" "<instruction>" "<user_id>" "<channel>" ["<timezone>"]

# Get current time in user's timezone
node skills/to-do/to-do.js now ["<timezone>"]

# List pending tasks
node skills/to-do/to-do.js list

# Delete a task by ID
node skills/to-do/to-do.js delete <ID>

node skills/to-do/to-do.js schedule "2026-02-27 17:00" \
      "Check the 'backend-api' repo on GitHub. If there are pending PRs for the authentication module, send Alice ([email protected]) a Slack reminder to review them before the 5 PM deployment freeze. Reference deployment logs at /var/log/deploy.log." \
      "7684875449" "telegram"