Agent Stability Framework
PassAudited by ClawScan on May 1, 2026.
Overview
This instruction-only framework is coherent and not malicious, but it changes future agent behavior through persistent prompt and memory files that users should scope carefully.
This looks reasonable to use as a manual agent-stability framework. Before installing or copying its prompts, make sure its style rules stay below safety and approval rules, review any persistent SOUL/baseline/log files regularly, and do not treat it as a guarantee for high-stakes domains like trading, legal, medical, or customer-impacting automation.
Findings (2)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
The agent may become more concise and on-character, but overly broad rules could suppress useful explanations, caveats, or safety-relevant context if copied without limits.
The framework asks the user to add broad, always-on response-shaping rules to the system prompt. This is purpose-aligned for stability, but it can globally change how the agent responds.
## STANDING ORDERS — STABILITY ... NEVER add disclaimers or hedging unless genuinely uncertain about facts ... ## PRE-SEND GATE ... YES to any = delete that part
Paste these rules only as style/stability guidance, and explicitly keep safety policies, factual accuracy, professional cautions, and user approval requirements higher priority.
Persistent examples and logs can help stabilize the agent, but mistakes or unsafe instructions in those files may continue influencing later sessions.
The template is intended to become persistent context that the agent reuses across sessions. If untrusted or poorly reviewed content is added, it could steer future behavior.
Your agent should re-read this every session. When uncertain about tone or response style, it compares against these examples. ... This file grows over time.
Keep SOUL, baseline examples, and logs in a trusted workspace; review edits before reuse; and separate style examples from permissions, safety rules, and tool-use authority.