Prompt Engineering

Prompts as Role Memory

In organisations where role holders are using AI tools effectively, the prompts they have developed are themselves a form of institutional knowledge. A delivery manager who has built a prompt that accurately summarises project risk from status updates, or a finance analyst who has calibrated a prompt to flag anomalies in spend data, has created role-level capability. When they leave, that capability leaves too — unless it is captured.

RolegacyAI treats prompt libraries as first-class role memories, capturing, versioning, and transferring them alongside decisions, lessons, and workarounds. But prompt engineering in RolegacyAI also has an internal dimension: the prompts used by the system itself to perform capture, classification, and brief generation are carefully engineered and versioned.

Capture Prompts

The capture pipeline uses prompts to extract institutional signal from unstructured content. These prompts are designed to identify the types of content that constitute role knowledge — decisions, lessons, workarounds, process patterns — without extracting personal, sensitive, or irrelevant material. Capture prompts are calibrated to handle the specific register and vocabulary of enterprise professional communication, which differs significantly from the content used to train general-purpose AI models.

Capture prompt performance is monitored through the Human Validation Loop: when a validator corrects or rejects an extraction, that signal feeds back into prompt refinement. Over time, the capture prompts improve their precision for the specific language patterns of the organisation and the role.

Classification Prompts

Classification uses a structured prompting approach that presents the memory candidate alongside the classification taxonomy and asks the model to assign types, domains, and a confidence assessment. The prompt is designed to encourage faithful classification rather than creative reinterpretation — the goal is to classify what the content actually says, not to infer what it might imply.

Successor Brief Generation Prompts

The Successor Brief Generator uses a multi-stage prompting approach. Retrieved memories are assembled into a structured context with source attribution. The generation prompt instructs the model to draw only on the provided context, to cite specific entries, to flag low-confidence or unvalidated content, and to organise the output according to the brief structure. The prompt includes explicit anti-hallucination instructions: the model is told to say "this information is not available in the current role memory" rather than generating a plausible but unsupported answer.

Role Holder Prompt Libraries

Beyond the internal prompts that power RolegacyAI's own pipeline, the system captures and stores the prompt libraries that role holders use in their own AI-assisted work. These are captured as AI Uplift memory entries: the tool used, the prompt pattern, the task it supports, and the calibration notes that make it work reliably for this specific role's context. When a successor takes the role, they inherit not just the history of what the role has done, but the AI capability it has developed.