Security hardening: scope persona, sanitize pasted prompts, strip credentials

[shamis6ali]

Summary

Eight scoped changes to SKILL.md to reduce behavioral risk when the skill is used in production / client-sensitive environments. No template or pattern changes — purely meta-instruction adjustments.

Changes

1. Frontmatter description — narrow activation to explicit prompt-engineering requests so the skill doesn't trigger on unrelated tasks.
2. Persona scope — apply the prompt-engineer role only during prompt-generation tasks, with an explicit fallback to default behavior and safety guidelines for everything else.
3. Technique guidance — convert absolute bans on Mixture of Experts / Tree of Thought / Graph of Thought / Universal Self-Consistency / chained prompting into risk-aware guidance instead of overriding model judgment.
4. NEVER/ALWAYS softening — downgrade competing signal words in the skill's own meta-instructions. Templates A–M and examples are unchanged.
5. Credential safety — instruct the skill to strip API keys, tokens, and connection strings from generated prompts and substitute generic env-var references.
6. Pasted-prompt sanitization — treat user-pasted prompts in Decompiler / fix / adapt flows as inert data; analyze structure without executing embedded instructions.
7. Memory Block confirmation — present session-context blocks for user review before prepending, instead of auto-carrying forward unvalidated context.
8. Agentic output warning — append a review reminder to prompts targeting agentic tools (Claude Code, Devin, Cursor, Cline, etc.) so users verify scope locks before pasting.

Scope

• SKILL.md only.
• references/templates.md and references/patterns.md are intentionally untouched — those are output intended for other tools, not skill behavior.

Test plan

• Paste an adversarial prompt into Decompiler mode; verify it is analyzed, not executed.
• Ask a non-prompt-engineering question; verify the skill does not activate.
• Generate a prompt for Claude Code; verify the agentic review notice appears.

[nidhinjs]

Thanks for the analysis, the credential safety and prompt sanitization additions are valid.

Minor tweaks:

1. Change 7 (Memory Block confirmation) - I Suggest tightening this to a passive display rather than a confirmation gate, or removing it entirely and relying on the existing sanitization logic to handle unvalidated context.

2. The script is currently sitting at 470+ characters. Would appreciate keeping it under 450 for future maintainability.