everyinfra-bulk-data-export Skill
通过 EveryData 导出有边界的结构化数据集,支持能力发现、成本控制、分页、去重,以及 CSV 或 JSON 校验。当用户需要大量记录、所有可用页面、数据集或 CSV/JSON 导出,而非单次 API 响应时使用。
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)
EveryInfra bulk data export
Use everyinfra_list_capabilities and everyinfra_call_api. This is a paid repeated-call workflow,
not a license to keep fetching until inventory is exhausted.
Before calls
- Identify the target record count, required fields, output format, and any time or cost ceiling.
- Discover the exact platform/action and read its required parameters, pagination fields,
max_limit, and current catalog price. Do not guess endpoint or cursor names.
- Estimate the maximum calls from the target count and live limit. State the estimate and obtain
the user's go-ahead before a material multi-call pull unless that exact bounded spend was already
authorized.
Fetch and stop
- Request the largest useful page within the live limit.
- Follow only pagination fields returned or declared by the capability. Stop when the response says
there is no next page, the requested row or call cap is reached, the budget is exhausted, or the
source stops advancing.
- Retry only transient failures with a small bounded retry count. Do not retry authorization,
parameter, inventory, or insufficient-balance errors as if they were transient.
- Deduplicate by a stable returned identifier. If none exists, document the composite key used;
do not silently drop similar-looking records.
Deliver and verify
- JSON should preserve the selected normalized fields and record types.
- CSV should use a stable header and flatten only fields chosen for the export; keep nested evidence
in JSON when flattening would lose meaning.
- Verify that the file exists, is non-empty, parses, and contains no duplicate deduplication keys.
- Report requested versus delivered rows, calls made, duplicates removed, truncation reason, output
path, and the sum of actual returned billing blocks. If billing is absent, say the actual charge
was not observable instead of substituting the catalog estimate.