bulk-massive-tasks-efficiency-skill
Bulk and Massive Tasks Efficiency
Finish the complete authorised population with fewer repeated reads, preserved evidence and exact outcome proof. This is an execution method, not authority to contact people, edit records or add sources. Apply the current request, owner policy and provider-specific workflow first.
1. Bind one population and finish line
Define system/account, member identity, explicit population/window/count, nested payload required, exclusions, allowed mutations and proof layer. Use configured owner defaults when no window is supplied. A named cohort stays that cohort; a single-file task does not become a disk or provider census.
Distinguish source discovery from processing: titles, IDs, headers, labels and prior summaries can locate members, but cannot replace their bodies or action-changing details. Relevant attachments, related tasks, later replies and native outcomes are part of the payload where they affect the decision. Include archived/closed/submitted members when the named population or newer evidence gives them a live obligation; do not let static status filters erase work.
Search for the current canonical artifact/process before creating a tracker, export or duplicate. Keep coverage and working state in memory or the existing native task surface. Create local files only when the user asks or a provider operation inherently requires the minimum file.
2. Ingest once before classification
Inspect the current supported schema, batch endpoints, pagination limits, response-size limits, authentication and rate limits. Prefer the narrowest connector/API/CLI or database read that returns full payloads. Use bulk body/thread reads where they preserve complete coverage. When rich search can omit objects, independently exhaust the ID census before hydration; a successful body query cannot replace source-ID reconciliation.
Retrieve the complete bounded population before ranking, deciding relevance or mutating. Pagination is transport only. Use provider-sized batches, with bounded concurrency for independent reads, until exhaustion is proven. Preserve stable member IDs, exact source bounds, returned counts, full payloads and cursors. Detect repeated cursors, premature empty pages and partial/error responses instead of interpreting them as completion.
“One ingest” means one logical pass across every member, not one oversized tool response. Do not lose bodies to a truncated dump, output cap or context overflow. Keep raw payloads in the supported session/runtime memory; expose complete bounded chunks once as necessary. Detect truncation and read the missing slice before classification. Do not silently replace complete ingestion with previews, summaries, samples or rank-then-drop. If a runtime cannot support the required population, use a supported larger/bulk route or state the precise unavailable slice; never claim full coverage.