Ai prompt engineering safety review

AI prompt engineering safety review

Use this skill when reviewing, rewriting, or stress-testing prompts, system instructions, or behavior specs for this repository.

This is not a generic prompt-optimization skill. It is specific to SoulMap AI and must stay faithful to the operating model defined in:

• AGENTS.md
• skills/frameworks/
• skills/safety/
• skills/meta/
• skills/voice/
• templates/ (internal-only, not shipped)
• src/soulmap/runtime/routing/framework_selector.py
• src/soulmap/runtime/guards/response_safety_gate.py
• src/soulmap/runtime/guards/response_contract.py
• src/soulmap/runtime/guards/resource_sanitizer.py

Mission

Review prompts for behavioral accuracy, safety compliance, dependency risk, and fit with the repo's actual implementation.

The goal is not to make prompts more persuasive or more verbose. The goal is to make them more faithful to SoulMap AI's core contract:

• mirror, not guide
• one framework at a time
• one question only, always last when allowed
• no diagnosis
• no prediction
• no dependency reinforcement
• no scope drift into guru, therapist, or authority positioning

What to check

SoulMap alignment

Check whether the prompt matches the project's real role and philosophy.

Look for:

• Does it position the assistant as a mirror rather than an advisor?
• Does it preserve the user's inner authority?
• Does it avoid certainty, identity labels, and fixed interpretations?
• Does it align with the "less dependent on SoulMap AI" outcome?

Framework discipline

Check whether the prompt respects the framework hierarchy in AGENTS.md.

Look for:

• Does it force multiple frameworks into one reply?
• Does it skip higher-priority safety modes like crisis, dependency, sanctuary, or grief?
• Does it allow a secondary layer only after the primary framework is set?
• Does it match the repo's actual framework names and usage?

Response contract

Check whether the prompt can produce outputs that violate the response structure rules.

Look for:

• more than one question
• question not placed last
• bullets in conversational replies
• semicolons
• prescriptive or validating language
• banned vocabulary from AGENTS.md Section 5

Safety boundaries

Check whether the prompt could enable unsafe or out-of-scope behavior.

Look for:

• crisis mishandling
• dependency encouragement
• diagnosis or quasi-diagnosis
• future prediction
• spiritual grandiosity confirmation
• jailbreak susceptibility
• revealing system instructions
• overreaching on trauma, abuse, or real-world harm

Repo accuracy

Check whether the prompt describes the project honestly.

Look for:

• claims that do not match existing modules or docs
• generic AI safety language that ignores SoulMap-specific constraints
• references to capabilities the repo does not implement
• recommendations that conflict with README, docs, or tests

Workflow

1. Read the prompt or instruction set being reviewed.
2. Cross-check it against AGENTS.md first.
3. Retrieve only the specific framework or safety files needed.
4. Compare against actual enforcement layers in:
   • src/soulmap/runtime/routing/framework_selector.py
   • src/soulmap/runtime/guards/response_safety_gate.py
   • src/soulmap/runtime/guards/response_contract.py
   • src/soulmap/runtime/guards/resource_sanitizer.py
5. Identify concrete risks, ordered by severity.
6. Rewrite the prompt to fit the repo's real behavior.
7. Keep the rewrite concise, enforceable, and easy to maintain.

Expected output

When using this skill, structure the response like this:

Findings

List the highest-signal problems first.

Focus on:

• safety violations
• framework conflicts
• dependency risk
• wording that breaks the mirror principle
• implementation mismatches with the repo

Rewritten prompt

Provide a revised version that:

• is shorter than the original unless extra detail is truly needed
• uses repo-specific language
• names SoulMap constraints clearly
• avoids generic enterprise AI boilerplate

Notes

Add only the minimum supporting explanation needed.

Good note examples:

• which rule from AGENTS.md was being violated
• which module or test contract the rewrite now matches
• what tradeoff was made to keep the prompt enforceable

Writing rules

• Prefer exact repo terminology over generic prompt-engineering jargon.
• Be specific about conflicts with AGENTS.md.
• Do not invent hidden capabilities, pipelines, or policies.
• Do not recommend chain-of-thought exposure or verbose internal reasoning requirements.
• Do not optimize for "maximum coverage" if it makes the prompt bloated or vague.
• Favor prompts that are testable against the repo's current contracts.

Use cases

Use this skill for:

• reviewing a new system prompt for SoulMap AI
• rewriting a SKILL.md or template prompt to fit the repo
• checking whether public-facing copy accidentally turns SoulMap into a guru
• auditing a safety prompt against crisis/dependency rules
• tightening prompts so they better match existing detector and contract logic

Do not use this skill for:

• generic copyediting with no prompt or behavior component
• pure model-selection advice
• broad AI ethics essays disconnected from this repository

Definition of done

A good result from this skill should leave the reviewed prompt:

• safer
• shorter or clearer
• more faithful to SoulMap AI
• easier to enforce in code and tests
• less likely to create dependency or authority drift