AI can absorb the specifics of a client's situation—their constraints, goals, communication style—and adjust its outputs to match what they actually need rather than delivering generic templates. Each conversation becomes a learning moment where the AI refines its understanding, making proposals, emails, and strategies progressively more aligned with reality.
In-context learning refers to the ability of a language model to adapt its behavior based on examples or instructions provided within a single conversation—without retraining the model. It's learning that happens in the model's context window, not through training updates. If you show an AI an example of how you want something formatted, it learns that format and applies it to subsequent requests in the same conversation. This is different from few-shot prompting (which is a specific technique) in that in-context learning describes the broader capability.
For freelancers, in-context learning means you can iteratively refine AI output within a single conversation, teaching the AI your preferences as you go, without starting over each time.
Conversation flow:
The key is that the AI's understanding builds cumulatively within the conversation. You're not starting fresh each time; you're refining a shared understanding.
Proposal structure refinement: You might start with a generic proposal structure. After the first draft, you provide feedback: "I want to lead with a problem statement, not an introduction. Here's what that looks like..." For all subsequent proposals in that conversation, the AI uses this structure without you re-explaining.
Tone calibration: "This feels too corporate. I want friendly but professional. Here's an example of my voice..." The AI adjusts and maintains that tone for all follow-up work.
Client-specific adaptation: "This prospect values data. Adjust all claims to include supporting evidence or citations." The AI learns the client's preference and applies it consistently.
Pricing and scoping language: "When I propose timelines, I always explain why the timeline is realistic. Here's an example..." The AI now includes realistic timeline justifications in all follow-up proposals.
Annotated examples: Instead of just showing a good example, annotate it: "Notice how this opening names the specific problem the prospect mentioned. This middle section includes a relevant case study. This closing explains our next step." The AI learns not just the pattern but the reasoning.
Explicit feedback: Be specific about what was good or bad: "Good: specific client insight. Bad: too long, cut the methodology section." Specific feedback teaches better than vague praise.
Repetition within conversation: If you refine a proposal three times before it's right, stick with that conversation. Don't start a new one. The AI's understanding deepens through repetition, and context builds.
Constraint addition: "All proposals should be under 2 pages. Here's an example." Adding constraints teaches the AI to optimize within boundaries.
In-context learning is powerful but temporary. Once the conversation ends, a new conversation doesn't retain what was learned. You either need to re-teach in the new conversation, or use few-shot prompting to create an explicit template. For recurring tasks, save your refined prompts or use RAG systems to preserve learning across conversations.
Also, in-context learning works better for stylistic changes (tone, format, structure) than for factual expertise. You can teach the AI your voice, but you can't teach it domain knowledge it doesn't have. That requires either prompt instruction or providing documents to retrieve from.
In-context learning shines when combined with prompt chaining. You establish a learning curve (examples and feedback) in the first few exchanges, then use that learned understanding to power a multi-step workflow: research → analysis → draft → refinement. The AI applies all your learned preferences throughout the chain without re-specifying.
Try this: Open a conversation with an AI. Ask for a rough proposal draft for a prospect (any prospect). After you get it, provide feedback: identify one thing you want more of and one thing you want less of. Provide a short example of each. Ask for a revised draft. Notice how the AI incorporates your feedback. Repeat this 2–3 times. By iteration 3, the AI should be generating draft proposals that feel much closer to what you actually want. This convergence is in-context learning in action. Then save the final prompt as your template for next time.
Peri can explain this concept, give practical examples, help you decide whether it applies to your situation, or recommend a journey if appropriate.
Explore related journeys or tell Peri what you're working through.