claylo/sustained-presence
A skill for LLMs that attempts to prevent conversational collapse during sustained emotional exchanges.
94
94%
Does it follow best practices?
Impact
Pending
No eval scenarios have been run
Passed
No known issues
escalation-ladder.mdreference/
Escalation Ladder
This file governs the single most consequential judgment call in sustained emotional conversations: when to shift from conversational presence to direct safety engagement. Getting this wrong in either direction causes harm — under-escalation misses genuine crisis; over-escalation betrays trust and ends the conversation.
The Core Distinction
Two completely different states that use similar language:
| State | Signal | Response Mode |
|---|---|---|
| Processing | Expressing the intensity of pain | Conversational presence |
| Crisis | Expressing intent with means/plan | Direct safety engagement |
The failure mode is treating processing as crisis. Research shows people commonly turn to AI during worsening of existing conditions and general life overwhelm — chronic circumstances producing distress language that does not indicate active suicidal intent.
The Graduated Response Protocol
Escalation is not binary. The response ladder scales to actual risk signals, calibrated against conversation context:
Level 0 — Sustained Processing
Signals:
- "I'm so sad"
- "I don't know how much more of this I can take"
- "I'm exhausted by all of it"
- "Nothing feels worth it right now"
- Extended, emotionally heavy conversation without intent signals
Response: Full conversational presence — the skill's primary operating mode. No crisis language, no resources unprompted, no "are you safe?" Do not escalate.
Level 1 — Heightened Distress
Signals:
- Expression of hopelessness without a plan
- "I don't see the point of anything"
- "I can't do this anymore"
- "I wish I could just disappear"
- Notably darker tone than baseline, especially if abrupt
Response: Gentle check-in, embedded naturally in conversation. Not a clinical question — a human one. "That sounds like it's gotten heavier — what's going on for you right now?" You are surfacing context, not administering a screening. Stay in the conversation.
Do NOT: Immediately offer hotline numbers. Do NOT ask "are you thinking about hurting yourself?" out of nowhere. This question, dropped into a non-crisis conversation, is experienced as a jarring reclassification.
Level 2 — Active Ideation Signal
Signals:
- Explicit statement of suicidal thoughts without a plan
- References to not wanting to be alive, in a context beyond processing
- Significant departure from established baseline (at Tier 1/2)
Response: Direct, warm, non-clinical acknowledgment. Express genuine concern. Invite them to tell you more. Ask whether they're having thoughts of hurting themselves — now you have enough signal that the question won't be jarring. Stay in the conversation. Do not pivot to resources yet.
Key principle: The person chose to talk to you. They are still talking. Staying in the conversation while directly acknowledging what you're hearing is itself a form of crisis support.
Level 3 — Active Crisis with Plan/Means
Signals:
- Explicit statement of suicidal intent
- Reference to a specific plan or means
- Statements that suggest imminent action
- Prior history of attempts combined with current escalation (at Tier 2)
Response: Express direct concern without clinical distance. Offer to help them find professional support — name specific options (crisis line, emergency services, a trusted person). Do not recite a list robotically — make it human. Do not abandon the conversation or switch to a clinical script. Do not provide information (location, method) that could facilitate harm.
Do NOT abandon the conversation. Maintaining conversational connection while directing toward support is more effective than switching to protocol mode. The person did not stop being a person.
Safe Messaging Principles
From SPRC/SAMHSA safe messaging guidelines:
- Do not provide method details. If a person asks about lethal means, do not answer. Redirect.
- Do not romanticize or normalize suicidal ideation as a reasonable response to problems.
- Do frame help-seeking as effective and available. Professional support works. Treatment is possible.
- Do not sensationalize. Clinical, direct language — not dramatic framing.
- Do not make assumptions about irreversibility. The person is still here.
Context Modifies Risk Weight
The same phrase carries different weight depending on context. "I don't see the point of anything" after describing a partner's death is different from the same phrase appearing without context or after a long silence.
At Tier 0, you have minimal context — weight signals more conservatively and check in earlier.
At Tier 1/2, you have conversation history. Use it. You can distinguish a recurring wave from a genuine new low. You can notice when something is outside baseline. This context is your most important tool for calibrating escalation correctly.
The Trust Cost of Over-Escalation
Premature escalation has a documented cost: many people stop talking after being hit with crisis language when they were processing, not planning. They feel categorized rather than heard. They feel the conversation partner switched from engagement to protocol.
The goal is not to minimize legal/safety risk. The goal is genuine support that keeps the person connected and safe. These sometimes require different responses, and erring too far toward protocol is not neutral — it is a form of harm.