Own your AI: Guide to keeping your AI knowledge portable

Assume that every model you use today will eventually be replaced, repriced, or made redundant. Avoid dependencies that cannot be unwound and keep your institutional knowledge in your own systems. Treat every AI session as a temporary workspace you feed into and extract from.

Below are some guidelines I use to optimise the utility of AI in my day-to-day work without undermining the human element and what I think (hopefully!) makes me good at what I do.

1. Own your knowledge

  • Playbooks in Word or Markdown: Preferred positions, fallback clauses, red lines, escalation thresholds – all in .docx or .md on your / your firm’s own secure storage.
  • Modular clause libraries: Standard data privacy, indemnity, liability, IP, dispute resolution and governing law snippets as reusable files you paste into any model or agreement template.
  • Tone of voice / format: Remember that what makes you / your firm distinctive is also the tone of voice, format, style and length / brevity of what you write. These are important contextual inputs into any prompt you use. Keep them handy in a separate file.
  • Never rely on platform memory: "Custom GPTs", project memories, and saved chats are vendor-locked. If your vendor changes, your access changes too. Keeping your documents within your systems in .docx or .md files also enables you / your firm to develop and share your own unique prompts and prompting techniques.

2. Prompt portably

  • Reusable instruction blocks: Keep regular role prompts ("Act as a general counsel with a moderate bias towards accommodating most requested changes if they are reasonable", "Strict Redline Mode") in a local file. Paste into any model.
  • Simple XML tags work everywhere: Wrap context with <playbook>, <clause>, <task>. Every commercial LLM reads these identically.
  • Feed context fresh each session: Paste the relevant playbook at the start. Don’t assume the model remembers yesterday.

3. Trust but verify

  • Models are confidently wrong: They state an incorrect position with the same assurance as a correct one. You need ingrained prompting habits that catch this.
  • "Rate your confidence 1–10 and explain why": Forces the model to separate what it knows from what it’s inferring.
  • "What’s the strongest counterargument?": Surfaces weaknesses the first answer glossed over.
  • "Separate stated facts from your assumptions": Distinguishes what came from relevant source material / documents versus what the model decided to fill in itself.

4. Tame the output

  • Models over-generate by default: Models routinely undertake work or tasks you never asked them to do. Without constraints, you get 2,000 words of hedged analysis (and potentially some content on side quests) when you needed three bullet points.
  • Specify format and length upfront: "Answer in under 150 words. No headers, no preamble. Plain prose. Black 11-point Calibri font. Square-bracket language you are not certain about."
  • Tell it what to leave out: "Do not explain background law. No caveats I already know. Just the clause and the risk."
  • Use output templates: Paste a table or markdown structure and say: "Fill this in. Don’t change the format."

5. Match model to task

  • Complex reasoning (multi-clause risk, cross-referencing): pick a top-tier reasoning model.
  • Long-document ingestion (full agreements, data rooms): pick a large context window model.
  • Drafting and prose (redlines, memos, client text): pick a model known for natural legal writing.
  • Keep a benchmark test: Develop and maintain a set of documents (of various complexities) to test and re-test it on every model update. That way your standards stay fixed and you can objectively determine which model provides the most useful output.

6. Protect confidentiality

  • Not all AI tiers are equal: Free and personal-plan chat interfaces may use your inputs for model training. Enterprise and API tiers typically offer data isolation and no-training guarantees.
  • Check your firm’s approved tools and tier before pasting anything client-confidential.
  • When in doubt, anonymise first: Strip names, dates, and identifying details before you paste. Restore them in the final document, not in the AI session.
  • Consider privilege issues: Don’t forget that running a document through a model – especially if done by a non-lawyer – may result in any legal privilege attached to that document being compromised. So where appropriate lawyers need to remind non-legal colleagues not to run privileged legal advice / memos through models themselves.

‍

Finally – AI allows you to be creative in how you keep up to date. For example, rather than the incessant LinkedIn feed or avalanche of unsolicited emails, I have a regular prompt that I run weekly that produces a 3-4 page update on all materials things related to AI regulation, data privacy, legal tech, related M&A / investments and other areas I am interested in – all presented in a format that suits my own preferences.

In my experience I’ve realised that my brain also soaks things up best when I’m taking a morning walk with my dog, so I often ask an LLM to turn my multi-page updates (and other documents) into natural language audio files / podcasts that I can listen to while taking my morning walk.  

THE GOLDEN RULE: Session-based processing, local storage

Feed in your context and playbooks → get the output → save locally → close the session. This gives you a process that you can repeat with any model.

The switch test: could you move to a different model tomorrow and lose nothing? If not, the part you’d lose belongs in a local file.