docs(skills): resolve the text format in using-bee and follow newer Backlog features

[lollipop-onl]

Stacked on #170 → #172 — review those first. The base moves down the stack as each one merges.

Summary

Updates both Agent Skills to current authoring guidance and to Backlog's newer features, and fixes agents treating Backlog notation (Backlog記法) as Backlog's only markup.

Why agents misread Backlog notation

backlog-notation's description triggered on any formatted text headed for Backlog, and on projects whose format was unknown. So it loaded on Markdown projects too, and its body then read as "how Backlog text works". #126 added a gate in the body, but the description (which is what decides loading) still won.

What changes

• using-bee owns format resolution ("Writing Text to Backlog"): a stated rule in the prompt or AGENTS.md → reuse this session's check → otherwise bee project view --json textFormattingRule once → ask the user if it can't run. It never guesses from existing text.
• Documents are always Markdown. The Add document API parses content as Markdown in every project, so the skills say so explicitly.
• backlog-notation only triggers once the format is known to be Backlog notation, when invoked by name, or when the user asks about its syntax / conversion in either direction. Its body repeats the check so it still works without using-bee.
• Newer Backlog features: the document commands (feat(document): add comments, count, add-tag and remove-tag commands #170) and related-issue commands (feat(issue): add related, add-related and remove-related commands #172); there is no document-update endpoint (don't delete and recreate without asking); newer spaces and projects have Wiki off, so long pages go to documents.
• Mentions: <@U{numeric id}> (also <@T{id}> / <@project>), the same in both formats; plain @Name is text. Checked on a live space: a mentioned member of the project is notified, a non-member is not, and saved mentions come back as @Name, so edited text must restore them. Document bodies posted through the API can't hold mentions (<@U…> stays literal), and there is no public API for document comments.
• Authoring guidance (Anthropic best practices, agentskills.io): third-person what + when descriptions with the main use case first, Japanese trigger words (バックログ, 課題, プルリクエスト…), near-miss exclusions ("product backlog", Jira). Frontmatter stays name / description only.
• bee api writes the user explicitly asked for no longer need a second confirmation, since that stalled non-interactive runs.
• AGENTS.md and the AI agent guide page describe the new split.

How the text was produced

The prose was run through gvzdv/claudish-to-english's Markdown hook (rewrite-md.sh, codex provider). Two sentences where the rewrite weakened the meaning were restored. A separate model (Fable) then reviewed trigger behaviour with should / should-not-trigger queries in English and Japanese, routing between the two skills, and facts against the CLI source. Its fixes are applied. Two suggestions were not applied: trimming the intentional repetition, and a low-confidence claim about > cell merging.

Not re-verified

The Backlog notation syntax table is unchanged from #126. The Backlog help site returned 403, so it could not be re-checked.

Test plan

• vp check passes
• vp run --filter @nulab/bee generate:skill:check passes
• Both descriptions ≤ 1024 chars (724 / 741)
• Try in Claude Code: a Markdown project, a Backlog-notation project, and a document post pick the right format

🤖 Generated with Claude Code