The lawyer who stopped re-explaining herself to ChatGPT

Picture a senior associate at a mid-size corporate law firm in London, sometime in late 2023. Every morning she opens ChatGPT, types a variant of the same paragraph: the firm's preferred contract language, the jurisdiction she works in, the fact that she reviews SaaS agreements and does not need American legal boilerplate. Every morning. For months. Then one day she just stops using ChatGPT at work entirely and switches back to keyword search and junior associate notes. Her reason, shared in a LinkedIn post that circulated widely enough to be picked up by legal tech commentators: "It kept forgetting who I was."

This story is not unique to her, and that is precisely the point.

What actually happened

OpenAI had built ChatGPT as a stateless system by design. Each conversation began from zero. The technical reasons were sensible: privacy, cost, the difficulty of managing long-running context at scale. But for professionals doing recurring, domain-specific work, this was a serious friction point. The lawyer's experience was replicated across finance teams, marketing departments, and consulting practices worldwide.

OpenAI's response came in stages. Custom Instructions launched in July 2023, letting users store a persistent system-level prompt that prepended every conversation. It was basic but genuinely useful: you could specify your role, preferred tone, output format, and constraints once, and the model would apply them across sessions. Then, in early 2024, OpenAI introduced Memory, a feature that allowed ChatGPT to store specific facts between conversations, either automatically extracted from dialogue or added manually. By mid-2025, Projects arrived in ChatGPT, allowing users to group conversations, upload persistent files, and apply custom instructions at the project level rather than globally.

Anthropic took a different architectural approach with Claude. Rather than building memory as a separate layer bolted on later, Claude's Projects feature (released in mid-2024) was designed from the start around the idea of a persistent workspace. You create a project, upload documents, write a system prompt for that project, and every conversation within it inherits that context automatically. A legal team could have one project for NDA reviews, another for M&A due diligence, with different instructions, tones, and reference documents in each.

The distinction between the two vendors matters in practice. ChatGPT's memory system, according to OpenAI's own documentation (a vendor source, worth checking against independent user research), stores discrete facts: "User works in contract law," "prefers British English," "does not want numbered lists." Claude's project-level context works differently: it holds full documents and a structured instruction set that functions more like a brief handed to a professional before a meeting. Neither approach is strictly superior. They suit different workflows, and professionals who have used both platforms seriously tend to develop a clear preference based on their specific task patterns.

Why it still matters

The lawyer's frustration points at something that took the AI industry a while to name clearly: the difference between a capable model and a useful tool for knowledge work. A model can be extraordinarily capable at reasoning, summarization, or drafting, and still be nearly useless in practice if it requires constant re-briefing. Cognitive overhead is a real cost.

The introduction of persistent context features changed the conversation inside organizations. Suddenly, AI tools could be configured to behave consistently with firm-specific terminology, house style, regulatory jurisdiction, or product constraints. A compliance officer at a European bank could set up a Claude project that always references MiFID II requirements. A product manager could give ChatGPT standing instructions to frame every output in terms of the OKR framework their company uses. These are not exotic use cases. They are the bread and butter of professional work.

There is a subtler point here too. When professionals control what context persists, they are, in effect, doing a form of prompt engineeringprompt engineeringPrompt engineering is the practice of designing and refining text inputs to guide large language models toward accurate, relevant, and reliable outputs.View full definition → at a higher level of abstraction. They are not crafting a clever one-shot prompt; they are designing a working environment. This shifts the skill set required. The question is no longer just "how do I phrase this request?" but "what does this tool need to know about my world to be consistently useful?"

Anthropic's design philosophy with Claude Projects reflected an explicit bet that professionals would respond well to a document-centric model of context. Upload your style guide, your product brief, your regulatory reference document, and the model draws on all of it. OpenAI's approach with ChatGPT has been more granular and user-driven, leaning on extracted facts and user-defined instructions. By mid-2026, both platforms have iterated significantly on these foundations, but the underlying design choices made in 2023 and 2024 still shape how each tool behaves.

The takeaway for you

The practical implication is simple enough: if you are using ChatGPT or Claude for recurring professional work and you have not spent thirty minutes configuring Projects, custom instructions, or memory settings, you are working harder than you need to. The setup is not complicated. Write a paragraph that describes your role, your constraints, your preferred output format, and anything the model consistently gets wrong about your context. Do it once per project type. The compound return over weeks of use is substantial.

There is a governance angle worth flagging for teams. Persistent context means persistent risk if the instructions are poorly written or outdated. A project configured with last year's regulatory requirements and never updated is worse than no project at all, because it creates false confidence. Treat your project instructions the way you would treat a shared team template: version it, review it periodically, and make sure someone owns it.

The lawyer in London eventually came back to AI tools, once Projects existed and she could brief the system properly before the first message of the day. Her public comment on the experience was brief and accurate: it felt like the difference between a new contractor and one who had read the file.