What's Next

You have a personal AI familiar. It runs on your machine, reads your email and calendar, talks to you on Telegram, remembers your preferences, has a personality, survives reboots, logs everything, and acts on a schedule. Nine lessons. One system.

Now what?

Make it personal

The claw you built is a starting point. The whole point is that it's yours. Some directions worth exploring.

More data sources

The gatherer has two adapters: email and calendar. Add more.

• RSS feeds. Pull headlines from blogs you follow. The claw summarizes what's new.
• GitHub notifications. PRs that need review, issues assigned to you, CI failures.
• Weather. The claw mentions rain before your outdoor meeting.
• Slack. Unread DMs and mentions from channels you care about.
• Bank alerts. Transaction notifications parsed from email or API.

Each adapter is the same shape: a function that returns structured data. The gatherer doesn't care where it comes from.

More channels

Telegram is one channel. The Chat SDK supports others with the same handler.

• Discord. A private server where you talk to your claw. Useful if your team is on Discord.
• Slack. A DM bot in your workspace. The claw joins your work context.
• WhatsApp. Requires the Business API, but the adapter exists.
• Web. A simple HTML page with a text input and a WebSocket. The claw in a browser tab.

Adding a channel is one adapter, one environment variable, and a few lines in channel.mjs. The handler stays the same.

More skills

Skills scale linearly. Write a SKILL.md, drop it in skills/, and the claw picks it up.

• Todoist. The claw manages your task list.
• Obsidian. The claw reads and writes notes in your vault.
• Git. The claw checks repo status, creates branches, summarizes diffs.
• Home automation. The claw dims the lights or sets the thermostat via HomeKit CLI.
• Finance. The claw queries your budget app and flags overspending.

The pattern is always the same: find a CLI tool, write a skill that maps natural language to commands.

More gates

Script gates are the most extensible part of the system. Every new gate is a few lines of bash.

• Weekly summary. Fires on Monday morning. The claw compiles a digest of the past week.
• Git hygiene. Checks if you have uncommitted changes or branches behind remote.
• Disk space. Alerts when a drive is above 90%.
• Stale digest. Fires if the gather cycle hasn't run in 6+ hours.
• Meeting prep. 30 minutes before a calendar event, the claw pulls context on the attendees.

Every gate is free. Every gate runs in milliseconds. Write as many as you want.

Challenges

These are harder. They push the claw into territory the course didn't cover.

Multi-model routing table

Right now, model selection is manual: --model haiku in the gatherer, --model sonnet in the analyzer, default in the claw. Build a routing table that selects the model based on the task context. A models.json file that maps task types to models, with cost tracking per route.

Conversation summarization

Sessions grow. Context windows fill. Build a summarizer that runs when a session exceeds a token threshold. It compresses the conversation into a summary, starts a fresh session, and injects the summary as context. The operator doesn't notice the transition.

Multi-claw

Run two claws with different souls. One for work (sharp, efficient, no small talk). One for personal (warm, curious, asks follow-up questions). Different SOUL.md, different IDENTITY.md, different Telegram bots. Same infrastructure.

Voice

Add speech-to-text input via Whisper or a similar tool. Record a voice memo on your phone, send it to the claw as a voice message on Telegram, and get a text response. The Chat SDK handles voice messages as attachments.

Web dashboard

Build a simple web page that reads .claw/logs/claw.jsonl and renders charts. Cost over time. Interactions per day. Error rate. Model usage breakdown. The data is already there. It just needs a frontend.

Shared skills

Package your best skills into a repository other people can install. The wizard-skills repo is one example. Build your own collection. A skill for your favorite CLI tool. A skill for a workflow only you understand. Skills are markdown files. They're trivially shareable.

Self-improving CLAUDE.md

The memory analyzer curates MEMORY.md. Build an equivalent for CLAUDE.md. A background process that observes how the claw behaves, identifies instructions that aren't working, and proposes updates. The claw improves its own instruction set.

The real lesson

The nine lessons taught specific techniques. But the real lesson is the pattern: identify a gap, build the smallest thing that closes it, wire it into the system, and move on. The claw gets better one piece at a time.

Every production claw system works this way. Andy's system has 193 skills because he added them one at a time over months. joelclaw has 76. They didn't start that way. They grew.

Your claw will grow too. The substrate is set. Build what you need, when you need it.