Skills, Custom GPTs, and Gems: Getting the Most Out of AI Agents
Every major AI platform now lets you customize how it works. Claude has Skills and Projects. ChatGPT has Custom GPTs and custom instructions. Gemini has Gems. The promise is the same across all three: teach the AI how you work, and it will stop giving you generic output. The reality is more complicated — and more time-consuming — than the promise suggests.
This is a guide to the customization layer available to everyday users of these platforms. No code required, no developer tools needed. Just the built-in features that Claude, ChatGPT, and Gemini provide for shaping how the AI handles your work.
What Each Platform Calls It
The terminology differs, but the core concept is the same: you write a set of instructions that tell the AI who it is, how to behave, and what kind of output to produce.
Claude uses Skills and Projects. A Skill is a set of step-by-step instructions — written in plain English — that tells Claude how to handle a specific type of task. You can create them through the Skill Creator in the sidebar, or simply write a markdown file (called a SKILL.md) that describes your workflow and drop it into your workspace. Projects let you bundle skills together with reference files so Claude has persistent context about your work. When you ask Claude to do something, it scans your available skills and automatically loads the right instructions.
ChatGPT uses Custom GPTs and Custom Instructions. A Custom GPT is a specialized version of ChatGPT that you configure through the "Create" interface — you define its role, upload knowledge files, and write instructions that govern how it responds. Custom Instructions are lighter-weight: global preferences (your role, your industry, your preferred output format) that apply across every conversation. In 2025, OpenAI introduced structured instruction templates using XML-style tags for identity, objectives, workflows, and constraints — all written in plain language, not code.
Gemini uses Gems and Global Custom Instructions. A Gem is a task-specific assistant you create through the Gems panel — you name it, describe its behavior, optionally upload files or connect a Google Drive folder, and save it for one-click access. Gemini will even help you write the instructions: type a rough description and click the rewrite button, and it expands your notes into a structured prompt. Global Custom Instructions work like ChatGPT's equivalent — persistent preferences about who you are and how you want responses formatted.
How to Set Them Up
The setup process across all three platforms follows the same basic pattern: define the role, describe the workflow, and provide reference material.
Define the Role
Start by telling the AI what it is. Not "you are a helpful assistant" — that is the default behavior and adds nothing. Be specific about the job. "You are a commercial real estate analyst who specializes in multifamily acquisitions. You produce deliverables for institutional investors." The more specific the role, the less you have to correct the output.
Describe the Workflow
This is where most people stop too early. A good set of instructions doesn't just describe the what — it describes the how. Instead of "create a lease abstract," you need to specify: which fields to extract, what order to present them in, how to handle missing data, what to flag for manual review, and what output format to use. Think of it as writing an SOP for a new hire — the kind of document you'd give someone on their first day so they can do the work without asking you twenty questions.
In Claude, this lives in the Skill's instruction section or a SKILL.md file. In ChatGPT, it goes in the "Instructions" field of the Custom GPT builder. In Gemini, it goes in the Gem's instructions panel.
Provide Reference Material
All three platforms let you upload files that the AI can reference. This is where you store your formatting templates, style guides, example deliverables, or any other material that defines what "correct" looks like for your firm.
Claude Projects and ChatGPT Custom GPTs both support uploading PDFs, spreadsheets, and documents directly. Gemini Gems can pull from uploaded files or connect to Google Drive. The key practice across all platforms is to include an index — a short document that lists every reference file and what it contains — so the AI knows where to look for specific information instead of searching blindly.
Where This Actually Helps
For repetitive, well-defined tasks, custom instructions make a real difference. If you write the same type of memo every week, a Custom GPT or Skill that knows your firm's structure, tone, and formatting conventions will get you to a usable first draft faster than starting from scratch each time. If you always want financial figures in a specific format, or you want the AI to always ask clarifying questions before producing output, global instructions save you from repeating yourself every session.
The value compounds when you share these configurations across a team. A Custom GPT that your entire group uses means everyone gets consistent output — the same column headers, the same formatting, the same approach to ambiguity.
Where It Breaks Down
The customization features on these platforms are good enough to help. They are not good enough to solve the underlying problem, which is that general-purpose AI does not understand your industry's deliverables natively. No matter how detailed your instructions are, you are still working around the fact that the model was not built for your specific job.
Instruction drift. The longer and more complex your instructions get, the less reliably the AI follows all of them. You add a rule about how to handle escalation clauses. Three weeks later, you realize it stopped following your formatting rules from section two because the instruction set is now 1,500 words long and the model is prioritizing the most recent additions. Every platform has this problem. The context window is large, but attention is not uniform.
Model updates break things. When Anthropic, OpenAI, or Google releases a new model version — which happens multiple times per year — the way the AI interprets your instructions can shift. A Custom GPT that produced clean sensitivity tables last month might start rounding differently or restructuring columns after an update. You don't get a changelog that says "your lease abstract skill will now behave differently." You find out when the output looks wrong.
Domain knowledge has limits. You can instruct the AI to extract fourteen fields from a lease. You cannot easily instruct it to understand that Amendment 3 superseded the escalation terms in Amendment 1, or that a "greater of" rent structure implies conditional logic that changes how you model the cash flows. These are judgment calls that require deep domain expertise — the kind that comes from doing the work, not from reading instructions about the work.
Scaling across a team is harder than it sounds. One person can maintain a well-tuned Custom GPT. A team of fifteen people using a dozen different skills across four workflow types is a different story. Someone has to own the instructions, update them when models change, onboard new team members, and troubleshoot when the output drifts. That maintenance work is not trivial, and it usually falls on whoever cares the most — which is not a sustainable staffing model.
The Real Cost Is Your Time
The customization features in Claude, ChatGPT, and Gemini give professionals more control than they had a year ago. That is genuinely useful. But the time you spend writing instructions, curating knowledge files, testing output after model updates, and maintaining consistency across your team is time you are not spending on the actual work. For an individual power user with a few well-defined tasks, the investment pays off. For a deal team that needs institutional-quality output across dozens of deliverable types, the customization treadmill becomes its own job.
This is the gap that purpose-built AI coworkers are designed to close. Tools like Lumetric take the raw power of frontier models and make them opinionated — pre-built for specific industries and deliverables so deal teams skip the configuration step entirely. AI that already understands rent roll structures, knows how to format a sensitivity table with 25bps increments, and produces IC memos that match institutional standards out of the box. Not a general agent you spend your evenings training. A coworker that was built for the job.