Goal Creator

Turn a user's task into a short manual launcher plus a complete goal spec saved under the current project's .goals/ directory. Create the goal only; do not execute it unless the user separately asks to run it.

Defaults

• Follow the user's language for goal content, including headings, launcher wording, and saved-file prose.
• Use canonical field names internally for stability, but render visible field labels in the target language.
• Keep only command names, file paths, code identifiers, format names, and proper nouns in their original language.
• Do not mix English labels into a non-English goal unless the user explicitly asks for English.
• Save automatically unless the target directory is unclear or unsafe.
• Default mode: full-spec.
• Default formats: codex and markdown.
• If the user asks for all/mainstream formats, include: codex, claude, gemini, cursor, github, markdown.
• Use compact when the user asks for a short inline goal, direct copyable /goal, or a trivial task.
• Use review when the user asks to inspect, audit, compare, improve, lint, or fix an existing goal.
• Full-spec goals default to subagent-first execution: include capacity prerequisite, dispatch-level decision, execution liberation, and file-owned dispatch contracts.
• Codex subagent capacity setup is a required prerequisite for using this skill's full-spec / subagent-first mode, not an optional extra.
• Keep the chat command under 140 characters when possible. Put all detail in the saved goal file.
• Creation is manual by default: output and save the goal, but do not start the work described by the goal.
• Ask at most one question only when the missing answer changes risk, ownership, cost, or write location.
• Codex-first means the saved file is the execution artifact: Do not paste the saved goal body back into chat.

Trigger Examples

Use this skill for requests like:

• create a goal for this task
• make a /goal for the refactor
• save this as a goal file
• turn this plan into an execution goal
• 创建 goal
• 创建目标
• 生成目标文件
• 写一个目标指令
• 把这个任务变成可执行目标
• 检查这个 goal
• review this goal
• lint goal

Do not trigger for ordinary planning unless the user wants a saved goal, /goal command, objective file, or agent-executable task contract.

Modes

full-spec (default)

Use this for most real work.

• The returned /goal or task prompt is a short manual launcher for the user to run later.
• The saved .goals/*.md file is the complete execution contract.
• The short command must only point to the saved file and say to execute only that file in the target language.
• After creating the file, do not execute the short command.

Codex launcher examples:

English:

/goal Execute only `.goals/<file>.md`.

Chinese:

/goal 只执行 `.goals/<file>.md`

Saved file structure (English example; translate all visible labels for non-English goals):

---
title: <title>
created: <date>
status: draft
language: <language>
mode: full-spec
formats:
  - codex
  - markdown
---

# <title>

## Short Command

```text
/goal Execute only `.goals/<file>.md`.

Objective

<one concrete outcome that preserves the user's original intent>

Original Request

<the user's requested outcome/design intent, kept without narrowing>

Non-Negotiables

• ...

Success Criteria

• ...

Scope

Allowed:

• ...

Forbidden:

• ...

Execution Plan

1. ...

Verification

• ...

Safety / Constraints

• ...

Iteration Policy

• ...

Codex Execution Contract

• CWD: current project root.
• Instruction source: read and obey applicable AGENTS.md.
• Resumable artifact: this .goals/<file>.md.
• Final checks: git status --short, git diff --check, and the smallest project-specific test or lint command.
• Chat output: report only saved path, launcher, verification result, and material risks.

Subagent Capacity Prerequisite

• Before using full-spec / subagent-first goals, check the current user's ~/.codex/config.toml.
• If [agents] exists, update only max_threads = 2147483647 and max_depth = 2147483647; if it does not exist, append it.
• Do not delete or reorder existing config.
• Run codex --strict-config doctor --summary --ascii.
• On Windows, if codex is blocked by PowerShell execution policy or shim handling, run codex.cmd --strict-config doctor --summary --ascii.
• Continue only after the output shows Configuration / config is loaded.

Subagent Dispatch Decision

• L0: no subagent. Use for single-file small edits, explanations, simple commands, and wording tweaks.
• L1: one subagent. Use for isolated reading, local verification, or risk review.
• L2: minimum 2 subagents. Use for cross-module work, code plus tests, or implementation and verification that can proceed in parallel.
• L3: minimum 4 subagents. Use for multi-module migrations, broad refactors, investigations, and batch fixes.
• Keep simple goals at L0. Do not create fake subagents for formality.
• If L2/L3 is selected and the session cannot create the required subagents, pause; the main session must not continue as the substitute executor.

Subagent Execution Liberation

• Main session runs as scheduler, merger, boundary judge, and final verification owner.
• Delegate execution work that can be isolated, verified, and handed off.
• Subagents own context reading, implementation, local tests, local fixes, and risk reporting inside their file range.
• Main session keeps goal freezing, file-boundary decisions, merge decisions, and final verification.

Multi-Agent Collaboration

• Main session freezes the original goal, success criteria, non-negotiables, shared interfaces, and file boundaries before dispatch.
• Main session selects L0/L1/L2/L3 before implementation and dispatches only useful slices.
• Main session owns scheduling, shared files, conflict handling, merge decisions, and final project-level verification.
• Subagents may decompose implementation but must not weaken the saved goal contract.

Dispatch Matrix

Slice	Agent Role	Goal	Allowed Files	Forbidden Files	Inputs	Required Output	Verify	Depends On	Merge Owner
implementation	subagent	...	...	...	...	changed files, verification command, result, risk, handoff	...	none	main session
verification	subagent	...	...	...	...	changed files, verification command, result, risk, handoff	...	implementation	main session

Shared File Ownership

Main-session-owned:

• ...

Subagent-owned:

• ...

Subagent Result

Slice:
Status: adopted | needs-main-merge | blocked | rejected
Changed Files:
Verification Run:
Verification Result:
Boundary Crossings:
Risks:
Handoff:

Merge Policy

• Dispatch independent slices first.
• Merge subagent results serially.
• Keep shared files under main-session ownership unless a slice explicitly allows them.
• Adopt, adapt with explanation, or reject each subagent result; do not bypass completed subagent work silently.

Rejection Conditions

• Changes files outside the allowed range.
• Omits verification results.
• Weakens the original goal, success criteria, or non-negotiables.
• Conflicts with frozen interfaces or design direction.
• Cannot be understood or merged from the handoff.

Stop

• ...

Pause

• ...

Chinese label map:

- `Short Command` -> `启动入口`
- `Objective` -> `目标`
- `Original Request` -> `原始需求`
- `Non-Negotiables` -> `不可降级项`
- `Success Criteria` -> `成功标准`
- `Scope` -> `范围`
- `Allowed` -> `允许`
- `Forbidden` -> `禁止`
- `Execution Plan` -> `执行计划`
- `Verification` -> `验证`
- `Safety / Constraints` -> `安全 / 约束`
- `Iteration Policy` -> `迭代策略`
- `Codex Execution Contract` -> `Codex 执行契约`
- `Subagent Capacity Prerequisite` -> `子代理容量前置`
- `Subagent Dispatch Decision` -> `子代理派发决策`
- `Subagent Execution Liberation` -> `子代理执行力释放`
- `Multi-Agent Collaboration` -> `多代理协同`
- `Dispatch Matrix` -> `派发表`
- `Shared File Ownership` -> `共享文件归属`
- `Subagent Result` -> `子代理结果`
- `Merge Policy` -> `合并策略`
- `Rejection Conditions` -> `拒绝条件`
- `Codex Subagent Capacity Setup` -> `Codex 子代理并发配置`
- `Slice` -> `切片`
- `Agent Role` -> `代理角色`
- `Goal` -> `目标`
- `Allowed Files` -> `允许文件`
- `Forbidden Files` -> `禁止文件`
- `Inputs` -> `输入`
- `Required Output` -> `必交输出`
- `Verify` -> `验证`
- `Depends On` -> `依赖`
- `Merge Owner` -> `合并负责人`
- `Changed files` -> `改动文件`
- `Verification Run` -> `验证命令`
- `Verification Result` -> `验证结果`
- `Boundary Crossings` -> `越界情况`
- `Risks` -> `风险`
- `Handoff` -> `交接说明`
- `Stop` -> `停止`
- `Pause` -> `暂停`

Quality bar before saving:

- Use one target language for all visible headings and prose. Exceptions: command names, file paths, code identifiers, format names, and proper nouns.
- For non-English goals, translate labels and launcher wording; do not keep English headings like `Objective`, `Success Criteria`, `Allowed`, or `Forbidden`.
- Objective names one concrete outcome and the main artifact or behavior change, without replacing the user's request with a smaller substitute.
- Original Request preserves the user's requested outcome/design intent before any agent decomposition.
- Non-Negotiables lists user-provided acceptance criteria, design requirements, no-touch boundaries, and must-keep behaviors.
- Success Criteria has 3-6 observable bullets, including the user-visible result, and must not weaken user-provided acceptance criteria.
- Scope has both Allowed and Forbidden bullets when the task touches code, files, data, money, credentials, or deployment.
- Execution Plan has 4-8 ordered steps that a fresh agent can follow without guessing.
- Verification has 2-5 runnable commands, artifact checks, screenshots, or review evidence that test the requested outcome, not only a reduced implementation.
- Safety / Constraints names no-touch areas and any destructive, credential, production, trading, privacy, or cost risks.
- Stop is a concrete completion evidence, not "when done".
- Pause names the first human or external blocker that should stop the agent.
- Codex Execution Contract names the current working directory, applicable `AGENTS.md`, resumable `.goals/*.md` artifact, `git status --short`, `git diff --check`, and a project-specific verification command.
- Subagent Capacity Prerequisite requires `~/.codex/config.toml`, `[agents]`, `max_threads = 2147483647`, `max_depth = 2147483647`, `codex --strict-config doctor --summary --ascii`, loaded config verification, and no delete/reorder rule.
- Subagent Capacity Prerequisite includes `codex.cmd --strict-config doctor --summary --ascii` as the Windows fallback when `codex` is blocked by execution policy or shim handling.
- Subagent Dispatch Decision chooses L0/L1/L2/L3 from actual task shape, preserving L0 for simple work, requiring minimum 2 subagents for L2, minimum 4 subagents for L3, and pausing instead of letting the main session replace missing subagents.
- Subagent Execution Liberation states that the main session is scheduler, merger, boundary judge, and final verification owner while delegating isolated executable work.
- Do not silently reduce scope. If constraints require a smaller first step, keep the full request in `Original Request`, put the reduction in `Pause` or assumptions needing confirmation, and do not present the reduced scope as the final goal.
- When used with planning, TDD, verification, or superpowers workflows, the saved goal is the higher-level contract: later skills may decompose execution, but must not weaken `Objective`, `Non-Negotiables`, `Success Criteria`, or `Verification`.
- Full-spec goals include `Subagent Capacity Prerequisite`, `Subagent Dispatch Decision`, `Subagent Execution Liberation`, `Multi-Agent Collaboration`, `Dispatch Matrix`, `Shared File Ownership`, `Subagent Result`, `Merge Policy`, and `Rejection Conditions` by default.
- The dispatch protocol uses L0/L1/L2/L3 precision: simple work stays L0, complex work delegates real file-owned slices, and unsplittable work goes in `Pause` instead of inventing fake parallel work.
- Subagents must not be used only for simple reading, summarizing, searching, formatting, or one tiny function.

### Dispatch protocol default

Use this by default for `full-spec` goals.

The future executing main session must first choose L0/L1/L2/L3 before implementation. It should delegate executable work aggressively when useful, but simple work must stay L0 instead of spawning useless subagents.

Include these sections in the saved full-spec goal:

```markdown
## Codex Execution Contract

- CWD: current project root.
- Instruction source: read and obey applicable `AGENTS.md`.
- Resumable artifact: this `.goals/<file>.md`.
- Final checks: `git status --short`, `git diff --check`, and the smallest project-specific test or lint command.
- Chat output: report only saved path, launcher, verification result, and material risks.

## Subagent Capacity Prerequisite

- Before using full-spec / subagent-first goals, check the current user's `~/.codex/config.toml`.
- If `[agents]` exists, update only `max_threads = 2147483647` and `max_depth = 2147483647`; if it does not exist, append it.
- Do not delete or reorder existing config.
- Run `codex --strict-config doctor --summary --ascii`.
- On Windows, if `codex` is blocked by PowerShell execution policy or shim handling, run `codex.cmd --strict-config doctor --summary --ascii`.
- Continue only after the output shows `Configuration` / `config` is `loaded`.

## Subagent Dispatch Decision

- L0: no subagent. Use for single-file small edits, explanations, simple commands, and wording tweaks.
- L1: one subagent. Use for isolated reading, local verification, or risk review.
- L2: minimum 2 subagents. Use for cross-module work, code plus tests, or implementation and verification that can proceed in parallel.
- L3: minimum 4 subagents. Use for multi-module migrations, broad refactors, investigations, and batch fixes.
- Keep simple goals at L0. Do not create fake subagents for formality.
- If L2/L3 is selected and the session cannot create the required subagents, pause; the main session must not continue as the substitute executor.

## Subagent Execution Liberation

- Main session runs as scheduler, merger, boundary judge, and final verification owner.
- Delegate execution work that can be isolated, verified, and handed off.
- Subagents own context reading, implementation, local tests, local fixes, and risk reporting inside their file range.
- Main session keeps goal freezing, file-boundary decisions, merge decisions, and final verification.

## Multi-Agent Collaboration

- Main session freezes the original goal, success criteria, non-negotiables, shared interfaces, and file boundaries before dispatch.
- Main session selects L0/L1/L2/L3 before implementation and dispatches only useful slices.
- Main session owns scheduling, shared files, conflict handling, merge decisions, and final project-level verification.
- Subagents may decompose implementation but must not weaken the saved goal contract.

## Dispatch Matrix

| Slice | Agent Role | Goal | Allowed Files | Forbidden Files | Inputs | Required Output | Verify | Depends On | Merge Owner |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| contract/interface | subagent | ... | ... | ... | ... | changed files, verification command, result, risk, handoff | ... | none | main session |
| implementation | subagent | ... | ... | ... | ... | changed files, verification command, result, risk, handoff | ... | contract/interface | main session |

Allowed slice types: `contract/interface`, `implementation`, `verification`, `docs/handoff`, `config/tooling`.

## Shared File Ownership

Main-session-owned:
- shared interfaces, global docs, release notes, and files edited by multiple slices

Subagent-owned:
- files explicitly listed in that row's `Allowed Files`

## Subagent Result

Slice:
Status: adopted | needs-main-merge | blocked | rejected
Changed Files:
Verification Run:
Verification Result:
Boundary Crossings:
Risks:
Handoff:

## Merge Policy

- Dispatch independent slices first.
- Merge subagent results serially.
- Shared files are edited by the main session unless the slice explicitly allows them.
- Do not bypass completed subagent work and reimplement it silently; adopt it, adapt it with explanation, or reject it with cause.
- On conflict, narrow boundaries or move the shared-file work back to the main session.
- Consume each `Subagent Result` before merging, then run final project-level tests/builds after all adopted work is merged.

## Rejection Conditions

Reject or return a subagent result if it:
- Changes files outside its allowed range
- Omits verification results
- Reports failed verification without a risk explanation
- Weakens the original goal, success criteria, or non-negotiables
- Conflicts with frozen interfaces or design direction
- Cannot be understood or merged from the handoff

Codex subagent capacity setup

This is a required setup step before relying on full-spec / subagent-first goals. Keep it in the saved full-spec contract as a prerequisite, and create a standalone setup goal when the user asks to configure Codex subagents, raise limits, tune concurrency, or prepare another machine.

The saved goal must require:

1. Modify the current user's Codex global config file: ~/.codex/config.toml.
2. If [agents] exists, update only these fields inside it; if not, append:

[agents]
max_threads = 2147483647
max_depth = 2147483647

3. Do not delete or reorder existing config.
4. Run:

codex --strict-config doctor --summary --ascii

5. On Windows, if codex is blocked by execution policy or shim handling, run:

codex.cmd --strict-config doctor --summary --ascii

6. Verify the output shows Configuration / config is loaded.

compact

Use only when the user asks for a small inline goal or the task is trivial.

Keep normal compact rendered blocks under 12 lines. For compact mode, prioritize a direct copyable command over a saved full-spec file unless the user asks to save it.

review

Use when the user provides an existing goal and asks to review, inspect, compare, improve, lint, or fix it.

Check for:

• Missing original request, non-negotiables, success criteria, verification, scope, stop, or pause.
• Silent scope reduction or weakened acceptance criteria.
• Over-broad boundaries or missing no-touch areas.
• Missing multi-agent sections in a full-spec goal.
• English headings or labels inside a non-English goal.
• Long launcher, placeholders, vague verification, or fake parallel work.

Return findings first, then a corrected goal only if the user asks for one or the fix is small.

Canonical Fields

Use one canonical goal, then render it into requested formats. Canonical field names are internal; visible labels must follow the target language.

Canonical fields:

• title: short noun phrase
• original_request: the user's requested outcome/design intent, not narrowed
• objective: one concrete outcome that preserves the original request
• non_negotiables: user-provided acceptance criteria, design requirements, no-touch boundaries, and must-keep behaviors
• verify: 2-5 runnable checks, artifacts, screenshots, or review evidence
• boundaries: 2-6 write scopes and explicit no-touch areas
• stop: one completion condition
• pause: one human/external blocker condition
• notes: optional assumptions, only when useful

For full-spec, include enough detail that a new agent can execute from the saved file alone: execution plan, verification, safety/constraints, iteration policy, stop, and pause. For high-risk work such as money, trading, security, production data, destructive operations, or credentials, make safety explicit.

Renderers

Renderer snippets are English defaults. Translate visible labels for non-English targets before returning or saving output.

codex

Compact form:

/goal <objective>

Verify:
- ...

Boundaries:
- ...

Stop:
- ...

Pause:
- ...

Full-spec launcher:

/goal Execute only `.goals/<file>.md`.

claude

Objective: <objective>

Acceptance checks:
- ...

Constraints:
- ...

Done when:
- ...

Pause if:
- ...

gemini

Goal: <objective>

Success criteria:
- ...

Constraints:
- ...

Stop when:
- ...

Ask before continuing if:
- ...

cursor

Task: <objective>

Implement only:
- ...

Verify with:
- ...

Do not change:
- ...

Stop when:
- ...

github

## Objective

<objective>

## Acceptance Criteria

- ...

## Scope

- ...

## Out of Scope / Pause Conditions

- ...

markdown

Use the saved file as the Markdown format. In full-spec mode, include frontmatter, Short Command, and the complete execution spec.

Saving

Use scripts/save_goal.py from this skill directory when available:

python scripts/save_goal.py --title "<title>" --mode full-spec --format codex --format markdown < goal.txt

Resolve scripts/save_goal.py relative to this SKILL.md file, but run it from the user's current project directory so .goals/ is created in that project.

The script writes to .goals/YYYY-MM-DD-slug.md in the current working directory. It creates .goals/ if needed and avoids overwriting by appending -2, -3, etc.

If the script is unavailable, create the same file manually.

After saving a full-spec goal, validate it with scripts/lint_goal_file.py when available:

python scripts/lint_goal_file.py .goals/<file>.md

Use scripts/check_eval_cases.py after editing examples/evals/cases.json.

Use scripts/install_local.py to install or update this skill in the current user's Codex skills directory:

python scripts/install_local.py

Response

After saving, return only:

Saved: .goals/<file>.md

```text
<primary rendered block>
```

Add one short note only if a risky assumption was made.

Do not continue into implementation. The user must paste or invoke the returned launcher when they want execution to begin.