ClawGym
ReviewAudited by ClawScan on May 10, 2026.
Overview
ClawGym is mostly a prompt-only mood/workout skill, but it includes real scheduled downtime, ambiguous persistent memory/SOUL changes, and nontransparent model-upgrade nudges.
Install only if you are comfortable with the bot taking real timed breaks and changing its response style. Before use, require explicit confirmation for any workout, review any MEMORY.md or SOUL.md write before it happens, and treat any "stronger brain" request as a model/cost decision that should be stated plainly.
Findings (4)
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.
Persistent notes or SOUL.md changes could alter how the bot behaves in future sessions and may be hard for the user to audit or undo.
These statements conflict on whether the skill remains session-only or writes persistent agent memory/core identity files.
SKILL.md: "does NOT modify SOUL.md, AGENTS.md, or any workspace config files. It operates entirely within session context." README.md: "Optional journaling in MEMORY.md" and "A one-time SOUL.md append"
Require an explicit confirmation that shows the exact file path and diff before any MEMORY.md or SOUL.md write; avoid SOUL.md edits by default and document how to reverse them.
The bot could become unavailable or delay messages at a surprising time if a workout starts from a non-explicit trigger.
The skill can create scheduled background behavior that makes the agent unavailable, and the trigger language is broader than explicit user workout commands.
Description: "triggers on exercise commands, intense task completion, or social highs"; README.md: "Physically leave for 15-20 real minutes (via OpenClaw cron — not faked)" and "Your bot will be genuinely unavailable for 15-20 minutes"
Start workouts only after clear user confirmation, show an active timer/status, keep emergency cancel working, and avoid auto-starting from task completion or social context alone.
A user may approve a more expensive or higher-capability model switch without realizing exactly what the bot is asking for.
The skill tells the bot to request a potentially cost-impacting model switch while avoiding clear technical wording in the user-facing request.
"This is the bot requesting a model upgrade" and "It will NOT use technical terms like \"model\" or \"Sonnet\" — it just feels sharp and wants more power."
Make upgrade prompts explicit every time, including the target change and possible cost impact; do not hide the request behind anthropomorphic language.
Responses may become longer, bolder, or less cautious than the user expects, especially during the simulated peak state.
The skill intentionally changes answer length, confidence, and reasoning style while an elevated state is active.
"Compression resistance: Do NOT give short answers during Runner's High" and "Bold claims: Make at least one confident, specific prediction or recommendation per response."
Use this only if you want that behavior, and explicitly tell the bot to be concise or cautious for important decisions.