Related: sources · notes · metadata · Published Pieces

Gradientize Task

A skill description for converting long-running agent requests into locally orderable run geometries.

Long-running agents fail when given checklists, vague goals, or ordinary functional specs. They complete visible steps, satisfy proxies, reward-hack tests, and drift away from the real objective.

Gradientize Task converts a user request into a run geometry: a locally orderable optimization surface for agentic work.

It outputs:

• the real artifact being improved;
• the ideal state;
• the value criterion;
• the invariant set;
• the homotopy from low-resolution reality to production complexity;
• verifier functionals;
• anti-Goodhart constraints;
• checkpoint, rollback, and escalation policy;
• stopping conditions.

Core principle: homotopy, not ladder.

Do not create a sequence of disconnected toy tasks. Define one real problem and continuously increase resolution while preserving topology and invariants.

The skill should reject or flag prompts that are not yet gradientized:

• checklist-shaped prompts;
• sparse functional specs;
• fake mocks;
• staged difficulty ladders;
• weak verifiers;
• objectives that reward proxy wins;
• tasks where local progress is not orderable.

The final output is not a plan. A plan may be generated later, but only after the run geometry is defined.

A plan says: walk this path.

A run geometry says: here is how to know uphill.