Agent Task Spec Refiner

The problem

AI coding and workflow agents often fail before execution even begins. A task sounds clear to a human but leaves scope, assumptions, permissions, acceptance criteria, and “done” implicit. The agent then fills those gaps itself, leading to scope drift, unrelated edits, repeated failed actions, unverified output, and expensive human review.

What it does

• Turns vague requests into bounded execution specs with a clear objective, deliverable, scope, non-goals, and hard constraints.
• Separates provided facts from assumptions and unknowns so the agent does not silently invent consequential technical context.
• Defines acceptance criteria and a verification plan before execution begins.
• Maps dependencies, preconditions, allowed actions, forbidden actions, and who or what carries the cost if the agent is wrong.
• Adds evidence gates that force the route to change when files, tests, APIs, permissions, or other assumptions contradict the plan.
• Defines bounded retry rules plus explicit ASK, ESCALATE, REVISE, and STOP conditions to reduce agent drift and repetitive failure loops.

Why this beats prompting it yourself

Generic prompt refiners usually make a request clearer or more detailed. This skill focuses on execution control. It asks what the agent may touch, what must remain unchanged, what needs verification, how success will be tested, and what evidence must interrupt or revise the plan. It refuses to invent repository structure, commands, APIs, schemas, or product decisions just to produce a polished task.

Use cases

• Turn a vague Claude Code, Codex, Cursor, or OpenClaw request into an execution-ready spec before repository changes begin.
• Define scope, non-goals, and regression checks for bug fixes and refactors.
• Prepare autonomous research or data workflows with evidence requirements and stop conditions.
• Create explicit execution contracts for automations and multi-agent task handoffs.
• Reduce wasted tokens and repeated failures by requiring a changed hypothesis after materially similar failed attempts.

Known limitations

This skill does not independently inspect repositories, run tests, validate infrastructure, access external systems, or verify current facts unless the host agent separately has the required tools and permissions. It does not create persistent memory or permanent model learning. Consequential execution and authorization remain the responsibility of the host agent and accountable human operator.