Build a Reverse Prompting System with Claude Code

The Problem

You open Telegram. You stare at the input field. You type "hey" and delete it. Your agent has your goals in MEMORY.md, knows your projects, and could tell you what deserves attention. But it just sits there, waiting for you to speak first.

What You're Building

An agent that speaks first. It reads your goals, checks your calendar, reviews recent conversations, and proactively suggests what you should work on next — delivered to your Telegram every morning.

Milestone 1: Define Your Goals and Priorities

Your agent needs to know what you're trying to achieve — not just who you are.

Update my MEMORY.md file to include a new section called "Current Goals". Add 3-5 goals with priority levels (high/medium/low), deadlines if they have them, and what "done" looks like for each one. Also add a "Current Projects" section with the status of each active project. Make it structured enough for the agent to parse but readable as plain markdown.

What Claude Code does: It adds structured goal and project data to your long-term memory. MEMORY.md isn't just for facts about you — it's the agent's understanding of what matters to you right now. Goals with priorities and definitions of "done" give the agent context for useful suggestions.

Try it: Ask your agent: "What are my top priorities right now?" It should list your goals in priority order from MEMORY.md.

Milestone 2: Connect to Your Calendar

Goals tell the agent what matters. Your calendar tells it what's happening today.

Create a skill that reads my Google Calendar events for today and tomorrow. Use a read-only connection — the agent should see my schedule but not modify it. Parse the events into a simple format: time, title, duration. Store today's schedule in a daily context file that the agent loads alongside MEMORY.md. If I don't have Google Calendar set up, create a local calendar.json file I can manually update.

What Claude Code does: It adds time-awareness to your agent. The calendar skill gives the agent a sense of "now" — it knows your meetings, free blocks, and time before your next commitment. Suggestions go from "work on Project X" to "you have 90 minutes before your 2pm call — finish the Project X draft."

Try it: Ask: "What does my day look like?" It should list events with times and free blocks.

Milestone 3: Build the "Suggest Next Task" Engine

Goals, projects, calendar — the agent has everything. Let's build the brain.

Create a "suggest next task" skill for my OpenClaw agent. When I ask "What should I work on?" or trigger it via command, the agent should: 1) Read my goals and priorities from MEMORY.md, 2) Check my calendar for available time blocks, 3) Review recent conversation history for context (what I worked on yesterday), 4) Suggest 1-3 specific tasks ranked by impact, with a reason for each and an estimated time. The suggestions should be concrete actions, not vague goals.

What Claude Code does: It creates a multi-source reasoning skill. The agent cross-references three data sources — priorities (what matters), calendar (what's possible), and history (what's in progress). "Finish the Q1 report draft (45 min) — you started yesterday, deadline is Friday" beats "work on something important."

Try it: Ask "What should I work on next?" — specific, actionable suggestions with time estimates and reasoning.

Milestone 4: Set Up Scheduled Check-Ins

Don't wait for yourself to ask. Have the agent reach out.

Set up a daily check-in using OpenClaw's cron system. Every weekday at 9:00 AM, the agent should automatically run the "suggest next task" skill and send me the results on Telegram. Use a cron expression for weekdays only. Also add an afternoon check-in at 2:00 PM that reviews what I've accomplished and suggests what to focus on for the rest of the day.

What Claude Code does: It creates recurring cron jobs that trigger your suggestion engine. The agent becomes proactive — it messages you at 9 AM with today's priorities and at 2 PM with a midday course correction. The cron expression 0 9 * * 1-5 fires weekdays only.

Try it: Wait for the next scheduled time or manually trigger. You should receive personalized suggestions on Telegram — without asking.

What You Built

Remember staring at the input field? Your agent now:

• Knows your goals, priorities, and definitions of "done"
• Sees your calendar and understands available time
• Cross-references priorities, schedule, and recent work
• Suggests specific, timed, actionable tasks
• Reaches out proactively every morning and afternoon

You flipped the prompt. The agent tells you what to do next.

Take It Further

• Add weekly reviews — every Friday, summarize goal progress and suggest next week's focus
• Track completion — log which suggestions you accepted and measure impact
• Connect to project tools — pull status from GitHub issues or Notion for better context