Start Reports

Before the Agent writes pages in bulk, it shows you a work-start report: what it found in the sources, which articles it plans, and how they will be organized. This is your cheapest moment to steer. Changing a plan costs a sentence; changing thirty written pages costs a revision round.

flowchart LR
  talk["Describe the goal<br/>and sources"] --> inv["Agent registers sources<br/>and investigates lightly"]
  inv --> rep["Work-start report<br/>.tmp/work-start-report.md"]
  rep -->|change scope| inv
  rep -->|confirm| write["Bulk writing"]
  write --> review["Knowledge Review"]

How you get there

Start by talking freely: what the knowledge should cover, where it comes from (repositories, modules, document links) and what form you expect — a package for Agents, a site, or both. The Agent researches representative material and asks only about gaps that would change the scope or the output. It does not run a fixed questionnaire, and fully managed mode does not authorize guessing your answers.

For a first build, the Agent registers and captures the confirmed sources, takes a code skeleton and document outlines, and only then writes the report. It does not deep-read every implementation first; broad scope calls for a horizontal survey.

What the report contains

The report is written in your language, under .tmp/, and adapts to the task. Typical sections:

Section	What to look for
Readers, goal and current delivery	Who uses the knowledge, the whole outcome, and what this first batch delivers
What the materials support	Checked topics versus provisional ones, and what is still uninvestigated
Articles and batches	Each reader topic, new or existing article, supporting sources, and its depth: an entry point, an explanation or a procedure
Reading organization	Proposed navigation; for a new site, a sketch of top-level groups, a representative menu and page outlines
Indexing guidance	Which Indexer Skills are used on which sources
Decisions and gaps	Confirmed choices versus recommendations, conflicts, exclusions and unchecked coverage

Material counts, article counts and investigated scope are kept separate, and unknown totals are never shown as zero. A first batch is not presented as the whole goal.

Answering it

Reply in plain language: confirm, or say what to change. Requested changes are made before the Agent records your confirmation.

You are asked when the scope is complex and touches more than five articles, or when a separate decision is open — most often a top-level navigation change, which always shows a before/after tree and needs its own explicit answer. When a task exceeds 30 distinct images without a policy, the report also asks how to handle them. Small bounded tasks — a few documents, one module, a revision — get a short proposal instead, and a specific earlier request can already cover it.

Silence, elapsed time, questions and objections are never approval. Approval of an unchanged plan carries across batches and retries; material changes or scope expansion bring the plan back to you.

Large projects

For work spanning more than 30 source documents or substantive investigation of two or more repositories, the context-plan Skill first produces a project PLAN that splits the work into stages. Each stage then gets its own work-start report within the approved plan. See Large Projects.

开工报告

Agent 批量写页面之前，会先给你一份开工报告：它在来源中发现了什么、计划写哪些文章、这些文章将如何组织。这是你调整方向成本最低的时机。改计划只需一句话；改三十篇已写好的页面则要多走一轮修订。

flowchart LR
  talk["描述目标<br/>和来源"] --> inv["Agent 登记来源<br/>并做初步调研"]
  inv --> rep["开工报告<br/>.tmp/work-start-report.md"]
  rep -->|调整范围| inv
  rep -->|确认| write["批量写作"]
  write --> review["知识审核"]

如何走到这一步

先自由描述：知识要覆盖什么、来自哪里（仓库、模块、文档链接），以及你期望的形式——给 Agent 用的知识包、站点，或两者都要。Agent 会调研有代表性的材料，只就会改变范围或产出的缺口向你提问。它不会走一套固定问卷，全托管模式也不意味着可以替你猜答案。

首次构建时，Agent 会登记并采集已确认的来源，提取代码骨架和文档大纲，之后才撰写报告。它不会先把每处实现都深读一遍；范围较广时应先做横向摸底。

报告包含什么

报告使用你的语言撰写，存放在 .tmp/ 下，并随任务调整内容。常见章节如下：

章节	关注点
读者、目标与本次交付	谁使用这些知识、完整的目标成果，以及第一批交付什么
材料能支撑什么	已核实的主题与暂定主题，以及尚未调研的部分
文章与批次	每个读者主题、新建还是已有文章、支撑来源，以及深度：入口页、讲解还是操作步骤
阅读组织	建议的导航；对于新站点，给出顶层分组草图、一个代表性菜单和页面大纲
索引指引	哪些 Indexer Skill 用于哪些来源
决策与缺口	已确认的选择与建议、冲突、排除项以及未核查的覆盖范围

材料数、文章数和已调研范围分开统计，未知的总数绝不显示为零。第一批交付也不会被当作整体目标呈现。

如何回复

用自然语言回复即可：确认，或者说明要改什么。你提出的修改会在 Agent 记录你的确认之前完成。

当范围复杂且涉及五篇以上文章，或有单独的决策待定时，Agent 会向你提问——最常见的是顶层导航变更，它总会展示变更前后的目录树，并需要你单独明确答复。如果任务涉及超过 30 张不同的图片且没有既定策略，报告也会询问如何处理。小而明确的任务——几篇文档、一个模块、一次修订——只会给出简短方案，而你之前提出的具体请求可能已经涵盖了它。

沉默、时间流逝、提问和异议都不算批准。对未变更计划的批准会延续到后续批次和重试；一旦计划有实质变化或范围扩大，就会重新交给你确认。

大型项目

如果工作涉及超过 30 份来源文档，或需要对两个及以上仓库做实质性调研，context-plan Skill 会先产出一份项目 PLAN，把工作拆成若干阶段。之后每个阶段在已批准的计划内各自产出开工报告。参见大型项目。