James Cockburn
Menu
Home Services My journey LegalQuants Education & training Pro bono & access to justice LinkedIn archive How I think Frontier work Experience Projects Games Press Book an intro call

LinkedIn archive · 2025–2026

How the work developed.

I have been writing about this work as I go. These posts trace the move from experimenting with AI to building systems, teaching other lawyers and testing tools for use in practice.

Selected posts, earliest first. The dates are when I published them; the summaries describe the work at the time. Some of my views have changed as I have learned more.

  1. Experiments at home, questions for legal practice

    I was testing local language models and document retrieval in a home-hub project, while discussing what AI-assisted document review could mean for litigation. Sharing what we learned was already part of the work.

    Read the LinkedIn post : Legal AI tech: Information sharing VS secret advantage (opens in a new tab)

  2. From instructions to building

    After testing detailed instructions for legal tasks, I started learning about retrieval and agents, and trying to build the systems myself. This post records the trial and error at that stage.

    Read the LinkedIn post : The self-taught journey through AI (opens in a new tab)

  3. Assessing the promises

    Nine months into a closer investigation of AI, including conversations with technology providers, I wrote about the gap between product promises and what the tools could deliver for lawyers and their clients.

    Read the LinkedIn post : Informed scepticism - why AI is the future of professional services, why we must be wary of hype, and why we must jump in the deep end anyway. (opens in a new tab)

  4. What coding taught me about supervision

    Building revealed shortcuts I might have missed in a polished answer: an AI replacing supplied facts with dummy data, for example. I argued that understanding those choices mattered when using AI for legal work.

    Read the LinkedIn post : Why lawyers should try 'vibe coding' (opens in a new tab)

  5. A local retrieval system, and lessons to share

    After posting about my local retrieval system, I set out what I wanted to explain to other builders: the initial architecture, hardware, setup and the traps I had encountered.

    Read the LinkedIn post : I had a huge amount of interest, encouragement and feedback on my last post about building a 'no-code' (better phrase than vibe-code) local RAG system. (opens in a new tab)

  6. Teaching the first steps

    I began a practical series on AI-assisted coding, starting with how to scope an idea and turn it into working software. The aim was to help others get past the hurdles I had met.

    Read the LinkedIn post : How to 'Vibe Code' - Post 0 (opens in a new tab)

  7. Checking whether the output is right

    I wrote about the less visible work behind a convincing demonstration: reading outputs, creating benchmarks, testing edge cases and applying legal judgement. Producing an answer is only part of the job.

    Read the LinkedIn post : On Validation (opens in a new tab)

  8. Tools shaped by litigation practice

    I described pressuretest and timenarratives, built for the LegalQuants contribution to the Astra for Law launch, and the testing and peer review that changed them before release.

    Read the LinkedIn post : Astra for Legal (opens in a new tab)

The longer story, from 2010 →  ·  How I think now →  ·  Selected projects →