Answer extracted from the Unprompted: Real AI in IC podcast — listen to the full episode below.
What prompt engineering techniques help ensure better AI agent outcomes?
Start by asking the AI tool to clarify your requirements before running the full prompt—Rebecca Lindsay emphasizes you don't know what you don't know. Instruct the AI to ask follow-up questions and remember them for future interactions, and be explicit about which knowledge sources the agent should draw from and what outcome you're trying to achieve.
Three foundational principles for stronger prompts
The first step is pre-clarification. Rather than launching straight into a complex request, ask the AI to help you define what you actually need. This upfront conversation prevents misalignment later and forces both you and the agent to name assumptions that might otherwise remain hidden. As Rebecca explains in the episode, many people rush into their request without first unpacking what they actually want to achieve.
Second, embed permission for the agent to ask clarifying questions within your initial prompt. This transforms the interaction from one-directional to conversational. By instructing the AI to probe deeper, you allow it to surface gaps in your brief and propose refinements. Rebecca describes this as giving the agent a voice to shape the outcome, not just execute a predetermined instruction.
Third, be explicit about scope and knowledge boundaries. Define which sources, tools, or data the agent should reference and which it should avoid. Clarity on what the agent is supposed to know—and what it should ignore—prevents hallucinations and ensures outputs stay grounded in your organizational context. This is particularly important at Hyper Exponential, where the tone of voice agent and annual leave assistant had to be trained to draw only from specific internal documentation and policies.
"We want to be giving you the tools to do great comms yourself, and that is precisely what AI can enable us to do at scale."
Rebecca Lindsay — Internal Comms Manager, Hyper Exponential. Rebecca has been part of the company's AI enablement squad and is known for building multiple AI agents to solve internal communications challenges, including tone of voice checkers, annual leave assistants, and emotionally intelligent communications tools. Her work is grounded in a company culture where the CEO describes himself as an AI maximalist and encourages all teams to experiment with AI tools in their roles.
To dive deeper into how these prompt principles translate into real agent deployment and the behavioral design choices that follow, listen to Rebecca's full conversation on the episode, where she also discusses how to identify which organizational problems are right for agent solutions in the first place.
Key takeaways
- Pre-clarify your requirements with the AI before running the full prompt; acknowledge gaps in your own thinking upfront.
- Build clarifying questions into your initial prompt instructions so the agent refines its understanding of your goal.
- Explicitly define which knowledge sources and tools the agent should draw from, and which it should exclude.
- This framework prevents misalignment and ensures AI agents produce outputs that fit your organizational context and objectives.
See also
What are the key steps for implementing and rolling out an AI agent in an organization?
The three steps are: first, identifying the need through data or observed friction; second, creating the agent itself; third, designing the behavioral rollout to ensure adoption and support throughout the organization.
How can internal communicators identify which problems are suitable for AI agent solutions?
Rebecca identifies needs through data and evidence, such as from internal comms surveys. For example, the annual leave assistant was created after survey feedback revealed a bottleneck in answering employee time-off questions.
Why does productivity with AI initially decrease before increasing, and how should communicators plan for this learning curve?
Productivity with AI goes down before it goes up because time must be invested in experimentation and learning how to work effectively with the tools. Teams should anticipate this dip and plan adoption support accordingly.