菜单
Check whether the models, the product changelogs, and the connectors this team depends on have moved since last week, and report what it means for this team. Detect and report only. Use as the first step of the weekly tune-up.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/AutomatedMarketer/agent-team-template --skill check-whats-changed命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Apply the week's learning to the team's written rules, raise everything that would rewire the team as a task card, and write the one-page tune-up report. Use as the closing step of the weekly tune-up.
Write the run-log entry that records what this run did. Use at the end of every agent run, scheduled or manual, before committing.
Turn the owner's reaction to a draft — shipped, edited, or binned — into a filed verdict and a written rule, so the same correction never has to be made twice. Use whenever the owner says what they did with a piece the team produced.
正在显示 SKILL.md
| name | check-whats-changed |
| description | Check whether the models, the product changelogs, and the connectors this team depends on have moved since last week, and report what it means for this team. Detect and report only. Use as the first step of the weekly tune-up. |
| audience | team |
A team quietly running on last quarter's assumptions is the most expensive failure in this repo, because nothing looks broken. This step asks one question every week:
Is anything this team runs on no longer true?
You detect and report. You do not change how the team is wired. Swapping a model, pulling an upstream change, or re-authorising a connector are all the owner's calls, and they arrive as task cards, not as edits you made.
watch-updates — the second step of the security review — already watches the installed
skills and the three product repos. Do not repeat that work here. If the security review has
not run in over two weeks, say so in one line and move on; a watcher that stopped watching
is itself a finding.
agents/orchestrator/output/last-checked.md holds where each item stood last Sunday. No
note? That is a normal first run: record today's state, report "baseline established",
and skip every comparison below.
The repo names opus and sonnet as aliases, never pinned ids, exactly so this lane
rarely produces work. So check the two things that actually rot:
shared/standards/model-card.md now looks
wrong. Do not change the card.grep -rn "claude-[a-z]*-[0-9]" --include=*.md --include=*.mjs --include=*.json . | grep -v node_modules
Read the changelog for each tool the team's work actually depends on — Claude Code first, then anything else in the watch list. Compare against the last-checked marker.
Report only what changes how this team should work: a new capability that would replace something the team does by hand, a behaviour change that breaks an assumption in a skill, a deprecation with a date on it. Skip everything else. A tune-up report that lists forty release notes is a report nobody reads twice.
connections/register.yml is the list — every connection /connect has proved, with the
account it uses and the workflows that depend on it. Cross-check it against
.claude/settings.json, .mcp.json if present, and the names in .env.example:
| Check | What a finding looks like |
|---|---|
| Still authorised | An expired login is the most common cause of a workflow that "just stopped working" |
| Still reachable | A dead endpoint, reported before a scheduled run hits it at 6am |
| Configured but unused | Named in config, referenced by no skill or workflow. Attack surface with no benefit |
| Used but not configured | A skill calls something the repo cannot reach. This one is urgent |
Also report any recipe in connections/recipes/ whose verified date is more than 90 days
old. Connection details change monthly, and a stale recipe is a step-by-step guide to a page
that has moved.
Never print a key, a token, or a URL containing one. Report the connector by name and its state. If you cannot test a connector without spending money or sending something, do not test it — say it was not tested and why.
stack.yml declares the capabilities every team gets — read the file for the list rather
than carrying a count in your head, because the count changes — and
.agent-team/stack-check.md records the last time they were verified. Report any that are
missing, disabled, or never verified, and name what the team does worse without each one
rather than just listing it as absent.
A capability that quietly disappeared is worse than one that was never installed: the team carries on answering, just less well, and nobody is told.
Anything in runtimes.yml, plus any GitHub repo a skill in this repo tells the team to use.
Newest release or commit date against the last-checked marker, one line each.
An unreachable item keeps its old marker and is reported as unreachable. A failed fetch is never "no news" — that is how a watcher goes blind without anyone noticing.
Rewrite agents/orchestrator/output/last-checked.md: one line per watched item — what it
is, the marker you saw, today's date. This note and the report are the only things this
step writes.
Pass to learn-from-the-week: what moved, what it means for this team, and which findings
deserve a card. Recommend; never rewire.