Managing Sylvie's work email

This is the lead planning document for Sylvie's Email Management. Updates since approval are marked (21 Oct); anything that did not fit in place is in the addendum.

Situation

In one sentence: Fae has taken over Sylvie's work email (~70 emails a week), cleared the backlog and runs a manual process, earlier attempts with an AI digest and labels require improvement.

• Scope: Sylvie's work email. Her personal email may follow later, using the same framework with different details, once access to it is resolved (21 Oct).
• Volume: ~70 emails a week. A rough scan suggests ~80% are automatic (notifications, calendar, meeting notes, newsletters, marketing), ~10% with requests or information on life itself processes or projects Fae could likely handle and ~10% need Sylvie.
• Done so far: backlog cleared (archived everything older than two months, with some spam deleted (21 Oct); last month checked with Sylvie). Fae deletes ~25 automatic emails a week using her own judgement, not yet confirmed with Sylvie.
• Access: Fae uses delegated access because direct login needs Sylvie's phone every time. AI can't read mail through delegated access, but a forwarding workaround can be used.
  • Rufus (PR #21 review): this is a big issue. Can Fae connect once without having to renew each time? Or, as Rufus does, set up Sylvie's machine to run the automated review, with Fae reviewing the results and handling the manual emails.
• Tried: an AI digest, which failed for lack of context (for Fae & AI) on what mattered to Sylvie, and labels, which were too many and used before sorting was finished, so Sylvie didn't know where to look. Sylvie does want labels, and a better digest might still work.
• How Sylvie works: she skims the inbox, keeps it as her reference and reminder list, and mainly worries about missing key actions.
  • Rufus (PR #21 review): weaning Sylvie off using her inbox as her task inbox would be really valuable. Keeping the inbox so she can quickly find something is one thing; using it as her task list is another. See if Fae can set up a task tracking system for her that works (perhaps one Fae manages).
• Now: Fae checks the inbox Mon/Wed/Thu, replies and archives what she can, logs actions in meeting notes for Sylvie along side her personal task management system, and keeps two weeks of email in the inbox.

Complication

In one sentence: there's no agreed goal, no clear status or ownership, and Sylvie's preferences aren't written down, so neither Fae nor AI can take on more of the work with confidence.

• No agreed goal: it isn't defined what Sylvie wants: to stop checking her inbox, inbox zero, confidence that nothing is missed or other.
• Different needs: Sylvie uses the inbox as a reference, while Fae needs inbox zero to keep an overview, so neither has a clear picture.
• Unclear status and ownership: neither can see what's done, who owns what, or what Fae has handled. Current system for tracking tasks resulting from emails risks fragmentation and could be more time efficient.
• Knowledge gap: what's relevant to Sylvie, and what Fae can handle, depends on sender and context, not just email type. That knowledge isn't written down, so it can't be reused by Fae or AI.
  • Rufus (PR #21 review): this is key. AI (and much human) delegation works when there are clear, measurable goals and clear parameters and constraints.

Question

How do we build a system for Sylvie's work email that learns her preferences, so that within about a month AI sorts ~80%, Fae handles ~10%, and Sylvie only sees the ~10% that needs her, with nothing waiting more than a few working days?

Hypothesis

A few clear labels and a written rulebook of Sylvie's preferences, applied by AI and updated with every correction, will let AI sort most email, Fae handle the routine, and Sylvie see only what needs her.

• Rufus (PR #21 review): it may be a bit early for a hypothesis. The first complication is the goal: what it is, and what parameters or constraints a hypothesis must satisfy. Resolving one complication often reveals the next.

Issue tree

• 1. Where are we now?
  • What actually arrives each week, by type and sender? ✅ Answer (provisional). ~70 emails a week. A scan of previews from one week (10–15 Sep) found ~38% platform notifications, ~33% calendar emails, ~11% AI meeting notes, ~11% people and project emails, ~7% newsletters and marketing. A full read of one week is still to do.
  • Which emails need Sylvie herself, and why? ✅ Answer (provisional). About 10% (Fae's estimate): decisions and commitments only she can make, e.g. a supplier's scope of work, a contractor's signed contract, requests for access to her documents.
    • Rufus (PR #21 review): distinguish threads where Sylvie is only cc'd (Rufus often cc's her). Those don't need her unless she is named; at most, a line in a summary.
  • Which of Fae's current deletions are safe, and which need checking? ✅ Answer (provisional). Likely safe: calendar emails (when the change is already in her calendar), newsletters and Substack activity, social media and Google Docs notifications, meeting-recording notices, marketing. To check with Sylvie: security and sign-in alerts, payment and paid-service notices, access and share requests, and updates from people or groups she works with.
    • Rufus (PR #21 review): once confirmed safe, turn as much of this off at the source as possible; e.g. calendar notifications can be switched off in calendar settings, as Rufus did.
• 2. What does Sylvie want?
  • What is the goal: not checking her inbox, inbox zero, or confidence that nothing is missed? ✅ Answer (provisional, 15 Oct check-in). Sylvie wants to read only the ~10% of mail that needs her own response, mainly replies to things she wrote, since she wants to see their reply. Fae moves those into a dedicated "to read" folder and Sylvie checks only that folder.
  • How does she want to keep a reference to key emails and documents? ✅ Answer (provisional). Today she uses the inbox itself as her reference for key documents, actions and reminders. Whether she wants to keep doing that is still to confirm.
  • How does she want to see what needs her, and what Fae has done? ✅ Answer (provisional). She wants labels. A better digest may also work; the first one lacked context. To confirm. (21 Oct: from 15 Oct, what needs her goes in the "to read" folder.)
  • Are 80/10/10 and "a few working days" the right targets?
• 3. How do we set it up?
  • Which labels (three or four, in Sylvie's words), and what stays in the inbox? ✅ Answer (provisional). Fewer than last time, when too many labels confused her, and only once sorting is finished so she knows where to look. Names still to agree. (21 Oct: one agreed so far, the "to read" folder.)
  • What goes in the rulebook: which senders and types get deleted, filed, answered by Fae, or passed to Sylvie? ✅ Answer (provisional). Start from Fae's current deletion list (see 1), once confirmed with Sylvie. (21 Oct: inputs are logged in rulebook inputs.)
  • What can Fae handle without asking? ✅ Answer (provisional, 15 Oct check-in). Wise emails are ops: Fae removes them without Sylvie seeing them. The rest is still to agree, through the rulebook.
  • How can AI get access to the email (forwarding, or something else)? ✅ Answer (provisional). Delegated access doesn't let AI read mail. Forwarding to Fae's inbox worked before but is now off. Other options to ask Rufus about. (21 Oct: Rufus's options are under Access in the Situation.)
  • How does AI apply the rulebook, and what happens when it's unsure?
• 4. How do we pilot and learn?
  • What should be captured in a rule book?
  • How are corrections captured and added to the rulebook? ✅ Answer (provisional, 15 Oct check-in). Screen-share check-ins with Sylvie to refine the categories over time; Fae logs each correction in the rulebook inputs.
    • What do we measure, and when do we review with Sylvie? ✅ Answer (provisional). Keep it light: a weekly label search for overdue items needing Sylvie, a rough split and hours estimate, and a note of anything missed, all checked in the existing weekly review.
• 5. How do we keep it safe?
  • Not missing important emails
    • Which senders and topics are always passed to Sylvie, whatever the rules say (a "never auto-sort" list)? ✅ Answer (provisional). Currently only from Armelle.
    • When AI or Fae is unsure, does the email default to Sylvie rather than being deleted?
    • Do we archive rather than delete, so anything sorted wrongly can be recovered? ✅ Answer. Rulebook should specify what can certainly be deleted, what should be removed at the source, and what should be archived and labelled to check with Sylvie before deletion.
    • How do we spot-check AI's sorting, especially early on?
  • Security and privacy
    • What access does each of us have (Fae: delegated; AI: forwarding or other), and is it the minimum needed? ✅ Answer (provisional). Fae has delegated access; AI currently has none (forwarding is off). Whether that's the minimum needed is still to review.
    • Which emails should AI never read or process (personal, medical, legal, financial, other people's private matters)? ✅ Answer (provisional). Sender or subjects to be agreed with Sylvie and could be triaged into a folder without the content being read by Claude.
    • What can Fae or AI send on Sylvie's behalf without approval, if anything? ✅ Answer. Nothing yet: Rufus's ground rule is "draft, don't send": nothing is sent, signed or paid without approval from Rufus or Sylvie. (21 Oct, from 15 Oct check-in: Fae drafts replies; Sylvie will demo her own drafting and her AI tone pass.)
    • How do we treat security alerts, sign-in warnings and suspicious (phishing) emails? ✅ Answer (provisional, 15 Oct check-in), partial. Authentication codes: Sylvie finds them herself by searching the tabs, since they often land in Social. Sign-in warnings and phishing still to agree.
    • Where are the rulebook and notes kept, and do they avoid personal data? ✅ Answer (provisional). Per Rufus's ground rule, no personal documents or data in the repo; anything personal goes in Drive and is linked. Where the rulebook itself lives is still to decide. (21 Oct: inputs are logged in the project folder, rulebook inputs.)
    • How is access reviewed or removed if roles change? ✅ Answer (provisional). This should be part of Asdra's offboarding process (on backlog) for email and AI passwords to be changed when colleagues leave organisation.

Plan of work

Weekly review routine: 15 minutes with Sylvie at the first check-in of each week: run the overdue search (e.g. label:sylvie-action older_than:3d), go through what needs her and what Fae closed, and note any rulebook changes. Record one line in the meeting notes: "Overdue this week: 0" or the items.

Check-in points

• Rufus (pull request, Thu 24 Sep): Is the question at the right level, with the right success criteria? Is the issue tree complete? What are the options for AI access, and is forwarding Sylvie's email to AI OK for privacy? Should this file live in docs/plans/ instead? (21 Oct: moved to docs/plans/. Rufus's review answers are inline above.)
• Sylvie (Mon 28 Sep): review: the goal, the targets, how she wants to see what needs her, and cover on days Fae doesn't work.
• Sylvie (Mon 5 Oct): review: the draft rulebook, label names and which deletions are safe.
• Sylvie (Wed 21 Oct): review: keep, adjust, or widen AI's role after the pilot.

Where actions are logged

• GitHub issues: life-itself/onboarding #103 and life-itself/planning #23–26. (21 Oct: moved to Beads, epic lip-kr1 (plan steps 2–9 are lip-kr1.5 and lip-kr1.7–.13), plus lip-87i. The GitHub issues are closed with links.)

Sources

• Briefing: Asdra and Fae planning practice
• First draft and Rufus's feedback (issue #12)
• Rufus's feedback on Asdra's inbox plan (PR #17)
• Rufus's review of this plan (PR #21)
• SCQA and Issue trees guides

Addendum (21 Oct 2026)

Added after Rufus approved the plan. It records what has changed and what did not fit in place above; it does not change the SCQH.

• Next step: the AI analysis (plan step 3). Forwarding is still off, so the full read of one week has not happened, and branch 1 and the rulebook depend on it. Turning forwarding back on needs Sylvie's OK. Tracked as lip-kr1.7, due Thu 29 Oct.
• Rulebook inputs. Agreed preferences from the 2 Sep and 15 Oct check-ins (onboarding #103 and lip-87i) are logged in rulebook inputs. Fae adds to it as preferences come up; it feeds plan step 4.
• Rufus's own rules. Rufus keeps rules for AI management of his own email in rufuspollock/planning email/RULES.md. That repo is private and Fae can't open it yet; ask Rufus for access or a copy to use as a model. Asdra's plan for Rufus's inbox is PR #17.
• No digest runs for Sylvie's mail. No AI digest for Sylvie's mail runs at present. An older email-triage-and-digest skill in life-itself/ops is not in use.
• Personal email. Out of scope until access is resolved: delegated access to Sylvie's personal Gmail was blocked by a 2FA prompt on 2 Sep (onboarding #103). The password shared temporarily during that setup still needs changing (Beads lip-kr1.1).
• Plan dates. The dates in the plan of work have passed. Done: step 1 (Rufus approved 30 Sep). Steps 2–9 to re-date once step 3 is done.