项目案例 — Portfolio

01 · Engineering judgment

独立研究Independent research

先看判断，再看案例。Judgments first, then cases.

下面三条是我反复使用的工作判断；两个主案例和后面的方法研究，都是这些判断的落点。The three judgments below are ones I reuse constantly; both flagship cases and the method study that follows are where they land.

AUT-01 · Agent 的判别标准AUT-01 · What makes an agent

不是调了 LLM，就算 Agent。Calling an LLM does not make an agent.

可调用的 tools、observe-act 循环、由模型自主判断的终止条件，三者缺一不可。单次生成是 LLM 应用，固定流程 DAG 是 workflow——框架或调用方式本身不决定边界。Callable tools, an observe–act loop, and a termination condition the model judges for itself: all three are required. A single generation is an LLM application, a fixed DAG is a workflow — neither the framework nor the calling pattern decides the boundary.

为什么重要：Why it matters: 先分清对象，才能决定该把哪一层做成 Agent、哪一层做成确定性的流程。Only once the object is clear can you decide which layer becomes an agent and which stays a deterministic process.

AUT-02 · Agent 与 Harness 的分工AUT-02 · Division of labor between agent and harness

Agent 决定做什么，
Harness 决定在什么约束下做。The agent decides what to do,
the harness decides under which constraints.

工具协议、循环边界、终止判断、状态恢复与审计追溯，从业务逻辑里剥出来，成为独立基础设施，让 Agent 行为可观测、可约束、可回归。Tool protocols, loop boundaries, termination rules, state recovery, and audit trails are peeled out of business logic into independent infrastructure, so agent behavior becomes observable, constraint-bound, and regression-testable.

为什么重要：Why it matters: 两个案例里，Agent 负责判断与执行，边界与审计由工程约束承接，而不是埋进提示词。In both cases the agent judges and executes, while boundaries and audits are carried by engineering constraints rather than buried in prompts.

AUT-03 · 证据优先的公约数AUT-03 · The common factor: evidence first

结论必须能回到证据，
不能证明的就保留状态。Conclusions must return to evidence;
what cannot be proven keeps its status.

证据优先、显式拒答或诊断状态、人在回路、可重复生成——这些是贯穿案例的公约数：不把不确定性伪装成事实，只把可验证的结论发布出去。Evidence first, explicit refusal or diagnostic status, human in the loop, reproducible generation — these are the common factors across the cases: never dress uncertainty up as fact, and publish only verifiable conclusions.

为什么重要：Why it matters: RAG 的拒答状态、SQL 血缘的诊断状态与人工审核，都是同一判断的落点。The RAG refusal status, the SQL lineage diagnostic status, and human review all land on the same judgment.

02 · Representative work

近一年的两个主力项目Two flagship projects from the past year

下面两条实证，分别把非结构化知识与结构化 SQL 收束为可核验、可审计的结论。The two cases below reduce unstructured knowledge and structured SQL into verifiable, auditable conclusions.

01 匿名化项目Anonymized project

Knowledge assets before retrieval

面向专业材料的
RAG 知识构建RAG knowledge building
for professional material

在 embedding 之前，以受控分类、主题归整、来源回链和入库门禁，把 PPT/PDF 页面转化为高密度知识单元；弱证据显式拒答，回答回到来源页。Before embedding, PPT/PDF pages become high-density knowledge units through controlled taxonomy, topic organization, source back-links, and ingestion gates; weak evidence triggers an explicit refusal, and every answer points back to its source page.

问题Problem: 页面切片碎片化，主题与来源脱节Fragmented page slices, topics detached from sources
方法Method: 类别级知识资产与分层索引Category-level knowledge assets and a layered index
证据Evidence: 来源页引用、合同校验与拒答状态Source-page citations, contract validation, and refusal status

02 匿名化项目Anonymized project

Evidence before lineage claims

项目 SQL 动态节点搜索
与字段级血缘治理Dynamic node search over project SQL
and field-level lineage governance

从目标节点出发，在项目 SQL 快照中自动搜索相关节点与路径，再在已证实的 scope 内追溯字段来源；无法唯一证明的关系保留为诊断状态，交给人工审核。Starting from a target node, related nodes and paths are searched automatically inside a project SQL snapshot, then field origins are traced within the proven scope; relations that cannot be uniquely proven stay in diagnostic status for human review.

问题Problem: 全局名称猜测会制造虚假字段来源Global name guessing fabricates field origins
方法Method: 节点图、目标驱动搜索与字段 DFSNode graph, target-driven search, and field DFS
证据Evidence: SQL 位置、诊断状态与人工审核SQL locations, diagnostic status, and human review

03 · Agent engineering study

独立研究Independent research

判断如何变成工具：
Plugins、Skills 与可验证编排How judgments become tools:
Plugins, Skills, and verifiable orchestration

上面的判断要落地，需要工具层承接。模型提供通用推理能力，但工程系统需要回答更具体的问题：什么时候触发哪套流程、主 Agent 把什么交给 Subagents、哪些工具可以使用、什么动作需要批准，以及一次修改有没有破坏原有能力。For the judgments above to land, a tool layer has to carry them. Models provide general reasoning, but an engineering system must answer sharper questions: which workflow fires when, what the main agent hands to subagents, which tools may be used, which actions need approval, and whether one change broke an existing capability.

为什么从 Skills 开始Why start with Skills

先把工作方法写成合同，
再把能力打包成产品。Write the working method as a contract first,
then package capability as a product.

Skill 定义任务边界、输入输出、步骤、资源、失败语义和验证方式；Plugin 再把一个或多个 Skills、连接器与工具依赖组织成可安装、可分发的能力包。这样，业务方法不会被埋进某个模型、提示词或客户端代码里。A Skill defines task boundaries, inputs and outputs, steps, resources, failure semantics, and how to verify; a Plugin then organizes one or more Skills, connectors, and tool dependencies into an installable, distributable package. That way the working method is never buried in a model, a prompt, or client code.

Skill工作流合同Workflow contract

说明何时触发、怎样执行、何时拒绝，以及输出必须满足什么。States when to trigger, how to execute, when to refuse, and what the output must satisfy.

Plugin能力分发Capability distribution

将 Skills、连接器、MCP 工具和呈现资产组合为可安装单元。Combines Skills, connectors, MCP tools, and presentation assets into an installable unit.

Harness执行系统Execution system

管理上下文、工具调用、沙箱、审批、进度、失败和跨轮状态。Manages context, tool calls, sandboxing, approvals, progress, failure, and state across turns.

Golden Evals发布证据Release evidence

用固定案例、评分器和人工校准证明新版本没有关键回归。Proves with fixed cases, scorers, and human calibration that a new version has no critical regressions.

Orchestration

Skill 如何编排主 Agent 与 SubagentsHow a Skill orchestrates the main agent and subagents

主 Agent 保留需求、合同和最终决策；Subagents 只接收边界清晰、可以独立验证的工作包。Skill 规定什么时候允许并行、子任务应返回什么，以及发生共享写入或合同冲突时如何降级为串行。The main agent keeps requirements, the contract, and the final decision; subagents receive only well-bounded work packages that can be verified independently. The Skill states when parallelism is allowed, what each subtask must return, and how to degrade to serial execution on shared writes or contract conflicts.

Main Agent 理解目标并冻结合同Understand the goal and freeze the contract

确认输入、输出、公开边界、依赖和成功标准。Confirm inputs, outputs, public boundary, dependencies, and success criteria.

Subagent A代码与调用链探索Code and call-chain exploration

Subagent B测试与失败样本检查Tests and failing-sample inspection

Subagent C文档、来源与边界核对Docs, sources, and boundary checks

Main Agent 解决冲突并集成验证Resolve conflicts and integrate verification

只接收证据摘要；共享合同、写入和最终判断回到主线程。Accepts only evidence summaries; shared contracts, writes, and final judgment return to the main thread.

适合并行：Good for parallel:只读探索、测试发现、日志分析、资料归纳。 read-only exploration, test discovery, log analysis, and material synthesis.必须串行：Must stay serial:共享合同、同一文件、线上状态、审批动作和最终发布。 shared contracts, the same file, live state, approval actions, and final release.

Evaluation

用黄金测试集验证 Skill，而不是凭感觉验收Validate Skills with a golden test set, not by feel

黄金集不只保存“正确答案”，还要冻结触发条件、不得触发的反例、预期工具轨迹、审批点、状态语义和输出合同。A golden set stores more than the right answer: it freezes trigger conditions, negative cases that must not fire, expected tool traces, approval points, status semantics, and the output contract.

01

建立案例集Build the case set

覆盖标准路径、边界、歧义、对抗输入、工具失败和应拒绝场景。Cover standard paths, boundaries, ambiguity, adversarial inputs, tool failures, and cases that must be refused.

02

在 Harness 中回放Replay in the harness

固定 Skill 版本、模型、工具、沙箱和输入，记录完整执行轨迹。Pin the Skill version, model, tools, sandbox, and inputs, and record the full execution trace.

03

分层评分Score in layers

确定性断言检查合同；规则或模型评分检查质量；关键样本由人工校准。Deterministic assertions check the contract; rule- or model-based scoring checks quality; critical samples are calibrated by humans.

04

设置发布门禁Set the release gate

比较基线版本；关键案例不得回归，新发现的失败追加到黄金集。Compare against the baseline; critical cases must not regress, and newly found failures are added to the golden set.

触发准确率Trigger accuracy合同满足率Contract fulfillment工具与审批轨迹Tool and approval traces任务正确性Task correctness拒绝正确率Refusal accuracy成本与时延Cost and latency

Examples

几个可以说明设计边界的 SkillsA few Skills that show the design boundary

下面列的是工作流样本，不代表公开其私有指令、凭据、会话内容或运行数据。What follows are workflow samples; they do not publish private instructions, credentials, session content, or run data.

openai-docs

受来源约束的技术研究Source-constrained technical research

先检索并打开官方资料，再回答产品、API 与 Codex 问题；来源范围和引用要求本身就是 Skill 合同。Retrieve and open official material first, then answer product, API, and Codex questions; the source scope and citation requirement are themselves part of the Skill contract.

研究路由 · 来源门禁 · 引用验证Research routing · source gating · citation checks

skill-creator

把稳定工作流变成 SkillTurn a stable workflow into a Skill

从目标、触发条件、输入输出和失败语义出发，生成最小的 SKILL.md 与必要资源。Starting from goals, triggers, inputs and outputs, and failure semantics, generate the minimal SKILL.md and the resources it needs.

需求冻结 · 渐进披露 · 结构校验Requirement freeze · progressive disclosure · structure checks

plugin-creator

把能力打包为 PluginPackage capability as a Plugin

组织清单、Skills 与可选依赖，验证安装和分发边界；Plugin 是包装层，不替代工作流合同。Organize the manifest, Skills, and optional dependencies, and verify install and distribution boundaries; a Plugin is a wrapper, not a replacement for the workflow contract.

Manifest · 能力清单 · 分发验证Manifest · capability list · distribution checks

shared-chrome-session

已登录浏览器的协作控制Coordinated control of a logged-in browser

以标签页租约、刷新检查和人工确认约束共享会话，避免多个 Agent 争用状态或越过外部副作用边界。Constrain a shared session with tab leases, refresh checks, and human confirmation, so multiple agents do not contend for state or cross external side-effect boundaries.

会话租约 · 审计事件 · 人工批准Session leases · audit events · human approval