forge-prompt-engineering

Prevent prompt injection by treating user content and retrieved context as data, clearly delimited from instruction layers. This skill provides anti-injection and data separation patterns to ensure system prompts remain isolated from adversarial user inputs.

Enforce structured output by applying strict JSON schemas and parser-validated outputs so responses fail safely instead of silently drifting. This skill guides you in implementing schema validation to guarantee reliable structured generation.

Version production LLM prompts by treating them like production interfaces with tracked versioning, pinned models, and temperature discipline. This skill audits prompt codebases to ensure proper hygiene, versioning, and observable deployments.

LLM output drifts when prompts lack layered structure, strict schemas, or defined refusal paths for low-confidence scenarios. This skill audits prompts against median, edge, and adversarial inputs to identify and resolve these failure modes.

Yes, you can use a shell verifier to flag common prompt mistakes in implementation files. This skill provides a verifier workflow to detect issues like boilerplate roles, unsafe f-string interpolation, and unpinned model aliases.