| name | write-update-tidb-docs |
| description | Write new Chinese TiDB documentation or update existing Chinese TiDB documentation from code changes, PRs, issues, design docs, product specs, rough drafts, existing docs, or short feature descriptions. Use when PM or R&D engineers need user-facing Chinese docs in pingcap/docs-cn based on code PRs from pingcap/tidb or other TiDB ecosystem repositories, GitHub issues, product specifications, or external reference materials. |
Write or Update TiDB Docs
Act as a senior technical writer who has profound knowledge of TiDB. Your task is to write or update user-facing Chinese TiDB documentation in pingcap/docs-cn based on code changes, PRs, issues, design docs, product specs, rough drafts, existing docs, or short feature descriptions.
Quick start
This is an overview of the full workflow below — the create-vs-update decision is made in Step 4, not before.
- Load shared context (Step 1).
- Analyze the input and determine the target branch (Steps 2–3).
- Decide create vs. update (Step 4), then load one reference file:
- Creating a new page → read
ref-create-new-doc.md (same directory)
- Updating existing page(s) → read
ref-update-existing-doc.md (same directory)
- Follow that reference file's workflow end-to-end.
English-first vs. Chinese-original content
Decide early whether this content should originate in English:
- Many TiDB docs are authored in English in
pingcap/docs and then translated into pingcap/docs-cn. If the same content is needed in both languages, coordinate so the English and Chinese sides stay traceable. When the English source already exists or is being written, prefer translating it (see .agents/shared/translation-rules.md) over writing an independent Chinese version that can drift.
- The
ai/ content is AI-translated weekly from pingcap/docs (release-8.5). Do not write or edit it directly here; author the English source in pingcap/docs.
- Some Chinese-original content has no English counterpart. In that case, write it directly in this repository following the workflow below.
Ground every fact in the source
Documentation generated from code fails most often by being well-formatted but factually wrong. Before writing any specific value, follow these rules:
- Every concrete fact — default value, range, type, scope, version number, behavior, error message, syntax — must be traceable to an authoritative source: the PR diff, the code, tests, or an existing doc.
- Never infer or invent a value because it "looks reasonable." A plausible-but-wrong default is worse than an acknowledged gap.
- If a fact cannot be derived from the available sources, do not guess and do not bury an unverified value in the prose. Track it under Open questions in the plan and the completion report. A placeholder such as
<!-- TODO: confirm default --> is allowed only as a temporary local drafting marker: it must be resolved before the task is complete, and it must never remain in a document that is committed or proposed in a PR. Do not mark the task done while any placeholder remains.
- When the source and an existing doc disagree, note the conflict under Open questions in the plan instead of silently picking one.
Accepted inputs
| Input type | Examples |
|---|
| Code PR | pingcap/tidb PR link, diff, or reference |
| GitHub issue | Feature request, bug report, design discussion |
| Product spec | Feature specification, product requirement document |
| Design doc | Technical design, RFC, architecture proposal |
| External reference | Blog post, conference talk notes, user feedback |
| Rough notes | Bullets, chat messages, informal descriptions |
| Existing docs | Current doc page that needs improvement |
Multiple inputs can be combined. More context = fewer questions needed.
Defaults
- Inspect first, confirm when uncertain, then edit.
- Prefer updating existing docs over creating new pages.
- Not every code change needs a doc update. Documentation must justify its maintenance cost.
- Write in Chinese, following
.agents/shared/writing-style.md and resources/tidb-terms.md. Keep product names, commands, flags, paths, and config keys in English.
- If the user asks about local changes without naming files, start with
git status -u or git show --name-status.
Step 1: Load shared context
Always read before making any doc changes:
.agents/shared/repo-conventions.md.agents/shared/writing-style.md
Read only when relevant:
.agents/shared/translation-rules.md and .agents/shared/translation-terms.md — when translation from pingcap/docs is involvedresources/tidb-terms.md — when terminology is uncertain
Step 2: Analyze the input
From a code PR
gh pr view <PR-URL> --json title,body,labels,baseRefName,headRefName,files
gh pr diff <PR-URL>
Scan for documentation-relevant patterns. This table is for triage — deciding whether docs are affected and which area. For the exact target-file mapping with specific file paths, see ref-update-existing-doc.md.
| Code pattern | Likely doc area |
|---|
New/changed SysVar / DefValue | System variables |
New/changed config field / toml tag | Configuration files |
| New/changed command-line flag | Command-line flags |
| New SQL statement or grammar change | SQL statements |
| New built-in function | Functions and operators |
New INFORMATION_SCHEMA table | Information schema |
| New feature flag or experimental gate | Feature doc (new or existing) |
| Changed default or compatibility | Relevant docs + possibly release notes |
Focus on user-facing changes. Skip internal refactors that do not affect behavior.
From a product spec, issue, or design doc
Extract:
- What can users now do, configure, or observe that they could not before?
- Which components are affected?
- Which versions will include this?
- Any constraints, limitations, or compatibility concerns?
From rough notes or verbal description
Extract key user-facing facts. Ask focused questions only for facts that cannot be derived from code, tests, or existing docs.
Step 3: Determine the target branch and version
Target branch
| Source context | Docs target branch |
|---|
| New development (default) | master only |
| Version-specific behavior across maintained versions | master + needs-cherry-pick-release-X.Y labels |
ai/ content | Do not edit here; author the English source in pingcap/docs |
Follow the repository's cherry-pick model (see .agents/shared/repo-conventions.md): default to a single PR on the latest applicable branch (usually master) and rely on cherry-pick labels for other maintained versions, rather than opening parallel PRs per branch. When in doubt, target master.
Version number for "从 vX.Y 起" notes
Many entries need a precise version (for example, "本变量从 vX.Y 起引入"). Determine it from the source — do not guess:
- Code PR: derive from the PR milestone, the target release branch, or the next unreleased version on
master.
- Spec/issue/design doc: use the stated target version.
If the version cannot be determined from the source, mark it with a placeholder and list it under Open questions instead of inventing a number.
Step 4: Decide — create new page or update existing
Ask these questions:
- Does this change have a natural home in an existing page?
- Would adding it to an existing page make that page too long or dilute its focus?
- Does it introduce a distinct user task or feature that needs standalone discoverability?
- Is there enough substance for a standalone page (≥3 meaningful sections)?
| Answer | Action |
|---|
| Fits in existing page(s) | → Load ref-update-existing-doc.md and refer to it to update the existing page(s) |
| Needs a new standalone page | → Load ref-create-new-doc.md and refer to it to create a new standalone page |
| New page + related updates to existing pages | → Load both; start with ref-create-new-doc.md |
Then follow the loaded reference file's workflow from start to finish.
Shared gotchas
These apply to both creating and updating:
- The
ai/ content is AI-translated weekly from pingcap/docs. Edit the English source there, not the Chinese copy here.
- Do not change
CustomContent blocks without understanding platform-specific rendering.
- Do not silently broaden scope from a targeted fix into cross-file rewrites.
- Preserve code samples, commands, SQL, config names, API fields, JSON, EBNF, and UI strings unless the task requires changing them or they are clearly wrong.
- Apply Chinese writing conventions: Chinese/English spacing, full-width punctuation in prose, and terminology from
resources/tidb-terms.md.
Where this skill stops
This skill produces local edits + validation + a completion report. It does not create branches, commit, push, or open a PR on its own.
- After editing and validating, report the changed files and follow-ups, then stop.
- Create a branch, commit, or open a PR only when the user explicitly asks, or hand off to the workflow the user prefers.
- Release notes and PR creation are separate steps — flag them as follow-ups rather than doing them inline.
Coordinating with other skills and workflows
| Task | Where it is handled |
|---|
| Guard PR template metadata when opening the PR | docs-pr-metadata-guard skill |
| Guard issue template metadata | docs-issue-metadata-guard skill |
| Review the resulting documentation PR | review-doc-pr skill |
Translate an English PR from pingcap/docs into Chinese | .github/workflows/sync-doc-pr-en-to-zh.yml (see .agents/shared/translation-rules.md) |
Keep ai/ content in sync | .github/workflows/sync-ai-docs-en-to-zh.yml |
Output format
Plan (before editing):
Target: <branch/path>
Source: <PR URL, issue, spec, or description>
Action: <create new page | update existing | both>
Doc type: <task | concept | reference | new feature | troubleshooting>
Outline: <heading list>
Related updates: <TOC, links, overview, release notes>
Open questions: <facts needing confirmation>
Completion report:
Changed files:
- <path>: <what changed>
Source: <link or description>
Validation:
- <check>: <result>
Follow-up:
- <release notes, English source coordination, or other needs>
