writing-cybernetic-goals

name	writing-cybernetic-goals
description	Use when an existing v1 chain requires goal.control.json or bounded runtime compatibility.

Writing Cybernetic Goals

Legacy v1 compatibility only. V2 objective and completion obligations live once in commitment.control.json.

Overview

Write the goal contract from confirmed requirements and any required design.

This skill does not analyze requirements, design the solution, write execution policy, review, compile runtime, or execute target work.

Detailed rules live in references/goal-writing-detailed-rules.md and references/control-contract-rules.md.

Runtime Limit

goal.control.json is not the user-entered runtime /goal. Runtime execution must use either:

• using-bounded-control-json for bounded_runtime; or
• using-control-json for compiled JSON pre-goal runs.

The user-entered /goal must stay pointer-only.

Goal Modes

Mode A: Complex Control Contract

Use for controlled_run JSON pre-goal orchestration. Requires approved requirements, any required design, and a completed information sufficiency gate: for schema 1.1.0+, approved_control.information_sufficiency_check is satisfied or reviewed not_required, has run-local evidence_ref values, and has approved independent counterexample_review. Otherwise route to RunInformationSufficiencyCheck.

Output:

docs/cybernetics/runs/<slug>/goal.control.json

The goal must preserve What the User Approved, source requirements, required outcomes, non-goals, authority, evidence, and final answer format.

Mode B: Bounded File Goal / Audit Goal

Use for bounded_runtime work with fixed meaning.

Output:

docs/cybernetics/runs/<slug>/goal.control.json
docs/cybernetics/runs/<slug>/runtime.control.json

The direct /goal pointer uses .agents/skills/using-bounded-control-json. This mode does not require requirements/design/plan/review.

Required Goal Content

For complex goal control, include:

• objective;
• human purpose;
• source of truth;
• rules that cannot change;
• required outcomes;
• what counts as done;
• where the result must show up;
• final answer format;
• work covered and allowed actions;
• forbidden live or irreversible actions;
• stop conditions.

For bounded goals, include:

• fixed objective;
• bounded scope;
• allowed and forbidden actions;
• required steps;
• evidence required;
• verifier command.

Checks

Evaluation tasks need an explicit rubric before goal writing. Output-sensitive tasks need final answer format. Multi-place result tasks need result placement. If any of these are missing, return to requirements or design instead of guessing.

Output Format

For complex goal control, output:

Created or updated goal:
`docs/cybernetics/runs/YYYY-MM-DD-slug/goal.control.json`

Status:
- `approved` / `candidate` / `blocked`

Response-only next step:
- return to `$orchestrating-cybernetic-pregoal`; include `允许使用 subagent 执行 review。`, or
- if manual fallback is being used, continue to `$writing-cybernetic-execution-policies`.

For bounded JSON goal, output the pointer:

/goal Use .agents/skills/using-bounded-control-json and execute docs/cybernetics/runs/YYYY-MM-DD-slug/runtime.control.json. If the bounded JSON is missing, invalid, inconsistent, or insufficient, stop and report the smallest required human decision.

Do not write conversational next-step prompts into goal control JSON.

Validation Checklist

• Requirements and any required design exist.
• Controlled runs have What the User Approved: Approved.
• Controlled runs pass the information sufficiency gate before goal writing.
• Goal does not weaken source requirements.
• Runtime /goal is pointer-only.
• bounded_runtime mode points to using-bounded-control-json.
• Complex mode does not let runtime write its own plan.
• Missing rubric/design/output requirements are routed upstream.