S6.04 - Structured_prompting

S6.04 ? Structured prompting

flowchart LR
    A[Ad hoc prompting
Inconsistent framing
Hidden assumptions
Weak uncertainty disclosure] --> B[RAIDT
Run-level evidence framework]
    H[Healthcare] --> C
    I[Public services] --> C
    J[Law] --> C
    K[Cybersecurity] --> C
    L[Enterprise productivity] --> C
    B --> C[[Structured prompting
Standardised task framing
Output constraints
Required uncertainty statements]]
    C --> D[Run-level evidence pack]
    C --> E[RAIDT score profile]
    C --> M[Governance move
Evidence over assertion
Reviewability
Contestability
Audit readiness]
    D --> F[Reviewer reconstruction]
    E --> G[Organisational learning]
    M --> N[Policy-aligned use controls]

? Star S6 - Influence Methods as Governance Interventions

Star context: Positions prompting, RAG, PEFT/LoRA, RLHF/DPO and stacked influence as governance-shaping interventions around the run. In RAIDT, structured prompting matters because it makes one part of model influence more inspectable, comparable and reviewable without mistaking prompting for governance by itself.

Academic picture

Definition / background

Structured prompting is the use of a deliberately designed prompt format that requires specified inputs, constraints, response sections, uncertainty statements or decision rules rather than relying on an open-ended free-text instruction. In practice, it may include templates, fielded instructions, output schemas, mandatory caveats, refusal conditions, escalation triggers, formatting rules or task-specific checklists.

Conceptually, structured prompting sits between ordinary prompt writing and more formal workflow control. It does not alter model weights, and it is not equivalent to retrieval, fine-tuning or post-hoc auditing. Instead, it shapes how a run is framed before the model produces an answer. That makes it an influence method: it constrains and channels model behaviour through the design of the instruction itself.

In GenAI governance, structured prompting matters because many risks arise not only from the model but from the way a task is posed. A poorly framed prompt can omit context, suppress uncertainty, invite over-claiming, or fail to require the distinctions a reviewer later needs. A structured prompt can reduce those weaknesses by making assumptions explicit and outputs more comparable across similar runs.

Within RAIDT, the importance of structured prompting is not that it guarantees a good answer. Its importance is that a prompt structure can be recorded as run-level evidence, inspected alongside input context and output, and assessed for its governance effect on Responsibility, Auditability, Interpretability, Dependability and Traceability. Structured prompting therefore belongs inside RAIDT because it turns a commonly invisible design choice into something reviewable within the evidence pack and score profile.

Why this concept matters

Structured prompting solves a practical governance problem: many organisations use prompts as if they were informal operator preferences, even though those prompts materially shape outputs, risks and reviewer confidence. If prompt structure is left ungoverned, the same system may behave differently across teams and time without any clear record of why.

It also avoids a common confusion between output quality and governance quality. A polished answer is not necessarily a well-governed answer. Structured prompting matters because it can require the model to separate fact from inference, disclose uncertainty, follow domain boundaries, surface missing information and produce outputs in forms that can be checked more systematically.

If structured prompting is absent, organisations may struggle to distinguish operator craft from reproducible process. That weakens internal assurance, makes incident analysis harder and reduces the organisation's ability to defend why a particular run was acceptable. In RAIDT terms, this means weaker evidence packs, less stable scoring and poorer organisational learning from repeated use.

Key idea: Structured prompting matters in RAIDT because it converts prompt design from tacit practice into inspectable run-level governance evidence.

What this item controls

It controls how a task is framed before generation begins.
It controls whether required context, boundaries and role expectations are explicitly included.
It controls whether the model must provide uncertainty statements, caveats or escalation signals.
It controls output comparability across similar runs by standardising response structure.
It controls part of the interpretive path from input context to output, making reviewer reconstruction easier.
It controls one layer of influence without claiming to control data provenance, model internals or downstream human decisions.

Practical example / likely audience question

Audience question

Is structured prompting really a governance intervention, or is it just better prompt engineering?

Answer

The concern behind this question is that prompting is often treated as a craft skill for improving outputs, whereas governance is expected to involve policies, controls and evidence. The direct answer is that structured prompting becomes a governance intervention when it is used deliberately to constrain behaviour and when its design is captured as evidence at the level of the run.

For example, an ad hoc prompt might ask a model to summarise a complaint and suggest next actions. A structured prompt, by contrast, may require the model to separate verified facts from inferred concerns, list missing information, state confidence, avoid legal conclusions, and route high-risk cases for human escalation. The prompt is now doing governance work because it embeds reviewable constraints into the interaction.

RAIDT handles this better than a generic AI governance approach because RAIDT does not stop at saying that prompt standards are desirable. It asks whether the exact prompt structure used in a specific run was logged, whether the resulting output can be reviewed against that structure, and how the run should score across the five pillars in light of that evidence. That is the move from abstract good practice to operational governance.

Practical example in RAIDT terms

Consider a public service team using a GenAI assistant to draft first-pass summaries of housing support requests. The run-level issue is that different staff may phrase the task differently, leading the model to omit vulnerability indicators, overstate certainty or fail to separate citizen statements from model interpretation.

A structured prompt for this use case could require five sections: case summary, directly stated facts, inferred concerns, missing information, and escalation recommendation. It could also instruct the model not to make eligibility decisions and to flag any safeguarding or homelessness risk for immediate human review.

The evidence needed in RAIDT would include the prompt template version, the populated prompt used in the run, the source text supplied, the generated output, any human edits, and reviewer notes on whether the mandatory structure was followed. The most affected pillars would be Responsibility, Interpretability and Auditability, with Traceability strengthened if prompt versions are logged and Dependability improved when repeated runs remain comparable.

This improves governance readiness because a supervisor can inspect not only the output but the control logic that shaped it. The organisation is therefore better placed to justify the run, revise weak templates and show that safeguards were embedded before generation rather than added only after a problem emerged.

Detailed link to RAIDT

Structured prompting links to RAIDT in four ways.

First, it links to RAIDT's core idea by showing that governance should attach to the run as actually configured, not to generic claims about a model or policy.

Second, it links directly to run-level evidence because the structure of the prompt is part of the configured use of the system for a specific task, at a specific time, in a specific context.

Third, it links to both the evidence pack and the score profile because a reviewer can inspect whether the prompt imposed useful constraints, disclosed uncertainty requirements and supported a defensible output form across the five pillars.

Fourth, it links to reviewability, contestability, audit readiness and organisational learning because prompt templates can be compared, challenged, refined and governed over repeated runs.

Structured prompting -> Run-level evidence -> Evidence pack -> RAIDT score profile -> Governance readiness

In this chain, structured prompting is not the end point. It is the influence mechanism whose governance value becomes visible only when it is documented, reviewed and connected to RAIDT's broader evidential logic.

Link to the five RAIDT pillars

Responsibility

Structured prompting can require the model to stay within task boundaries, declare uncertainty and signal when human review is needed. This supports more responsible use by reducing the chance that the system presents speculative material as settled advice.