Truth-grounded execution protocol for AI agents.
Really is a self-contained agent skill for careful, evidence-based work. It helps an AI agent avoid guessing, overclaiming, unsafe side effects, and polished but unsupported answers.
Facts before conclusions. Evidence before confidence. Verification before completion.
中文名：真实模式 / 求真模式 / 严格开发模式 / 可信执行模式
真查、真想、真验证、真说明限制。
Really turns a broad agent task into a disciplined execution loop:
- Lock the user's actual intent.
- Choose the right evidence depth.
- Inspect real sources before claiming facts.
- Separate facts, inferences, unknowns, and recommendations.
- Execute only within scope.
- Verify before claiming completion.
- Report limits clearly.
It is especially useful when the task touches code, configuration, debugging, product behavior, UI, architecture, files, data, releases, credentials, deletion, or other high-risk state.
Many AI agent failures are not syntax failures. They are execution-discipline failures:
- explaining source code without reading it;
- saying a bug is fixed without running a relevant check;
- editing files when the user only asked for analysis;
- treating a likely guess as a verified fact;
- pushing through risky actions without confirming scope;
- hiding uncertainty behind confident wording.
Really makes those failure modes explicit and gives the agent rules for avoiding them.
| Feature | Purpose |
|---|---|
| Intent Lock | Classifies whether the user wants an answer, analysis, plan, scan, execution, verification, or artifact delivery. |
| Evidence Budget | Uses Light, Standard, or Strict evidence depth depending on task risk. |
| Truth Discipline | Separates fact, inference, unknown, and recommendation. |
| Source Audit | Requires inspection before source changes and review after source changes. |
| Verification Gates | Matches checks to code, UI, state, data, files, and artifacts. |
| Stop Conditions | Stops when evidence, permission, or scope is insufficient. |
| Concise Reporting | Reports conclusions, evidence, verified areas, unverified risks, and next steps. |
Without Really:
User: This feature is broken. Fix it.
Agent: I found the issue and fixed it.
Common problem: the agent may not have inspected the real source, reproduced the issue, or verified the fix.
With Really:
Intent lock: Execute allowed.
Evidence budget: Standard.
I will inspect the relevant files and runtime path first, then make a scoped change and run the closest available verification.
Verified:
- The failing path is in src/...
- The fix passes npm test ...
Unverified:
- Browser behavior was not checked because no local dev server is configured.
The difference is not verbosity. The difference is that claims are tied to evidence.
Use Really when:
- the task involves source code, configuration, files, logs, UI, data, or system state;
- the user asks for careful or evidence-based work;
- the user asks the agent not to guess, appease, or jump to conclusions;
- the task has meaningful risk, such as releases, credentials, deletion, billing, migrations, or repeated failures.
Do not use it for:
- pure casual conversation;
- quick creative ideation;
- very small answers where the user explicitly wants only a brief opinion.
Copy SKILL.md into the skill directory used by your agent environment.
For Codex-style skill systems, keep the file as:
really/SKILL.md
Then trigger it by asking for careful, evidence-grounded execution, or by naming the skill directly if your environment supports named skills.
- SKILL.md: the full protocol
- launch-posts.md: launch copy for GitHub, Hacker News, Product Hunt, X, V2EX, and Chinese developer communities
- LICENSE: MIT license
Really 是一套面向 AI Agent 的可信执行协议。
它的核心目标是让 Agent 在代码、配置、调试、发布、删除、迁移等任务中，先确认用户意图，再检查真实来源，区分事实、推断和未知，执行后说明验证情况。
它不是让 Agent 变慢，而是让 Agent 的结论有依据、动作有边界、完成声明可验证。
MIT