email-iteration-loop

Skill: Email Iteration Loop

Use this skill in the weekly retrospective to improve triage policy and drafting behavior from real outcomes.

Goal

Make the system safer and more useful over time without silently becoming more aggressive.

Review Inputs

• archived or trashed emails later rescued
• unsubscribed senders later manually re-engaged or manually re-subscribed
• messages left alone that the operator later archived, deleted, or marked spam
• drafts created by the system and their final sent versions
• threads repeatedly escalated because the model was unsure
• operator notes from the audit log

Questions To Answer

1. Where was the system too aggressive?
2. Where was it too timid?
3. Which senders or domains need explicit overrides?
4. Which categories need threshold changes?
5. Which draft patterns sound unlike the operator?

Required Outputs

• policy update proposals
• voice rule update proposals
• contact-tier update proposals
• allowlist or blocklist additions
• threshold change recommendations
• changelog entry

Safe Auto-Applied Updates

Allowed without approval:

• raise thresholds
• add new protected contacts
• add new blocklist entries for repeated junk
• add unsubscribe candidates
• append retrospective notes and changelog entries

Approval-Required Updates

Never auto-apply:

• lower destructive thresholds
• expand trash coverage
• expand draft classes
• enable auto-unsubscribe
• lower auto-unsubscribe thresholds
• shorten unsubscribe inactivity windows
• relax protected-contact rules
• remove risk boosters

Review Method

Rescue analysis

• count archived threads later moved back to inbox
• count trashed items later restored
• count unsubscribed senders later manually re-engaged
• identify protected-class mistakes immediately

Draft analysis

• classify draft edits as light, medium, or heavy
• extract whether edits changed tone, facts, structure, or commitment level
• update voice rules only when patterns repeat

Confusion cluster analysis

• group repeated ambiguous threads
• identify missing rules or missing contact metadata
• propose one rule change at a time

Output Format

{
  "period": "2026-04-01 to 2026-04-07",
  "safe_auto_updates": [
    "Raise file_min_confidence from 0.985 to 0.99"
  ],
  "approval_required_updates": [
    "Expand prepare drafts to customer logistics"
  ],
  "voice_rule_changes": [
    "Reduce default gratitude openers for vendor ops"
  ],
  "tier_changes": [
    "Add example.com to Tier 0"
  ],
  "confusion_clusters": [
    "Founder intros with forwarded context"
  ],
  "changelog_entry": "..."
}

Changelog Discipline

• timestamp every recommended change
• include evidence count
• note whether change is safety-increasing or convenience-increasing
• require explicit approval for any convenience increase that raises risk