Tal Raviv

Overview

Tal Raviv is an AI experimenter and builder whose work sits at the intersection of coding agents, product workflows, and real-world AI automation. Across the newsletter mentions, he appears as someone who both prototypes agentic systems and publishes practical artifacts—demos, frameworks, open-source tools, transcripts, and filesystem snapshots—that make those systems easier for others to study and reproduce.

For AI Product Managers, Raviv matters because his projects are unusually concrete. He has shown support agents using tool calls for real business actions, argued that Claude can automate core PM workflows, proposed reusable “LLM primitives” for building agents, launched Familiar for ambient work-context capture, and shared a long-running Claude Code hardware experiment on a Raspberry Pi with sensors and actuators. Together, these examples give AI PMs a useful lens on workflow automation, context engineering, memory, evaluation design, and the limits of autonomous systems.

Key Developments

• 2026-03-18: Tal Raviv said he uses Anthropic’s Claude to automate core PM workflows such as drafting specs, prioritizing backlogs, and generating roadmaps, arguing that Claude now outperforms him on much of this work.
• 2026-04-02: Colin Matthews spotlighted Raviv’s demo of a support agent that used system prompts to call `get_order` and `issue_refund` through an application server, automating order lookups and refunds for lost orders.
• 2026-04-14: Raviv advocated for “context engineering as a team sport,” proposing that each teammate’s AI assistant should benefit from a shared knowledge base.
• 2026-04-16: He compared AI for PM work to the self-serve analytics wave led by tools like Mixpanel and Amplitude: empowering, but also prone to incorrect conclusions without expert judgment.
• 2026-04-28: Raviv launched Familiar, an open-source app that captures screen and clipboard data every 4 seconds as Markdown so local AI agents can use live work context.
• 2026-05-02: He outlined four “LLM primitives” for building agents: simple chat threads; chat plus tools; chat plus tools plus skills; and a file system layer (Memento) for persistent memory across sessions.
• 2026-08-15: Raviv described OpenAI’s “Computer History” as OpenAI’s version of Familiar, while emphasizing Familiar’s offline, local, open-source, free, and model-agnostic positioning.
• 2026-08-27: He shared an experiment testing whether Fable could grow food using Claude Code on a Raspberry Pi with a water pump, lamp, camera, sensors, and actuators, including a control group and mission prompt.
• 2026-10-07: Raviv recapped the 40-day “can Fable grow my food” experiment, saying he gave Claude Code three tools—a water pump, lamp, and camera—and used roughly 50 million tokens. He reported that he did not eat any radishes, then published a deep dive, open-sourced transcripts, and the Raspberry Pi filesystem for others to inspect.

Relevance to AI PMs

1. He offers practical agent design patterns. Raviv’s examples move beyond abstract AI talk into deployable structures: tool-calling support agents, layered agent architectures, persistent memory, and shared context systems. AI PMs can use these patterns when scoping internal copilots or customer-facing agents.

2. He shows how context becomes product infrastructure. Familiar, shared knowledge bases, and Memento all point to the same PM lesson: model quality often depends less on prompting alone and more on how context is captured, structured, refreshed, and shared across sessions and teammates.

3. He models transparent experimentation with clear artifacts. The Fable hardware workflow is relevant not because it “worked perfectly,” but because it exposed setup details, token usage, evaluation framing, open questions, and raw artifacts. That is a useful template for AI PMs running capability tests, pilots, or postmortems.

Related

• Claude / Anthropic / Claude Code / Opus 4.6: Raviv’s experiments and workflow claims are closely tied to Anthropic models and coding-agent behavior, especially Claude used for PM work and Claude Code used in hardware automation.
• Familiar: Raviv launched this open-source context-capture app to give local AI agents live screen and clipboard context.
• Memento: Mentioned in Raviv’s framing of agent-building primitives as the file-system layer for persistent memory across sessions.
• System prompts / get_order / issue_refund: These connect to Raviv’s support-agent demo, a concrete example of tool use and application-server orchestration.
• Context engineering / shared knowledge base: Core to his argument that AI effectiveness should be treated as a collaborative team capability, not just an individual prompting skill.
• PM workflows / PM: Raviv is frequently cited in the context of using AI to automate specs, prioritization, and roadmapping.
• Mixpanel / Amplitude: These appear in his analogy about the benefits and pitfalls of self-serve AI for product teams.
• OpenAI / Computer History: Raviv compared OpenAI’s Computer History to Familiar, helping frame the emerging category of screen-history and work-context tools.
• Fable / Raspberry Pi / sensors and actuators: These are central to his long-running hardware experiment testing whether an AI system could manage a food-growing workflow.
• Colin Matthews, Marily Nika, Aman Khan, Ben Erez: Related people in the broader AI PM and agent tooling ecosystem referenced alongside Raviv in the source material.