Section 4 of 5 · 10 min read
System Instructions
The context layer you build every session can be made permanent. System instructions are the standing brief you give the AI — who you are, how you work, what you care about. Set them once; they load with every conversation.
The standing brief
Every major AI platform lets you set instructions at the account level — a standing brief that loads automatically into every conversation you start. Think of it as the “how to work with me” document you'd hand to a new colleague on their first day. Everything you'd otherwise retype at the start of every conversation belongs here: who you are professionally, your organization, your preferred communication style, recurring constraints, and anything else that consistently affects how you need AI to respond.
Most people skip this setup entirely. They rebuild context from scratch at the start of each session, spending two minutes typing background that the AI should already know. Over a year of daily use, that adds up to hours of wasted setup. More importantly, it means every AI interaction starts from zero familiarity — which limits what you can accomplish before the conversation context decays.
The flip side: global instructions affect every conversation. Keep them to information that's genuinely universal to your work. Information specific to a particular domain or project belongs at the project level, covered below.
Global instructions are the “how to work with me” document you'd hand to a new colleague. Write them once; every conversation starts with that foundation.
What belongs in global instructions
Here is a working template for a climate professional. Adapt it to your actual situation — the goal is specificity, not completeness.
Example global instructions
I'm a policy analyst at [organization], focused on [specific domain — e.g., energy access in sub-Saharan Africa / nature-based carbon markets / urban climate adaptation].
My audience for most work: [e.g., policymakers and program officers, not technical specialists]. Assume familiarity with climate concepts but not with technical jargon.
Communication style: direct and evidence-focused. I prefer shorter answers with clear structure over comprehensive answers with filler. Flag uncertainty explicitly — never paper over it with confident-sounding hedges.
I work primarily in English. Outputs should use American English conventions.
Never add unsolicited preambles (“Great question!”), summaries I didn't ask for, or suggestions to consult a professional at the end of responses.
Projects: the biggest context unlock
Global instructions handle your baseline. Projects handle everything domain-specific. A Project (called Gems in Gemini, Projects in Claude and ChatGPT) creates a dedicated workspace with its own instructions, memory, and files. This lets you maintain completely different contexts for different types of work without cross-contamination.
A climate professional might have separate Projects for: grant writing, policy analysis, stakeholder communications, research synthesis. Each Project has its own system instructions — defining the AI persona appropriate to that work — plus persistent files like organization background documents, style guides, template outputs, and reference materials. Every conversation you start inside a Project has access to all of that, automatically.
Project instructions override global instructions when they conflict. Within a Project, the AI is your grant writer, your policy analyst, or your communications specialist — not a generic assistant who happens to know your name.
Claude
—Projects
Strong context retention across conversations. System prompts support markdown structure well. Best for extended, document-heavy work.
ChatGPT
—Projects
Similar to Claude Projects. Integrates with memory system. Canvas mode useful for iterative document editing within a project.
Gemini
—Gems
Gems define a persona and behavior, linked to Google Workspace. Strong choice if your team is Google-integrated and you need shared AI workflows.
Writing effective project instructions
Project instructions are where you define the AI persona for this specific domain of work. Unlike global instructions (which are about you), project instructions are primarily about the role AI should play: what expertise it should bring, what constraints apply to this specific work, and what files in the project serve what purpose.
A concrete example for a grant writing project:
Example project instructions — grant writing
You are a senior grant writer specializing in climate and environmental programs. You understand program officer psychology: they read hundreds of proposals, they respond to specificity and evidence, they are skeptical of unsupported impact claims.
Files in this project: [org-profile.pdf] is our organization background; [previous-grants.pdf] are successful proposals for style reference; [theory-of-change.md] is our current program logic.
When drafting, always: lead with the problem before the solution; ground impact claims in specific evidence; use active voice; avoid mission-speak jargon.
When reviewing drafts I provide: identify the weakest logical link first, then the three strongest elements, then suggest the single most impactful revision.
Memory: what the AI learns on its own
Beyond explicit instructions, modern AI platforms maintain a memory layer — facts and preferences the model learns about you from conversations and stores for future sessions. This is distinct from conversation history: memory is a synthesized set of persistent facts, not a raw transcript.
Memory is valuable but requires maintenance. Periodically review what the AI has stored about you. Correct anything wrong; delete anything outdated. An AI working from incorrect memory compounds the error across every subsequent session. In Claude: Settings → Capabilities → Memory. In ChatGPT: Settings → Personalization → Memory. In Gemini: Settings → Personal Context.
The hierarchy across all platforms is the same: conversation-level instructions override project-level instructions, which override global-level instructions. More specific and more recent context takes precedence. If you tell the AI to be concise in your global settings but verbose in a specific conversation, the conversation wins. This hierarchy isn't absolute — it's probabilistic — but it's a reliable default.
A practical starting stack
The most important thing is not to get paralyzed by platform choice. The differences between Claude, ChatGPT, and Gemini matter less than how well you configure and use whichever one you choose. That said: pick one primary platform and go deep on its context management features before adding complexity. Enable memory. Set global instructions. Create your first project. Enable web search and code execution. Do those five things and you'll have more context leverage than the vast majority of professional AI users.
Add voice input (native or via a tool like Wispr Flow) when you want to think faster than you can type. Add a transcription tool for meetings. These compound over time. The goal isn't the perfect AI stack — it's a working one you actually use consistently.