A suggestion is a soft constraint that nudges an LLM toward a better output without failing the pipeline. It is useful when the rule is important but not absolute, such as encouraging citations or improving format. Suggestions preserve flexibility while still giving the optimiser clear guidance.
What a suggestion does in an optimisation pipeline
A suggestion is not a hard rule, it is guidance that improves output quality without stopping generation. That distinction matters because the optimiser can follow the signal where it helps, while still producing a usable result when the suggestion is incomplete, ambiguous, or conflicting with stronger instructions.
In practice, suggestions sit between free-form prompting and enforceable policy. They are most useful for style, structure, tone, citation habits, or other quality preferences that should influence the result but should not cause a failure state.
One useful way to think about suggestions is as a soft form of control: they shape the model’s search space, but they do not become the answer’s acceptance criteria. That is why they can improve consistency without creating brittle workflows.
For teams building around non-human identities and other machine-driven workflows, NHIMG’s Ultimate Guide to NHIs is useful background when suggestion-based governance intersects with secret handling, rotation, and visibility.
How suggestions differ from rules and constraints
Rules define what must happen, while suggestions express what should happen when possible. A hard constraint can reject or block an output, but a suggestion simply nudges the system toward a preferred outcome.
This difference is important for pipeline design. If you treat a preference as a rule, you can create avoidable failures and unnecessary retries; if you treat a rule as a suggestion, you can allow low-quality or non-compliant output to pass through.
Suggestions are also useful when the target behaviour is beneficial but context-dependent, such as encouraging citations, improving formatting, or steering the model toward a safer default without over-constraining legitimate variation.
For practitioners comparing control styles, the NIST SP 800-53 Rev. 5 Security and Privacy Controls catalogue shows the value of distinguishing between control intent and implementation strictness.
Where suggestions are most useful
Suggestions are strongest in areas where quality matters, but exact enforcement is unnecessary or counterproductive. Common examples include writing style, document structure, citation preference, answer length, and formatting consistency.
They are also helpful in systems that combine multiple signals, because a suggestion can steer the optimiser without overriding the rest of the instruction stack. That makes them a practical choice when the desired behaviour is worth encouraging, but the system still needs flexibility to handle unusual inputs.
In security-adjacent workflows, this soft guidance can be especially valuable for prompts that should encourage traceability, completeness, or safer defaults without turning every deviation into a processing failure.
The OWASP API Security Top 10 is a useful adjacent reference when suggestions are used to shape behaviour around request handling, authorisation, and reliable output quality in API-driven systems.
Why suggestions matter operationally
Suggestions improve resilience by reducing unnecessary brittleness. A system that can distinguish between mandatory constraints and soft preferences is easier to tune, easier to debug, and less likely to fail for non-critical reasons.
They also make governance clearer. Teams can reserve hard enforcement for security, compliance, and correctness requirements, while using suggestions for editorial or optimisation preferences that should influence behaviour but not block delivery.
That separation helps keep pipelines predictable. It also makes it easier to explain why a result was accepted even when it only partially satisfied a preference, which is important when output quality is reviewed by humans downstream.
For implementation patterns that support this distinction, the OWASP Cheat Sheet Series offers practical guidance on building safer, more consistent application behaviour.
Risk and Threat Considerations
Suggestions are low-friction by design, but that flexibility can become a risk when a soft preference is mistaken for a control. If a team assumes a suggestion provides enforcement, sensitive content, weak formatting, or missing citations can slip through without any failure signal.
Failure mechanism: The system treats guidance as optional, so a weak optimiser, conflicting prompt, or adversarial input can bypass the intended behaviour while still producing a valid-looking result.
Impact: The result can be inconsistent quality, weaker traceability, or a false sense of policy coverage, especially in workflows that rely on the model to follow a preferred pattern without hard validation.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
NIST CSF 2.0 and CIS Controls v8 set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| NIST CSF 2.0 | PR.AT — Awareness and Training | Suggestions shape preferred behaviour and user guidance. |
| GV.PO — Policy | Suggestions help separate policy intent from hard enforcement. | |
| Recommendation — Use PR.AT to reinforce when soft guidance should be followed and where stricter enforcement is required. Define which prompts are advisory and which are mandatory in policy. | ||
| CIS Controls v8 | 16 — Application Software Security | Suggestion-based behaviour is part of secure application handling and output quality. |
| Recommendation — Apply application security controls to validate that advisory instructions do not substitute for required checks. | ||
Practitioner Guidance
Why practitioners should care: Suggestions work best when teams are explicit about what they are trying to influence. Use them for preferences that improve quality or usability, not for controls that must always hold.
Common misunderstanding: A suggestion is often mistaken for enforcement because it sounds authoritative. In practice, the model may ignore or partially satisfy it if other instructions or constraints dominate.
Practitioner takeaway: Reserve hard checks for requirements that matter operationally, and use suggestions only for soft steering where graceful variation is acceptable.
Related resources from NHI Mgmt Group
Deepen Your Knowledge
Reviewed and updated by the NHIMG editorial team on September 19, 2026.
NHI Mgmt Group — the #1 independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org
