How Sandro Andric Built AgentHandover to Teach Agents Your Workflows
Built by sandroandric (GitHub) · · Source: GitHub
Result: According to the README, on a single 16 GB test Mac, Skill generation fell from 549 seconds to 60 seconds, and Memory search ranked the right screen first 90% of the time across 52 questions. The repo shows 703 stars.
What they built
Sandro Andric built AgentHandover, an open source macOS tool that watches how you work and turns repeated workflows into Skills that AI agents can run. According to the README, each Skill records the steps, strategy, selection criteria, decision logic, guardrails and your writing voice. Anything the system isn't sure about gets listed as an open question instead of a guess.
The Skills work with Claude Code, Codex, OpenClaw, Hermes and any MCP client such as Cursor or Windsurf. The repo is Apache-2.0 licensed and shows 703 stars.
How it works
The README describes a 12 stage pipeline that runs entirely on your Mac, with models served through Ollama. It says cloud mode was removed in version 0.5.0.
- Capture: A Rust daemon records the focused window and a Chrome extension adds page context. Privacy zones are checked before capture. Password managers, banking, payroll and sign-in pages are blocked by default.
- Understand: A small local model annotates what happened, mostly from text. The pipeline sorts work from noise, groups sessions and spots tasks that recur.
- Write the Skill: A local model drafts the Skill, then each step is checked against the recording. Steps it can't confirm become questions for you.
- Review: Nothing reaches an agent until you approve it. A Skill also needs no unknown required slots, a trust level above observe, and to be fresh. Freshness halves every 30 days since the workflow was last observed.
- Hand over: Claude Code gets Skills in its skills folder plus an MCP server, Codex gets a block in AGENTS.md, and OpenClaw and Hermes get Skills synced to their own folders. Agents report back after each run, which the README says improves Skills with use.
Version 2.0 added Memory, a private searchable record of what was on your screen, and Commonplace, one search across Skills, Memory, daily summaries and your profile. Memory is off for agents by default.
Setup needs macOS 14 or later and Ollama. You install the package or the Homebrew cask, record a task from the menu bar, answer up to three questions, approve the Skill, then connect your agent with one command. The README recommends a 16 GB Mac running Qwen 3.5 4B. It also notes that data is not encrypted at rest apart from Memory thumbnails, and recommends turning on FileVault.
The result
The numbers come from the author's own tests on a single 16 GB Mac.
- Skill generation dropped from 549 seconds to 60 seconds on the same kind of session.
- Time from recording to questions went from about 50 minutes with Gemma 4 12B to about 2.5 minutes with Qwen 3.5 4B.
- In Memory search, across 52 questions over 26 screens, the right screen ranked first 90% of the time and in the top five every time. The README notes this is a small sample.
How to build something similar
- Pick one task you repeat every week. Write the steps, the decisions you make along the way and the things that must never happen into a skill file your agent can read.
- Tell the agent to list what it doesn't know as questions for you, so it asks before it guesses.
- Keep a human approval step. No new skill should run until you have read it.
- Have the agent log what happened on each run, and use those logs to update the skill. Retire skills you haven't used in a while.
The original post
Want to build something like this?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week.
Join Shipping Skool — FreeMore use cases
- How Vinay Moved an AI Second Brain from Claude Code to Grok BotPersonal Productivity
- How Evan Kirstel Uses Claude Cowork as a Help Desk for Mac and iPhonePersonal Productivity
- How Khe Hy Built an Ambition-Upgrade Skill for Prompts and TasksPersonal Productivity
- How Sherwood Uses Hermes as a Personal Agent for Inbox TriagePersonal Productivity