Agentic Email Processing
How a unified config file turns your inbox from a pile of unrelated messages into a prioritized, project-aware briefing. The system drafts, but never sends. That is a design decision, not a limitation.
The Supervisor-Worker Pattern
One triage agent reads every email and routes it, and specialized workers handle enrichment, prioritization, and drafting. A daily digest surfaces what matters, and it is your one checkpoint: you approve or adjust suggested drafts there. The human checkpoint is the point of the architecture.
One File. Every Automation.
The config is not a settings file; it is the system's understanding of your practice. When you update a project here, every downstream automation inherits the change. This is the Setup Frame applied to systems, not conversations. The config is version-controlled and access-controlled like any other critical system file. Changes are committed to Git (a system that records every change and who made it), reviewed before deploy, and auditable after the fact.
A 1-10 score that influences digest ranking. The Death Star assault outranks supply logistics. When two emails arrive at the same time, the system knows which one you see first.
Per-contact tone guidance. Mon Mothma gets formal briefings. Han Solo gets directness. The drafter reads these notes before generating a response. Same principle as giving Claude context about your audience.
The trust calibration. "Low" means the system produces a skeleton with placeholders. "High" means it drafts a near-ready response. Strategic relationships stay at "low" because the cost of a wrong tone is higher than the time saved.
What the enrichment worker uses to pull relevant context. When an email from Ackbar arrives, the system knows to surface the exhaust port analysis and squadron readiness status alongside it.
What the System Produces
The emails are invented, and so is this run: it shows how the design handles a typical morning. Five emails arrive overnight. The triage agent classifies each one, the enrichment worker pulls context, the prioritizer ranks them, and the drafter generates responses where confidence allows. Click each email to see what the system surfaces.
Death Star Assault Planning
Open deliverable: "Attack run briefing" is in-progress and due before Yavin approach. Exhaust port analysis thread is active.
9.4 / 10 - Highest-weight project, deliverable deadline, requires response from strategic leader.
Death Star Assault Planning
Ackbar's comm style: direct operational language, dislikes ambiguity. X-wing squadron readiness is an active thread.
8.1 / 10 - Same project, deliverable-related, but review request rather than decision-required.
Independent Contractor Pipeline
Lando is a warm referral. Charming but cautious. Wants to know what is in it for him. Relationship is "cultivating."
5.2 / 10 - Medium-weight project, scheduling request, no deliverable attached.
"Thursday works for me. Same time? Looking forward to catching up on the Cloud City logistics."
Base Operations (auto-matched by domain)
No active threads or deliverables connected. Routine operational update.
1.8 / 10 - Low-weight project, informational, no response needed.
"Received, thank you. Filed for reference."
None. Contact not in config. Email flagged for manual review.
No prior history. Possible referral, possible noise. System cannot determine intent.
Unscored. Unknown contacts bypass the priority algorithm.
Notice what the system does not do. It does not reply to Mon Mothma, guess at Maz Kanata's intent, or skip the human checkpoint for high-stakes relationships. The value is not in what it automates. The value is in how it organizes your attention so you spend your judgment where it matters most.
Build Sequence
Each phase adds one capability. You can stop at any phase and still have a useful system. The sophistication of the architecture follows the complexity of the problem, not the other way around.
Config Migration
Move from per-client files to one unified config. Existing transcript processing reads from the new file. Validate that output quality is identical. Change nothing else.
Email Triage
Build the triage agent. It classifies email against the config and writes a digest. No drafting. Backtest against three days of history, then run live for a week. Target: 85% classification accuracy.
Enrichment Layer
When an email matches a project, pull active threads, deliverable status, and contact communication style. The digest now includes a "Context" block per email.
Draft Responses
Generate drafts where confidence allows. All drafts go to the Gmail drafts folder. Track: used as-is, edited then sent, or discarded. Target: 40% usable with minor edits. Zero relationship-damaging drafts.
Daily Digest
Connect all workers to a morning briefing. "Here is what you are working on today, ranked by priority." The inbox becomes a curated briefing rather than an unfiltered stream.
Config Review Cadence
A config is only useful if it reflects current reality. Three rhythms keep it honest.
What This Teaches
Every design decision in this system maps to a principle from the program. The same skills you are building in conversation with AI apply to building systems with AI.
Config as Setup Frame
The config file is a Setup Frame for an automated system. The quality of the context you provide determines the quality of every output.
Backtest as the Loop
Running three days of historical email through the triage agent and refining the rules is the Loop applied to systems. Test, evaluate, adjust, repeat.
Polished Drafts Are Dangerous
The drafter produces plausible responses. The more polished they look, the less critically you evaluate them. The human checkpoint exists because AI predicts what is plausible, not what is true.
Trust Calibration
The max_draft_confidence setting is explicit trust calibration. High confidence for scheduling acknowledgments. Low confidence for strategic relationships. You set the threshold before the system runs, not after.
Intentional Friction
"Drafts, never sends" is a design decision. The goal is to remove the noise between the human and the decision, not the human.
Start Manual, Then Isolate
Phase 1 changes nothing except how the config is structured. Each subsequent phase adds one capability.