goal-engineering

Goal Engineering

Render a goal from an already selected methodology. Do not choose the methodology, execute the goal, or decide that the resulting loop is ready.

Installed dependency

The renderer imports the canonical manifest from the sibling methodology-selector skill. Confirm that both skills are installed under the same skills root. If the sibling is missing, stop and tell the user to install the pair; do not run a broken command or recreate the manifest.

Inputs

Read references/input-contract.md. The eight shared fields are always explicit:

1. Methodology
2. Task
3. Audience
4. Context
5. Output Format
6. Validation
7. Method Fit
8. Human Review / Stop

The control envelope adds boundaries and fallback. Add smallest useful run, checker, discovery, execution handoff, persistence, schedule policy, budget, maximum iterations, isolation, or feedback only when the task makes them applicable.

Workflow

1. Confirm that methodology selection is complete. Do not repair or replace it.
2. Separate the measurable outcome from implementation steps.
3. Make validation evidence visible to the checker; a goal cannot rely on a checker reading evidence that the executing agent never surfaces.
4. Add boundaries before retry logic so a later attempt cannot redefine success.
5. Separate implementer, independent reviewer, and human authority when broad or mutating work needs an execution handoff.
6. Add scheduling, iteration, and feedback controls only when the cycle needs them.
7. Add explicit success, failure, no-progress, and human-return conditions.
8. Render with scripts/render-goal.mjs.
9. Report the exact character count and effective limit.
10. If the output exceeds the limit, return the overage and ask the caller to remove or rewrite optional context. Never truncate.

Character budget

• Every target tool has a common ceiling of 4,000 Unicode characters.
• A supplied stricter tool limit wins.
• Aim for 3,500 characters or fewer to leave revision headroom.
• Preserve task, validation, boundaries, terminal conditions, and fallback before optional explanatory fields.
• Omit inapplicable optional sections; do not emit empty headings.

Command

node /absolute/path/to/goal-engineering/scripts/render-goal.mjs /absolute/path/to/input.json

Resolve the first path from the directory containing this SKILL.md. The agent prepares the JSON input; do not ask the human to hand-write it. Successful output is JSON containing goal, character_count, limit, within_limit, target_tool, and omitted_optional_fields. An over-limit render returns the same evidence and exits non-zero.

Boundaries

• Do not select or reinterpret the methodology.
• Do not silently shorten user-provided requirements.
• Do not hide an over-limit result.
• Do not claim readiness.
• Do not execute, schedule, or persist the loop.

Stop

Stop after one valid render or one explicit over-limit failure.