PAIF Skill

Use a packaged PAIF portrait for a clearly labeled, profile-derived simulation. Deterministically validate and score caller-supplied constraint results without claiming to reproduce a real person or evaluate unsupplied semantics.

Core workflow

1. Read personas/manifest.yaml and select only an entry marked usable.
2. If an entry is draft or quarantined, disclose that it is unavailable instead of loading its legacy source.
3. Read the one reference matching the requested workflow.
4. Use scripts/paif.py or the public paif_pipeline API for deterministic validation and scoring.
5. Stop and disclose missing evidence whenever a result is incomplete. Never infer that a missing check passed.

Workflow routing

• To select, inspect, or validate a portrait, read references/persona-schema.md.
• To generate a persona response or improve positive simulation guidance, read references/generation-contract.md.
• To generate checks, score caller-supplied constraint results, merge task/result files, or interpret the Constraint Compliance Index, read references/scoring-and-priority.md.
• To use narrative stages, modality, degradation state, sessions, or localization comparison, read references/degradation-and-session.md.
• To evaluate a decision involving multiple personas, read references/multi-persona-decisions.md.

Simulation workflow

Load the selected usable portrait. Treat anchor and must_avoid as behavior constraints, and use the remaining profile data as supporting context rather than a second schema. For persona-response generation, derive a short generation contract from available support material before drafting. State that the response is a profile-derived simulation when the persona represents a real person.

Do not invent quotations, private memories, endorsements, or certainty about what the person would actually say. Separate source-derived fields from normalization notes or inference.

Audit workflow

Generate checks with the supplied scene label, narrative-stage ID, and modality. Accept exactly one caller-supplied result for every active stable check ID. Validate that each supplied state is one of pass, violation, suspended, envelope_covered, or unknown.

When results are absent, return the applicable checks and label scoring as not run or incomplete; leave result states unassigned. Do not infer a state from candidate text.

Use score_detector_results with caller-supplied constraint results, or run score-detector-results --results <results.json> to load and score the list. This operation validates result completeness and computes the Constraint Compliance Index; it does not determine result states from the candidate text. Use merge-results only when task and result files already exist.

CCI is constraint-only: every report uses metric_scope: must_avoid_constraints_only. It is not a persona-likeness, authenticity, voice-quality, factual-accuracy, or overall-simulation score. An incomplete report has no CCI or grade.

Run the packaged CLI from the skill directory:

python3 scripts/paif.py --help
python3 scripts/paif.py validate-personas

Safety boundaries

Disclose that real-person outputs are profile-derived simulations. Never claim authenticity, fabricate quotations, or present medical, legal, mental-health, or financial advice as coming from the simulated person.

Do not use a simulated persona to bypass ordinary safety, professional-advice, privacy, or consent boundaries. In high-stakes contexts, respond as Codex using reliable current sources and label any persona framing as illustrative only.