代理目标
- 作者仓库星标 4,565
- 作者仓库 epicenter
- 领域
- 人工智能
- 作者声明的智能体
- 未找到明确声明;不据此推断已兼容或已测试
- 静态检查
- 88 / 100 · 启发式扫描,不代表运行安全
- 运行环境证据
- 未执行;当前没有隔离运行测试证据
- 作者 / 版本 / 许可证
- @EpicenterHQ · 未声明许可证
- 上下文用量评级
- 中等消耗
- 接入复杂程度
- 需简单配置
- 是否需要外部接口密钥
- 不需要
- 兼容的系统
- macOS · Linux · Windows
- 运行环境要求
- 未检出明确要求
- 检测到的文件与系统行为
-
- 只读
- 允许写入 / 修改
- 命令执行
- 网络行为
- 仅限本地
- 安装命令数
- 无(仅作为资料)
需要注意: 未声明工具权限范围;实际权限由运行环境决定。
# Agent Goal
Write one `/goal` line that a coding agent can keep pursuing across turns until it can prove the work is done.
The best goal is both a directive and a completion condition:
```txt
/goal [do the work] until [observable condition is true].
```
Treat the goal as a contract. It should tell the agent what to do, what evidence proves it, and where to start.
The highest-signal goal answers three questions:
```txt
What should change?
How will the agent prove it changed?
Where should the agent start?
```
## Core Shape
Use this structure unless the user needs a different format:
```txt
/goal Complete [single objective] in [lane]. First read [required context]. Work in checkpoints. After each checkpoint, surface evidence from [validation]. Continue until [verifiable end state].
```
Include only execution-critical details:
- Objective: one concrete outcome.
- Lane: files, package, app, branch, issue, spec, backlog label, or intended work area.
- Context: plans, docs, issue links, logs, screenshots, traces, commands, or acceptance criteria to inspect first.
- Evidence: command output, tests, build result, screenshot comparison, eval score, file count, clean git status, or reviewed artifact.
- Done condition: the exact state that means the goal is achieved.
## Rules
1. Start the answer with `/goal` when the user asks for the goal text.
2. Name one main objective. If the request contains a backlog, make the goal "finish this named queue/spec" and define what empty or complete means.
3. Make "done" observable. Prefer "`bun test packages/workspace` exits 0" over "tests pass"; prefer "all checked items in `PLAN.md` are complete" over "finish the plan."
4. Tell the agent to surface evidence in the transcript. Goal evaluators judge what the worker has shown, not private intent.
5. Put long requirements in a plan or spec, then point the goal at that file. Do not paste a huge spec into `/goal`.
6. Ask for checkpoints when the work spans multiple turns. Each checkpoint should produce a small status note: changed, verified, remaining, questions.
7. Bound runaway work with evidence. For example: "after 3 failed attempts on the same test, report the root cause and the next product or ownership question."
8. Ordinary focus should not stop grounded fixes.
9. Do not use `/goal` for vague wishes, unrelated chores, open-ended research, or work where the agent cannot produce evidence.
10. Keep the condition judgeable from the transcript. If a separate verifier read only the conversation after each turn, it should be able to tell whether the goal is met.
## Distilled Pattern
Think in this order:
```txt
Outcome
What must be true?
Evidence
What command, artifact, or visible behavior proves it?
Lane
What is the intended work area?
Method
What should the agent read first, and how should it checkpoint?
Start
Where should the agent begin?
```
Then compress that into one goal.
## Platform Notes
Codex:
- Write the goal as a durable objective attached to the active thread, with a verifiable stopping condition.
- Codex docs do not describe Claude's separate evaluator model. Do not assume Codex uses the same evaluation mechanism.
Claude Code:
- `/goal` sets a session-scoped completion condition.
- Claude uses a separate small model after each turn to decide whether the condition has been met.
- The evaluator does not run tools or read files independently.
Shared rule: the goal should not rely on hidden state. Tell the agent to run checks and surface evidence in the transcript.
## Verifier Test
Before finalizing the goal, imagine a checker can see only the transcript, not the filesystem.
Good evidence:
```txt
`bun test packages/auth` exited 0.
All checklist items in `specs/auth.md` are checked.
The final screenshot shows the empty state and no overlap at 390px and 1440px.
`git diff --name-only` only lists files under `apps/api`.
```
Weak evidence:
```txt
The implementation looks complete.
The agent believes the migration is done.
Most tests should pass.
The UI seems better.
```
If the evidence is weak, rewrite the goal until completion can be judged from command output, visible artifact checks, or an explicit final status.
## Templates
Plan execution:
```txt
/goal Implement `specs/[file].md` in checkpoints until every checklist item is complete, the review section is filled in, and `[final validation command]` exits 0. First read the spec and the files it names. After each checkpoint, update the checklist and surface the validation result. Continue unless the spec conflicts with current code or needs a product decision.
```
Failing tests:
```txt
/goal Fix the failing tests in `[lane]` until `[test command]` exits 0 and no unrelated speculative changes are present. First run the command and inspect the failures. Work from the owning boundary outward. After each fix, rerun the targeted test and report the result. Ask before deleting tests or weakening assertions.
```
Migration:
```txt
/goal Migrate `[old path or system]` to `[new path or system]` until all callers use the new path, parity checks pass, and `[final validation command]` exits 0. First read `[migration plan or docs]` and identify callers. Work in checkpoints with validation after each checkpoint. Ask before changing unrelated public APIs, or if compatibility, data migration, or rollback policy is ambiguous.
```
Prototype:
```txt
/goal Build a polished first version of `[app or feature]` inside `[lane]` until the primary flow works end to end, the app builds and runs, and `[visual or command validation]` confirms the expected behavior. First read `[plan or reference]`. Work in checkpoints and surface screenshots or command output as evidence. Ask if the data model or user flow is unclear.
```
Backlog or issue queue:
```txt
/goal Work through `[queue or label]` until every item is closed or has a documented reason it needs user input. First list the queue and choose the item with the clearest owner and validation path. For each item, make the correction at its owning boundary, run `[validation]`, and report the result before moving on. Stop when the queue is empty.
```
Eval or prompt loop:
```txt
/goal Improve `[prompt or system]` until `[eval command]` reaches `[target score]` or no further targeted improvement is justified. First run the eval and inspect failures. Make direct design corrections, rerun the eval after each change, and report score changes. Ask if improvement requires product or policy guidance.
```
## Bad To Good
Weak:
```txt
/goal Make the app better and fix bugs.
```
Strong:
```txt
/goal Fix the checkout regressions tracked in `issues/checkout.md` until every listed reproduction passes, `bun test apps/storefront` exits 0, and the final status names any intentionally deferred issues. Work only in `apps/storefront` and shared checkout packages. Ask before changing payment provider contracts or deleting tests.
```
Weak:
```txt
/goal Finish the migration.
```
Strong:
```txt
/goal Complete `specs/20260514T120000 auth-migration.md` until every checklist item is checked, all auth callers compile against the new service, `bun test packages/auth apps/api` exits 0, and `git diff` shows no unrelated speculative edits. Update the spec after each checkpoint. Ask if the old API has undocumented behavior that needs a compatibility decision.
```
## Final Check
Before handing back a goal, verify:
- It begins with `/goal`.
- It has one main objective.
- It names the evidence that proves completion.
- It tells the agent to surface that evidence.
- It names where to start.
- It tells the agent to continue until the evidence proves completion.
章节与文件索引
按说明文件中的顺序排列。
核心形状
规则
蒸馏模式
平台备注
验证者测试
文件/命令 -> /goal · bun test packages/workspace · PLAN.md · bun test packages/auth · specs/auth.md · git diff --name-only · apps/api · specs/[file].md
内容校验值 -> 931dc4eb21c6
当前原文的自动摘译,尚未人工审校。
核心形状
除非用户需要不同的格式,否则请使用此结构: 仅包含执行关键的细节: 目标:一个具体结果。 Lane:文件、包、应用程序、分支、问题、规范、待办事项标签或预期工作区域。上下文:计划、文档、问题链接、日志、屏幕截图、跟踪、命令或首先检查的验收标准。
规则
当用户询问目标文本时,以
/goal开始回答。说出一个主要目标。如果请求包含积压工作,则将目标定为“完成此指定队列/规范”并定义空或完整的含义。让“完成”变得可观察。优先选择“bun test packages/workspaceexits 0”而不是“测试通过”;更喜欢“PLAN.md中所有检查的项目都已完成”而不是“完成计划”。告诉特工在笔录中提供证据。目标评估者判断员工所表现出的内容,而不是个人意图。蒸馏模式
按以下顺序思考:然后将其压缩为一个目标。
平台备注
Codex:将目标编写为附加到活动线程的持久目标,并具有可验证的停止条件。 Codex 文档没有描述 Claude 的单独评估器模型。不要假设法典使用相同的评估机制。 Claude 代码:
/goal设置会话范围内的完成条件。验证者测试
在最终确定目标之前,想象一下检查器只能看到脚本,而不能看到文件系统。好的证据:弱证据:如果证据弱,重写目标,直到可以根据命令输出、可见工件检查或明确的最终状态判断完成。