The true power of AI Agents in Shibumi isn’t just in what they do — it’s in how they’re guided. Prompts act as the instruction set for the Agent, shaping the quality, consistency, and usefulness of every output. A well-crafted prompt can transform raw solution data into clear narratives, actionable recommendations, and automated insights.
Crafting strong prompts isn’t about overloading the AI with information — it’s about intentional structure and clarity. By following a thoughtful process, you can guide the Agent to deliver outputs that are not only accurate, but aligned with your program’s operational needs, reporting expectations, and decision-making moments.
1. Start with the End in Mind
Before writing a single line of a prompt, it’s important to define what “good” looks like for the output. Ask:
- Who will read or use this output?
- What decisions or actions will it support?
- How should the output be structured — narrative, bullet points, risk statements, or recommendations?
This clarity helps focus your prompt on the desired outcome, rather than simply describing the input.
2. Anchor the AI in Real Context
Shibumi prompts are most effective when they reflect the structure of your solution. This means:
- Naming the level of work item (Program, Workstream, Initiative, Activity).
- Referencing relevant attributes (e.g., status, planned dates, progress, benefit values).
- Framing the scenario in plain language so the AI understands why the information matters.
This step ensures the AI is not operating in a vacuum — it’s responding to a defined business moment.
3. Use Injected Expressions Strategically
Injected expressions are how you bring live solution data into your prompt. They allow the AI to dynamically reference fields without manual updates. When used thoughtfully:
- The AI works from real-time context.
- Prompts remain reusable across similar work items.
- Outputs stay consistent even as underlying data changes.
Rather than adding every attribute available, focus on the ones that shape meaningful insight — e.g., Status, Forecast Dates, Risks, Resource Needs.
4. Define a Clear Output Structure
An ambiguous prompt yields an ambiguous response. A strong prompt clearly states:
- The format (e.g., 2–3 bullet points, a short paragraph, or a one-line recommendation + rationale).
- The tone (executive summary, operational detail, risk alert, etc.).
- What to include (e.g., accomplishments, blockers, recommendations) and what to leave out.
This structure doesn’t restrict the AI — it guides it to produce something consistent and useful every time.
5. Leverage Descendant Work Item Data When Appropriate
Not all insights live at a single level. Sometimes the most powerful outputs come from summarizing or analyzing data from descendant work items (e.g., Initiatives rolling up to Workstreams). This approach allows the AI to:
- Identify patterns and risks across multiple records.
- Aggregate narratives to reflect real program health.
- Deliver reportign that leadership can act on without navigating multiple layers.
6. Iterate, Test, and Refine
Good Prompting is rarely perfect the first time. It’s a process:
- Write → Test → Refine → Lock In.
- Review the AI’s output critically: Is it on-message, structured correctly, and business-relevant?
- Adjust the wording, expressions, or structure to close any gaps.
Over time, these prompts become reusable templates that reinforce consistency across teams and programs.
From Thoughtful Design to Advanced Prompts
Once this foundation is in place, it becomes far easier to design prompts that don’t just extract information — they generate insight.
In a later section, we’ll walk through an advanced prompt example, breaking down how each part of the prompt:
- Established context,
- Leverages attributes,
- Directs the AI toward a structured output, and
- Scales effectively across multiple work items.
This approach gives your AI Agents a clear “north star” to follow, ensuring every run delivers results you can trust and act on.