Anyone who has used AI tools such as Claude or ChatGPT has likely noticed a recurring frustration: each new conversation starts from zero. The user must re-explain who they are, what they do, who their audience is, and how they prefer information presented. For someone who regularly creates presentations, articles, or other recurring work product, this is an inefficient use of time.
There are now three ways to address this problem, each serving a somewhat different purpose: standing orders given within a conversation, memory that persists across conversations, and Projects. Understanding the difference helps users choose the right approach for the task at hand.
Standing Orders
The simplest approach is to give the AI explicit instructions at the start of a conversation, or at any point during one. These are sometimes called standing orders or custom instructions. Examples relevant to legal work might include:
- Only cite cases or statutes you are highly confident exist; if uncertain, say so rather than guessing
- Flag any claim that should be independently verified before I rely on it
- If you do not know the answer, say so, do not fill in gaps
- Use plain language and avoid unnecessary jargon
- After each response, list the sources you relied on
The value of instructions like these is that they push back against one of AI’s core tendencies: AI is trained to guess rather than acknowledge uncertainty. Explicit instructions that reward honesty over guessing help counteract that tendency. The limitation is that standing orders given in a single conversation disappear when that conversation ends. The next session starts from scratch.
Memory
Many AIs offer a memory feature. When I last wrote about standing orders, I noted that Claude did not have this feature, so it was necessary to paste the standing orders into each conversation. Fortunately, Claude now offers a memory feature that persists across conversations. Once enabled, Claude synthesizes context from past interactions, including professional role, current projects, and working preferences, and carries it forward into future sessions. A user who has told Claude they are a solo practitioner focusing on employment law, who prefers plain language and concise bullet points, should not need to repeat that information in future conversations. Memory is currently available on all Claude plans, including the free tier, and can be enabled in Settings. Users retain full control over what is stored and can edit or delete memory entries at any time.
You can provide standing orders to Claude by going to “settings” and “general”. The below image shows what the screen looks like and where to paste your instructions. If you aren’t sure what to paste, ask Claude or whatever AI you are using to help you create the instructions. Provide information about yourself and tell the AI to ask you questions until it gets the information it needs. Then paste the resulting orders into the proper location.
Memory is well suited for personal preferences and background context that apply broadly across all of a user’s work. It is less suited for task-specific instructions or reference materials tied to a particular project, which is where Projects come in.
Projects
In Claude, a Project is a dedicated workspace with its own standing instructions and uploaded reference materials that carry over across every conversation within it. Unlike memory, which applies globally, a Project is organized around a specific type of work. An attorney might create separate Projects for CLE presentations, client-facing materials, and legal research, each configured differently. For example, before Projects a single PowerPoint often required multiple conversations because the context window would fill up before the work was done. Project eliminates this problem.
Within a Project, the user sets custom instructions once. Those instructions apply automatically to every conversation that occurs within that project. Reference materials such as prior presentations, style guides, articles, and preferred clause structures are uploaded once and remain available without re-uploading. The standing orders described above can be embedded directly into the Project’s instructions, so they apply automatically without the user needing to repeat them each session.
Choosing the Right Approach
The three tools are complementary rather than mutually exclusive:
- Memory handles who you are and how you generally prefer to work, background that applies across everything
- Projects handle task-specific instructions and reference materials, the standing orders and documents relevant to a particular type of work
- Standing orders given in conversation handle one-off instructions specific to a single session
For someone who creates presentations regularly, the most efficient setup is a dedicated Project with standing instructions already embedded. Memory handles the broader context. For any single session where something specific is needed, such as a particular audience or an unusual format, additional instructions can still be given in the conversation itself.
Confidentiality in Claude and Other AIs
One important note for law firms: memory can be scoped to individual Projects, which means Claude retains distinct memories for each matter or workspace without those memories crossing over into other contexts. A firm can keep one client matter’s memory entirely separate from another’s. For attorneys with confidentiality obligations under Rule 1.6, this separation matters. That said, attorneys should still review their AI tool’s privacy settings and terms of service to confirm how memory data is stored and whether it is used for any purpose beyond the session. Turning memory off entirely remains an option for some AIs, and on Enterprise plans, administrators can disable the feature organization-wide.
A Note on Limitations: You Must Still Verify
None of these tools eliminate the need for human review. Memory and Project instructions improve consistency and reduce repetition, but the AI’s output still requires verification. This is especially true in legal work. An attorney using AI for research must still read the cases. An attorney using AI to draft documents must still review the substance. The efficiency gains are real, but they do not change the underlying professional obligations.