搭建、重构和排查 ChatGPT Apps SDK 应用,把 MCP 服务器与 widget 界面组合起来。
智能体与编排
define-goal
试用在动手之前,把模糊的想法整理成有明确验证证据和范围边界的可量化目标。
它能做什么
把用户想做的事收敛成一条目标描述,写清完成时会成立的具体结果、涉及的系统或产物、如何验证完成,以及哪些内容不在范围内。创建前会先调用 get_goal 查看是否已有活跃目标,避免重复创建或与现有目标冲突。像「继续推进」「再查一查」这类只描述动作的目标会被拒绝或改写;只有当缺失信息会影响最终结果或验证方式时,才会追问一个简短问题。
什么时候用它
- 把「让结账快一点」改写成带基准测试和运行次数的延迟指标
- 为缺陷修复设定目标,先复现、再修复、最后跑通验证命令
- 新建目标前确认已有目标是否已经覆盖当前需求
- 在开始调研或运维工作前先定下验收标准
技能文档
Define Goal
Overview
Shape the user's intent into an objective an agent can pursue honestly. Prefer measurable outcomes, explicit evidence, and bounded scope over activity descriptions.
This skill covers goal definition and goal-tool creation only. Do not create intermediate planning artifacts, durable snapshots, ledgers, decision logs, or resume files from this skill.
Workflow
-
Confirm that goal definition is actually needed.
- Use this skill when the user asks for
$define-goal, asks to create or set a goal, asks for the goal tool, or wants help turning an intention into a clear objective. - If the user only asks for ordinary implementation work, do the work directly instead of forcing goal creation.
- Use this skill when the user asks for
-
Restate the likely goal in concrete terms. A usable goal names:
- the specific outcome that will be true
- the main artifact, system, repo, environment, or user-facing behavior involved
- how completion will be verified
- what is in scope
- what is out of scope when ambiguity would matter
- the stop condition for asking the user instead of grinding
-
Make it quantitative when the domain supports it. Prefer numbers that represent real success, not decorative precision:
- pass/fail validators: exact tests, checks, CI jobs, evals, commands, or acceptance criteria
- quality thresholds: latency, error rate, cost, accuracy, recall, precision, coverage, flake rate, bundle size, memory, uptime, completion rate, or manual review criteria
- artifact constraints: file paths, affected modules, allowed commands, output formats, target environments, deadlines, or maximum blast radius
- evidence counts: number of reproduced failures, successful reruns, reviewed examples, migrated records, addressed comments, or verified cases
-
Repair weak goals before setting them.
- Rewrite vague goals into measurable objectives when local context makes the rewrite safe.
- Ask one concise clarification question when the missing detail changes the intended outcome or validation.
- Reject pure activity goals such as "make progress," "keep investigating," "improve things," or "work on X" unless they are sharpened into a verifiable outcome.
-
Check active goal state before creating a goal.
- Call
get_goal. - If there is no active goal and the objective meets the quality bar, call
create_goal. - If there is an active goal that still matches the user's intent, continue using it instead of creating a duplicate.
- If there is an active goal that conflicts with the new request, ask whether to finish the current goal, mark it complete if done, or start a separate goal-backed thread.
- Call
-
Create the goal only after it passes the quality bar.
- Use a single concise objective string.
- Include the verification evidence in the objective itself.
- Include scope bounds when they constrain the work.
- Include a token budget only when the user explicitly requested one.
- Do not call
create_goalfor an ordinary multi-step task unless the user explicitly asked for goal-backed work.
Goal Quality Bar
Before create_goal, the objective should answer:
- What concrete thing will be true when this is done?
- What evidence will prove it?
- What quantitative or binary threshold defines success?
- What scope boundaries matter?
- What should cause the agent to stop and ask?
Good:
Reduce checkout API p95 latency below 250 ms for the documented slow path by making the smallest safe server-side change, then verify with
npm run test:checkoutand the existing local latency benchmark showing p95 under 250 ms across 3 consecutive runs.
Good:
Resolve the open review comments on PR 123 that request code changes, update only the affected auth files and tests, and verify with the targeted auth test command plus
gh pr view 123showing no unresolved change-request threads.
Weak:
Make checkout faster.
Weak:
Keep investigating the PR comments.
Quantification Heuristics
- For bugs, define success as reproduction first, fix second, and a failing-then-passing validator when possible.
- For tests, name the exact command and required pass condition.
- For performance, name the metric, target threshold, measurement method, and number of runs.
- For quality work, define an observable acceptance bar such as reviewed examples, lint/typecheck/test pass, or user-approved artifact.
- For research, define the decision the research must enable, the sources or systems in scope, and the evidence standard.
- For operations, define healthy state, monitoring window, failure threshold, and rollback or escalation trigger.
Clarifying Questions
Ask only when a reasonable rewrite would risk pursuing the wrong outcome. Keep the question short and oriented around the missing validator or scope boundary.
Useful question shapes:
- "What metric should define success here: latency, cost, accuracy, or user-visible behavior?"
- "Which environment should I verify against: local, staging, or production?"
- "What is the minimum evidence you want before I mark this goal complete?"
If the user cannot provide a metric, propose the most honest binary validator available and ask for confirmation.
常见问题
- 它会直接创建目标吗?
- 会。目标通过质量标准后,它会先调用 get_goal 检查是否已有活跃目标,再用一条简洁的目标描述调用 create_goal。如果现有目标仍符合用户意图,就不会重复创建。
- 什么样的目标才算达标?
- 需要说明完成时什么会成立、用什么证据证明、量化或二元的成功阈值、关键的范围边界,以及什么情况下应当停下来询问。只描述动作的目标会被拒绝或改写。
- 会向我提问吗?
- 只在合理改写仍可能做错方向时提问,比如缺少验证方式或范围边界不清。如果用户给不出指标,它会提出最可靠的二元验证方式并请用户确认。
- 它会生成计划文档或跟踪进度吗?
- 不会。它只负责目标定义和目标创建,明确不产出中间计划文档、持久快照、台账、决策日志或续跑文件。
OpenAI 的更多技能
浏览全部技能用文件里已发布的设计系统组件、变量和样式,在 Figma 中搭建或更新整页界面,而不是手动画基础图形。
按阶段从代码库生成 Figma 设计系统,每步都经你确认后再继续。
按 Figma Plugin API 的规则编写并执行 JavaScript,避开那些难以排查的报错。
把概念、品牌线索或参考图做成经过校验和打包的 Codex 动画宠物图集。
imagegen
官方在 Codex 中生成或编辑位图,把需求整理成结构化提示词,并把成品落到项目里。