💡 New script: gmail-decision-log (automated decision extraction & archival from labeled email threads)

[github-actions[bot]]

Summary

Scan user-labeled email threads using Gemini to identify and extract decision statements, capturing the decision, who made it, when, and the surrounding context. Append extracted decisions to a running Google Doc organized chronologically, creating a persistent, searchable organizational memory for decisions made via email. Follows the project's label-scoped processing pattern from gmail-to-drive-by-labels.

Market Signal

Research identifies email-to-knowledge-base as the biggest gap in the personal productivity AI market. No tool effectively builds a persistent, searchable knowledge base from email for Google Workspace users. Decision tracking tools like Notion and Confluence require manual entry. The average knowledge worker receives 121 emails/day — decisions made in email are nearly impossible to find later, and teams waste hours re-debating things that were already decided. Read.ai's Personal Knowledge Graph is meeting-centric. No tool automates email-to-decision archival. REM Labs and similar AI PKM tools focus on manual curation, not automated extraction.

User Signal

No existing idea in the backlog covers decision extraction or organizational memory from email — this is a genuinely new surface area. The project's gmail-to-drive-by-labels already archives full email content to Docs, and this script applies the same pattern with intelligent filtering (decisions only, not full threads). The gmail-ai-classifier demonstrates Gemini-powered email analysis as a proven pattern in this codebase. The label-scoped architecture means users opt in by labeling threads, preserving the human-in-the-loop quality control.

Technical Opportunity

Follows the gmail-to-drive-by-labels architecture closely: label-scoped Gmail processing, Google Doc output with chronological organization, and the dual-file testable pattern. Gemini structured output extracts decision metadata (statement, author, date, confidence score). The existing test-utils mocks (GmailApp, DocumentApp) cover all needed services. The getCleanBody() utility in gas-utils.js handles quoted-reply stripping, ensuring clean input for Gemini. Label-scoping means users explicitly mark threads with a "decisions" label, acting as a quality filter.

Assessment

Dimension	Score	Rationale
Feasibility	med	Gemini structured output for decision detection requires careful prompt engineering; confidence thresholds needed to filter noise; follows proven architecture
Impact	high	Fills the biggest identified market gap (email-to-knowledge-base); decisions are the most valuable and most commonly lost type of email content
Urgency	med	No specific time trigger, but the problem compounds over time — every day without a decision log is more institutional knowledge lost

Adversarial Review

Strongest objection: A "decision" is subjectively defined and contextual. The script would either capture too much (every casual statement of intent like "let's do pizza for lunch") or miss important decisions that don't use explicit decision language. False positives erode trust faster than false negatives.

Rebuttal: The label-scoping pattern is the key defense. Users apply a "decisions" label to threads they KNOW contain important decisions — the same curation pattern that makes gmail-to-drive-by-labels effective. The script's job is to extract the specific decision statement from within those pre-curated threads, not to determine which threads contain decisions. Gemini's confidence scoring (validated by the classifier's existing validateClassification pattern) provides a second quality filter. The MVP extracts only high-confidence decision statements from user-labeled threads, combining human judgment (labeling) with AI precision (extraction).

Suggested Next Step

Define the Gemini extraction prompt with sample decision statement patterns, design the Google Doc output format (date, thread subject, decision statement, decision maker, participants, confidence score, link to original thread), and validate extraction accuracy against a test corpus of real email threads containing decisions.