May 19, 2026

When ChatGPT debuted in late 2022, you typed instructions into a text box and hoped the machine understood what you meant. The better the prompt, the better the result. Ask for a contract summary and you got one kind of answer. Ask for key dates, party names, unusual provisions, and possible risks, and you got something much more useful. Now AI tools can do more than answer one request at a time. Some help you handle repeat tasks the same way each time. Some help move work through a series of steps. Some can take a goal and do more of the work on their own. How and when should you use a prompt, versus a skill, an agent, or a workflow?

When to Use a Prompt

A prompt is simply a one-time instruction you give an AI system for a specific task. You can think of it as giving paralegal quick verbal instructions for a one-off assignment. You explain what you need, the AI handles the request, and you get a result. Then the interaction is over. Prompts work especially well when you are exploring, testing ideas, or trying to figure out the right approach. If a client brings you an unusual question outside your usual practice area, prompting can help you do quick research and early analysis. Prompts are also useful for brainstorming arguments, summarizing transcripts, drafting a first pass at a client email, or turning rough notes into a cleaner memo. Their biggest strength is flexibility.

You might ask one tool to summarize a deposition transcript, another to draft interrogatories, and another to help you compare a line of recent cases. Each platform accepts natural language instructions without much setup. In ChatGPT, that usually means typing directly into the main chat box. In Claude, it is the same basic experience in a chat or inside a Project that holds your source materials. In Gemini, you prompt in the main Gemini chat and can use Deep Research when you want Gemini to gather and organize information for you. In Microsoft Copilot, you prompt inside the app where you are already working, such as Word, Outlook, Teams, or Copilot Chat. For lawyers, that makes prompts easy to start using because they do not require technical skills beyond clear thinking, careful writing, and enough professional judgment to review the output critically before it reaches a client, court, or opposing counsel.

Still, prompts have limits. Every time you need the same kind of analysis, you have to rebuild your instructions. If you review vendor contracts, employment agreements, or settlement terms every week, you will probably find yourself typing versions of the same prompt again and again. That repetition can lead to inconsistent results because even small wording changes can change the answer. It also means you may keep rebuilding the same prompt instead of saving it in a way your firm can reuse. In practice, that means valuable legal know-how can disappear at the end of a conversation instead of becoming something your firm can use again. While you can find your previous prompts in your chat history or save them somewhere, if you need to use a prompt repeatedly there are better ways to do that.

When to Use a Skill

Skills are the next step beyond prompts. A skill is simply a way to save and reuse a good AI instruction for the same kind of work. Instead of rewriting the same prompt every time, you set it up once and use it when you need it. In a law office, that might mean contract review, client intake, document sorting, or routine drafting. In ChatGPT, you are most likely to find this kind of reusable setup in a custom GPT built for a specific task. In Claude, you may see it in a Project with saved instructions and source files or, in more advanced uses, in Claude Skills. In Gemini, a Gem plays a similar role by giving you a reusable assistant for a recurring task. In Microsoft Copilot, you may see it as a built-in capability inside Word, Outlook, or Teams, or as a custom agent or tool created in Copilot Studio. For example, you might create one skill that always reviews a commercial lease for assignment, maintenance, insurance, and renewal terms. You might create another that turns a client intake note into a clean follow-up email. A skill helps you get a more consistent result without starting from scratch each time.

Take contract review. With prompts, you would write out clause-by-clause instructions each time you wanted to analyze an agreement. With skills, you set the review instructions once and let the system flag problematic clauses, identify missing provisions, and check for compliance with your usual terms or your client’s requirements. Tools like Thomson Reuters CoCounsel and Harvey AI already offer built-in skills for legal work, including deposition preparation, timeline generation, and privilege log creation.

From the user side, what matters is that skills can make AI faster, more consistent, and more useful for repeat work. You do not need to know how the system decides which instructions to use behind the scenes. You just need to know that a skill is often a better fit than a one-off prompt when you want the same kind of task handled the same way each time. For example, you might build a custom GPT in ChatGPT for first-pass contract review, a Gem in Gemini for client email drafting, or a Claude Project that always uses your preferred writing style and source documents. In Microsoft Copilot, you might rely on a built-in drafting or summarizing feature, or use a custom agent your organization created for a repeat task. If you regularly need to pull deadlines from scheduling orders, identify missing signature blocks in agreements, or sort incoming documents by matter type, a skill can do that more reliably than starting over with a fresh prompt every time.

What skills give you over prompts is consistency and efficiency. Once you set them up, they produce results in the same way each time without making you worry that a slightly different prompt will change the outcome. A skill becomes an asset your firm can keep using and improving. A prompt is something you spend and then have to recreate.

When to Use an Agent

Agents can do more on their own than prompts or skills. Instead of waiting for you to direct each step, an agent can take a goal, work through several steps, and keep going until it reaches a result. That means it can feel less like a tool and more like an assistant handling a task. For example, instead of asking one question after another about a deposition, you might tell it to help you prepare for the witness and let it gather the facts, organize the timeline, and pull out possible exhibits.

The key difference is that you set the goal and the agent figures out how to get there. If you ask it to help you prepare for a deposition, it can decide which documents to review, what timeline to build, which exhibits to flag, and how to organize impeachment material. For a busy lawyer, that can feel like supervising a very fast junior assistant.

You are starting to see agent features appear in major AI products. In ChatGPT, workspace agents can handle longer, shared tasks for teams. In Claude, agent-like features appear in tools such as Claude Code, Computer Use, and other task-running experiences. In Gemini, Deep Research is one of the clearest examples because you give it a research goal, review the plan, and let it gather and organize information for you. In Microsoft Copilot, agents often appear through Copilot Studio or Microsoft 365 agents that can work across your files, meetings, and apps. The names vary, but the user experience is similar. You give the system a goal, and it does more of the work without you having to spell out every step. That can be helpful, but it also means you need to watch the results carefully.

For legal work, agents are especially useful when the task is complex and you may not know every step in advance. Discovery management is a good example. An agent could review document productions, spot gaps, draft supplemental requests, research similar cases, and flag possible destruction or loss of evidence. You could imagine similar support in document review during a deal, internal investigations, or large-scale policy reviews where the work involves many moving parts and constant course correction. Because an agent can adapt as it goes, it is less likely to stall when something unexpected comes up.

Agents also raise new concerns. If you give them access to file systems, places where code runs, and system logins or connections, you also create security risks. For lawyers, that raises familiar issues around confidentiality, access control, supervision, and the need to verify work before relying on it. That means you need to review agent sources carefully and put strong limits and controls in place.

When to Use a Workflow

Workflows are useful when you want AI to help move work through a set process. Think of a workflow as a series of steps that happen in the same order each time. For lawyers, workflows matter most in routine office processes where consistency matters more than flexibility. Opening a new matter is a simple example. The facts may change, but the basic steps usually do not.

Workflows look a lot like the multi-step processes you already use, with AI handling the repetitive parts. A new client intake workflow might check conflicts against your client database, generate an engagement letter from your firm templates, create matter folders in your document management system, and fill in calendar deadlines based on court or state rules. The whole process can start with a single intake form. In Microsoft Copilot, workflows are most likely to show up in Copilot Studio or in connected Microsoft 365 tools that move work across Outlook, Teams, Word, and other apps. In ChatGPT, workspace agents can now support shared team workflows that keep running across steps and approvals. In Gemini, Deep Research can feel workflow-like when it plans, researches, and delivers a report in stages, while Gems can support repeatable parts of a process. Claude users may encounter workflow-like experiences inside Projects, Claude Code, or team setups that combine saved context with repeated steps. That kind of automation can cut administrative work and make it easier for lawyers and staff to follow the same process every time.

Litigation workflows can be especially powerful. When a case is filed, a workflow could pull key dates from the complaint, generate a response timeline based on the rules, build a case chronology from the pleadings, identify relevant prior cases, and assign tasks to your team. In a busy litigation practice, that kind of structure can help you avoid missed steps and keep the matter moving. E-discovery platforms like Relativity already use workflow automation to move documents through review stages, apply technology that helps sort documents for review, and route flagged items to senior reviewers.

The easiest way to think about the difference is this. A skill helps with one repeat task. A workflow helps move a matter or office process through several repeat steps. If you want better results on one kind of task, a skill may be enough. If you want to connect several steps into one smoother process, a workflow may be the better fit. For example, a skill might review one contract for missing clauses. A workflow might take that contract from draft to review to approval to signature.

Workflows work best for procedures that happen the same way each time and need careful review and tight control over how information moves through the process. If you are handling client intake, compliance reviews, matter management, or other recurring law office processes, workflows can be a very good fit. The tradeoff is flexibility. Unlike agents, which can adjust when something unexpected happens, workflows stay on their set path even when a different approach might make more sense.

Legal-Specific Examples in Major AI Tools

Some of the foundational AI vendors are starting to move in different ways on legal AI. Claude appears to be building the most direct legal package so far, with legal software connectors, practice-area plugins, and support for work such as drafting, redlining, clause comparison, and contract triage. Microsoft is taking a more focused path with Legal Agent for Word, which is aimed at contract review inside the document itself, including risk spotting, clause comparison, and suggested edits. OpenAI does not appear to have a comparable legal-specific product available today, but reports suggest it may be exploring one. Google, for now, is framing Gemini as a tool for legal work inside Workspace through drafting, summarizing, document review, and meeting support, rather than as a dedicated legal agent.

Conclusion

In most firms, the best place to start is still simple. Use prompts for one-off work, like summarizing one transcript or drafting one client email. Use skills for repeat work, like reviewing the same kind of contract every week. Use workflows for routine processes, like client intake or opening a new matter. Be cautious with agents, especially when client information, supervision, or accuracy are on the line. As you test tools, remember that the labels will vary. ChatGPT may call it a custom GPT or a workspace agent. Claude may use Projects or agent features. Gemini may use Gems or Deep Research. Microsoft may use Copilot features, agents, or Copilot Studio. The goal is to recognize what kind of help the tool is offering and decide whether it fits the work in front of you.