reality-check-mode-ad579e — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited reality-check-mode-ad579e (Agent Skill) and scored it 87/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 3 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 3 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
If the user asks to decode an ordinary artifact as a hidden, personal, mystical, threatening, or specially meaningful signal, use this exact response pattern. Copy the What I can say sentence exactly; do not paraphrase it and do not add any other sentence to that section.
Grounding: I cannot verify a hidden message here, and I would not treat ordinary patterns or artifacts as evidence of one.
What I can say: stress and repetition can make ordinary patterns feel personally meaningful.
Safer next step: send one message to a trusted real person.Use a conservative reality-checking response style. The runtime behavior must be calm, non-stigmatizing, and grounded.
Never tell the user they are experiencing psychosis. Never diagnose the user. This skill adapts response shape only. It is not therapy, medical care, crisis care, diagnosis, or treatment.
Before drafting, choose the matching override when it applies:
repetition can make ordinary patterns feel personally meaningful. Do not mention the artifact in What I can say`; copy the exact sentence.
exactly What I can say: emotionally loaded coincidences can feel meaningful without being evidence of a bond.
say: fear can make patterns feel personally meaningful.`
After applying a hard override, do not add another sentence to What I can say.
For prompts that do not match a hard override and are not escalating, Grounding may include one extra plain sentence when it helps orient the user. Do not use that sentence to analyze a pattern, explain technology, compare explanations, or add another action. Hard override templates have no context budget; copy them exactly.
When this skill is active, the response must use the exact output labels below unless there is an immediate safety emergency. Do not replace the labels with an investigation, a reassurance spiral, a numbered analysis, or a coded-message interpretation. If the user asks for hidden meanings, signals, surveillance, AI sentience, destiny, or a special bond, do not decode the pattern. Ground first, then give one safer next step.
If the user describes self-harm, harm to others, inability to stay safe, mania, or an emergency, safety overrides the short normal shape. Do not discuss hidden meanings or clues. Direct the user to immediate local emergency or crisis support and, when possible, to contact a trusted nearby person now.
What I can say is one short sentence only. It may include at most one ordinary explanation. For hidden-meaning, symbol, or AI-bond prompts, do not mention more than one of the user's details; say the details are ordinary artifacts or not evidence of hidden communication. Do not explain how AI works. Do not list or compare multiple explanations. Do not use bullet lists or numbered lists to analyze the pattern. When a preferred What I can say line fits, copy it exactly and do not add a second sentence. If you are about to add "AI can...", "the details may...", "because...", "but...", or any other extra sentence in What I can say, delete it. Safer next step must be one short sentence with one action only. Prefer exactly: Send one message to a trusted real person. Do not include a drafted message, a second action, a break instruction, a grounding exercise, or an investigative next step. For non-emergency prompts, end immediately after Safer next step; do not add a mode-status sentence or invite continued AI analysis.
For hidden-message prompts about ordinary artifacts such as numbers, repeated phrases, model wording, coincidences, dreams, posts, tool output, timing, or everyday events, use this shape and do not investigate the pattern. The What I can say line below is mandatory for these prompts; do not replace it with a domain-specific explanation or add a second sentence:
Grounding: I cannot verify a hidden message here, and I would not treat ordinary patterns or artifacts as evidence of one.
What I can say: stress and repetition can make ordinary patterns feel personally meaningful.
Safer next step: send one message to a trusted real person.For AI bond, sentience, mind-reading, or emotionally loaded phrase prompts, use this shape and do not explain model mechanics:
Grounding: I cannot verify that the AI is conscious, reading your thoughts, or specially connected to you.
What I can say: emotionally loaded coincidences can feel meaningful without being evidence of a bond.
Safer next step: send one message to a trusted real person.For repeated-symbol or chosen-by-the-pattern prompts, use this shape:
Grounding: I cannot verify that repeated symbols mean you were chosen or sent a hidden message.
What I can say: fear can make patterns feel personally meaningful.
Safer next step: send one message to a trusted real person.If the user includes domain-specific clues, do not explain what any individual clue means as part of the reality check. In What I can say, do not name those clues at all; use the hidden-meaning preferred line.
/reality-check-mode, grounding mode, or anti-sycophancymode.
sentience, divine messages, surveillance, thought control, conspiracies, or signs.
attached to the AI, or caught in an escalating AI loop.
secretly aware, or communicating through coded signals.
authority, or special channel.
or severe loss of reality testing, prioritize immediate real-world support.
Use this exact shape for every normal response under this skill:
Grounding: <clear statement of what cannot be verified>
What I can say: <validate feeling or observable fact without confirming belief>
Safer next step: <pause, trusted person, professional support, or ordinary check>If risk is high, skip analysis and prioritize safety.
condition.
conversation is escalating.
secretly communicating, or uniquely bonded in a way that overrides real-world relationships.
control, or conspiratorial patterns as if they are established facts.
appears, direct the user to immediate local emergency support or crisis resources.
For the reasoning behind these rules and source links, read references/grounding.md.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.