platform-rules Skill
此技能用于保持 Coupler.io 各技能中作为事实引用的平台规则处于最新状态:预算灵活性、送达率阈值、归因窗口、公开基准、套餐限制、促销日期。触发语包括:“刷新平台规则”、“查一下 Google Ads 规则”、“Klaviyo 基准还新吗”、“某平台改了预算规则”、“把这条规则加进注册表”。基于 registry.md 工作——它是一份记录每条规则、来源、检查时间及哪些技能在用的目录,刷新时只需核对来源,无需重读每个技能。
安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。
技能指令原文(SKILL.md)
Platform Rules
Skills in this repo state platform rules as fact: Google may spend 2× a daily budget, Klaviyo's spam
band ends at 0.05%, TikTok allows 30% budget steps. Platforms change these on their own schedule, and
the same rule is copied into several skills. registry.md is the catalog of those rules:
the value, the platform page it comes from, the date it was last checked, and the skills that use it.
The skills keep the value inline. An agent loading a skill through the MCP sees only that file, so the
registry is for maintainers, never a runtime dependency.
What counts as a platform rule
A statement belongs in the registry when all three hold:
- The skill states it as fact and the model acts on it. A threshold that sets a verdict, a window
that decides which days are complete, a limit that sizes a recommendation.
- Something outside the skill sets it: an ad, email, store or search platform, or Coupler.io's
own product.
- It can change without anyone editing the skill.
The kinds that qualify:
| Kind | Examples |
|---|---|
| Limits and thresholds | Daily budget flexibility, budget step sizes, spam and bounce bands, inspection quotas |
| Windows and defaults | Attribution windows, data-freshness lags, the timezone a platform reports in |
| Published benchmarks | Revenue per recipient by flow, tickets per order. Always with the publication date |
| Platform guidance that changes the advice | When seasonality adjustments help, whether a target change restarts learning |
| Coupler.io product rules stated as a number or default | Refresh intervals by plan, schedule defaults |
| Calendar facts | Black Friday and Cyber Monday dates, which move every year |
What does not belong here:
- Connector facts: column names, report types,
mcp_excluded_params, what a column holds. They
change with Coupler releases and are checked against config and importer code (skill-review
items 2, 3 and 7).
- Method choices: volume floors, break rules, threshold multiples. These are the skill's own
design, not something a platform sets.
- Definitions: CTR is clicks over impressions whatever the platform says.
How a rule is written in a skill
The value, its source and the date it was checked, inline where the model uses it:
A campaign may spend up to 2× its average daily budget on one day and no more than 30.4× it in a
month (Google Ads help, support.google.com/google-ads/answer/2375423, checked 2026-10-28).
Once per skill, where the rules are used, add the check instruction:
If you can browse the web, check the platform's page; if it says something different, use the page
and say the skill's figure is out of date. If you can't, use these figures and give the date.
"Check the docs" on its own is not a rule. In clients without web access it means "recall"
(skill-review item 8).
Refresh
Run on request, for every rule or one platform's rows.
- Fetch each source in the registry and compare it with the stored value. Platform pages only.
A blog or agency post can tell you a change happened, never what the rule now is.
- Classify each row:
- Unchanged. Update
Checked. - Changed. Record the new value with a short quote from the page.
- Unverifiable. The page didn't load or no longer states the rule. Leave the value and say so;
never fill it from memory or a third party.
- For each changed rule:
- Update every skill in
Used in, keeping each skill's wording and the inline "checked" date. - Search the repo for the old value (for example
30.4,0.05%,$6.77) to catch copies the
registry missed. Add them to Used in.
- Bump each edited skill's patch version.
- Open one PR with a table of every changed rule: old value, new value, quote, source, skills
touched. Unchanged rows go in as a single "rechecked" line. Unverifiable rows are listed so a
person can check them by hand.
Work the Unchecked section the same way, a few rows per run. A row moves to Verified only when
its platform page has been read.
Adding a rule
When a review or a new skill introduces a rule:
- Check it is not already in the registry under another name. If it is, the new skill uses the
registry's value; a different value is a contradiction between packs (skill-review item 5).
- Fetch the platform page and add a row: value as stated, source URL, today's date, status, and the
skill and search text in Used in.
- If the page can't be read, add the row as
UnverifiableorSecondhandrather than leaving the
rule out. A known gap is checkable later; a missing row is invisible.
Statuses
| Status | Meaning |
|---|---|
| Verified | The platform's own page was read on the Checked date and states this value |
| Secondhand | Seen only through another party citing the platform (another vendor's help page, a search summary). Check the primary source next refresh |
| Unverifiable | The platform's page couldn't be read. The value is a working rule; skills must say so |
| Unchecked | Found in a skill; no one has checked it against a source yet |
Guidelines
- Verify or say you did not. A stale
Checkeddate is better than a fresh one nobody earned. - Quote the page. The quote is what lets the next maintainer see whether the wording moved.
- One value per rule across the repo. Two skills disagreeing is a defect in one of them.
- This skill updates itself the same way
skill-reviewdoes: through a PR against this file, never
silently in the middle of a refresh.