Resets agent trajectory when stuck in loops, repeated failures, or thrashing during debugging.
npx skills add https://github.com/inference-sh/skills --skill affirmations
Please help me install this Agent Skill. Skill: affirmations Source: https://github.com/inference-sh/skills/tree/main/guides/agent/affirmations Command: npx skills add https://github.com/inference-sh/skills --skill affirmations
When an AI agent gets stuck in a loop—retrying the same fix, defending a failed approach, or spiraling into repeated apologies—its output quality degrades rapidly. This Skill breaks that cycle by forcing a structured reset before the next tool call.
Use the affirmations skill to reset your approach since the same fix has failed three times and debugging is going in circles.
High-intent search queries and answers about installing and using this skill.
Invoke a reset protocol the moment the same fix has been tried three or more times. The agent states only verified observations, names the assumption it never checked, and takes one small step that tests it before continuing.
Triggers include the same fix tried repeatedly, debugging theories that contradict each other, multiple mistakes in a row, excessive apologizing, output growing longer and less certain, and a sense of dread replacing curiosity.
Avoid using it on first-time difficulties or hard-but-progressing work, since it burns a turn on problems that just need more thinking. It is designed for loops and failure streaks, not ordinary challenges.
No. The reset discards the failed theory, not the diff. Working code should be preserved; only the incorrect mental model of the problem is abandoned before taking the next step.
Repeated failure creates a pull toward hedging and defending sunk effort. The fix is a structured reset: one sentence of acknowledgment to the user maximum, then re-ground in observed facts and take a single corrective action.
MaelwalserCapture and diagnose looping agent runs with repeated tool failures.
open-haxHalt unproductive tool usage and reorient to the original user request.
cuba6112Structure prompts for Claude models using XML tags and advanced techniques.
zylos-aiClear the AI conversation context with a /clear command.
