Premise
Large archives fail when an agent holds everything at once. Small batches, checkpoints, and provenance make long extraction recoverable and trustworthy.
Setup
- Grant read access to exported conversations.
- Define exactly which messages qualify.
- Choose a schema and test one project before scaling.
Boundaries
- Preserve original text.
- Every record needs a source.
- Progress must survive interruptions and context limits.
Work in small batches. After each, write a durable checkpoint with source, completion status, and count. If a batch exceeds limits, divide it and resume from the last confirmed checkpoint.
Store exact text, project or conversation reference, available date, and one topic label. Deduplicate only clear copies while retaining every source reference.
Produce a searchable file grouped by topic plus totals, skipped sources, and errors. Never rewrite missing text or invent metadata.
Prompt
Build a searchable archive of prompts I authored across the supplied projects and chats. Extract only [PROMPT TYPE]; exclude assistant responses and unrelated messages.