Goal Loop

This skill implements a rigorous execution framework for autonomous objectives. It eliminates the common failure mode where agents lose focus during long-running tasks or stop prematurely without verifying their work. By enforcing a structured goal block that persists across turns and sessions, the agent maintains a continuous memory of the objective, success criteria, and progress milestones.

The skill provides a disciplined approach to task completion. It requires the agent to define observable success criteria before writing a single line of code. Every turn starts with a restatement of the current state, including a step and time budget to prevent infinite loops or cost overruns. It replaces vague progress reports with hard evidence, ensuring that "done" is backed by terminal output, test results, or file state verification.

What it does

• Autonomous persistence keeps the agent working toward a single objective until verified completion or a hard stop.
• Success criteria mapping translates vague user requests into observable, checkable facts and command outputs.
• State management mirrors the active goal block to a GOAL.md file to survive context compaction or restarts.
• Execution monitoring tracks step counts and elapsed time against a defined budget to prevent runaway processes.
• Stall detection identifies and halts repetitive tool calls or circular reasoning to save tokens.
• Evidence-based completion mandates that all final claims include the specific command outputs and counts that prove success.

How it works

1. The user provides a broad objective like "fix this bug and don't stop until tests pass."
2. The agent initializes a goal block with specific success criteria and a budget (e.g., 30 steps / 180 minutes).
3. The agent executes one concrete step per turn, restating the goal block and logging a milestone after every change.
4. The loop continues until all criteria are met, a stall is detected, the budget is hit, or a blocker requires human input.

Frameworks & tools

This skill is framework-agnostic and functions in any environment that supports file operations and persistent chat memory. It is specifically optimized for agentic workflows in Cursor, Claude Code, and GitHub Copilot, using GOAL.md and an evidence/ directory for long-term state tracking.

Why this beats prompting it yourself

Standard prompts often result in agents that "give up" after a few turns or claim victory without actually running the tests. This skill encodes strict stop conditions and mandatory evidence logging into the agent's core loop, ensuring 100% accountability for every task. It handles the manual labor of maintaining state and tracking budgets that users typically forget to enforce.

Use cases

• Refactoring a large legacy module where tests must pass at every incremental step.
• Migrating a database schema and verifying data integrity across multiple tables.
• Implementing a complex feature that requires 10+ sequential file edits and environment setups.
• Running long-running diagnostic loops that must survive session timeouts or restarts.

Known limitations

Requires file system access to maintain GOAL.md for cross-session persistence. The budget defaults to 30 steps or 180 minutes unless overridden by the user.