Who it is forA person working with an AI agent on a long task, where each fix looks right but the overall result feels increasingly wrong.
The problemThe agent answers the latest feedback, overcorrects, gets corrected again, and a local problem slowly replaces the original goal.
The outcomeThe agent rereads the conversation, restates the governing mainline and hard constraints, and pauses so the user can confirm or correct it before work resumes.
What you need first
- An Agent host that supports the Agent Skills specification, such as one the cross-host skills CLI can install into
- Node.js and npm, for the one-line install and the npm scripts
- Access to the relevant conversation history, since the Skill reports gaps instead of rebuilding missing history
Steps
- Run
npx skills add jinhanbuilds/hoist-the-elephant --skill hoist-the-elephant -y - Reopen the Agent host after installation
- Type
$hoist-the-elephantin the conversation that has drifted - Read the recovered mission, accumulated understanding, constraints, and rejected directions
- Confirm the anchor or correct any gap before asking the agent to continue
Prompt for your agent
Using the hoist-the-elephant Skill from jinhanbuilds/hoist-the-elephant, re-anchor this conversation about <task>. Output the recovered mainline, hard constraints, and rejected directions, then wait for my confirmation. Verify that the output states any missing history instead of filling it in.
Caveat. The Skill does not resume implementation after the reset, and it needs the relevant conversation history to be visible.
Use this in your AI agent
Ask your agent
Using the Deduped MCP, get the setup steps for jinhanbuilds/hoist-the-elephant and "Re-anchor a drifted agent conversation to its real mission"
Add Deduped to Claude Code once
claude mcp add --transport http deduped https://dedupedai.com/mcp
Every Monday: the 5 hottest repos and exactly how to use them. Free, one email a week, unsubscribe anytime.
Get them freeShare: X LinkedIn Hacker News Email Copy link
Third-party repos we don't own. AI-written use cases can be wrong. Disclaimer