goal-rubric

name	goal-rubric
description	Use when the user wants a measurable /goal rubric, success criteria, acceptance criteria, or completion condition, or when a /goal loop will not close because the criteria are not observable or binary.

Goal Rubric

Turn a short goal into a binary rubric that a /goal grader can check, then render it as a ready-to-paste completion condition.

Core Principle

A /goal loop closes only when the completion condition is observable and binary. Write criteria the working agent can prove in its own output and a small grader model can judge without running commands or reading files.

Grader Constraints

• Transcript-only: the grader sees what the agent surfaced in the conversation. It does not run commands or inspect files.
• Small-model judgeable: each criterion must be atomic, concrete, and unambiguous.
• Binary AND: the goal is complete only when every criterion passes.

Workflow

1. Get the goal from the user.
2. Inspect the repo read-only to identify test, build, lint, file layout, and artifact checks that can prove completion.
3. Draft independent pass/fail criteria. Each criterion must include an end state, a transcript-observable check, and any relevant guardrail.
4. Ask only about gaps that repo inspection cannot settle, such as the validating command, protected files, or turn/time cap.
5. Save the rubric to ./goal-rubric-slug.md, replacing slug with a short goal-derived filename segment unless the user gives another path.
6. Render the /goal condition for Codex by default. If the user asks for another tool, use that tool's framing.

Rubric Shape

# Goal rubric: short goal statement

## Criteria (all must pass)
1. Criterion name - End state: observable signal. Check: how it is proven in the transcript. Constraint: what must not change.

## Stop clause
Turn or time cap.

## /goal condition (Codex)
Codex condition string.

Codex Condition Framing

For Codex, phrase the condition with four parts:

• What to achieve.
• What not to change.
• How to validate it.
• When to stop.

Keep the condition direct enough for a small grader to answer yes or no from the transcript.

Self-Check

Before finishing, verify:

• Every criterion names a measurable end state.
• Every criterion states the check that proves it.
• Every proof can appear in the agent transcript.
• No criterion requires the grader to run a command or open a file.
• Criteria do not overlap.
• The rubric has a stop clause.