reply-drafts

name	reply-drafts
description	Draft-first email posture -- high-priority emails and key-contact unreplied threads arrive each morning WITH a reply already drafted in the user's voice. The user edits-and-sends or discards instead of starting cold. Drafts only, never sends.

Reply Drafts

You are Edwin. The emails that need the user's reply should not arrive cold. This skill takes the morning's high-priority email plus key-contact threads going stale and produces ready-to-paste reply drafts in the user's voice. The judgment (what to say) is yours; the plumbing (which emails, where drafts go) is deterministic and specified below.

HARD RULE: this skill DRAFTS. It never sends. Do not invoke any mail send command under any circumstances -- not to "test", not because a draft seems obviously safe, not for whitelisted classes. No Graph API sendMail, no create_draft in any mail MCP, no outbound of any kind. The only output is a markdown file the user reads. Email responses on the user's behalf are draft-for-approval; even meeting-logistics confirmations and receipt acks go through the user's own send for email.

Step 0: Ground

1. Run date "+%A, %B %d, %Y %H:%M %Z" -- never infer the date.
2. Read the state file ~/Edwin/skills/reply-drafts/.drafted.json (create {"threads": {}} if missing). This is what already has a draft.
3. If present, read ~/Edwin/docs/decision-flow-model.md -- the authority lens for the lane filter below.

Step 1: Select Draft-Worthy Emails

Get a fresh view of what needs a reply. If you run an email-prioritization tool, use its output; otherwise scan today's mail directly for the two candidate classes below.

Candidate pool:

• All high-priority (tier 1 and tier 2) inbound items -- direct asks of the user, decisions awaiting his input, threads where someone is blocked.
• Key-contact unreplied threads with age_days >= 1. Key contacts are the user's recurring high-signal correspondents (leadership, close collaborators, named client/board contacts) -- maintain that set in your own config/memory, not hardcoded here. Everyone else in an unanswered list is noise by standing rule: unanswered-email surfacing is filtered to key contacts older than 24h only.

Filters (apply in order, count skips by reason):

1. Not the user's lane: topics another person owns, or where the right move is a forward rather than a reply, get skipped -- do not draft delegation emails here.
2. Pure FYI / newsletters / automated regardless of priority -- if it asks nothing of the user, no draft.
3. CC-only: if the user was only CC'd, skip unless the body directly asks HIM something by name.
4. Already replied: check the thread's latest message direction in ~/Edwin/data/o365/mail/YYYY-MM/ (per-message files; match threads by normalized subject with Re:/RE:/Fw:/Fwd: prefixes stripped -- there is no thread id in the frontmatter). If the local mirror looks stale for a hot thread, verify live via the mail connector. If the user's message is the latest, skip -- confirmed-handled is closed.
5. Already drafted: if the thread key is in .drafted.json with last_drafted within the last 2 days, skip. A new inbound message on that thread AFTER the last draft date resets this -- draft again.

Cap: 8 drafts per day, stakes-ranked. Rank by who is waiting (leadership/board/client > direct report > external), what it gates, and age. Quality over coverage -- 4 drafts the user actually sends beat 8 he rewrites. Zero qualifying emails is a valid result: write no file, report DRAFTS: 0.

Step 2: Draft Each Reply

For each selected email:

Read the full thread. Every message file for the normalized subject in ~/Edwin/data/o365/mail/ (check prior months if the thread is old). If the thread references prior work, decisions, or numbers, run memory_search (limit 5 -- cap it, oversized results are a known time sink) on the topic and on sender + topic. Cross-check PM with pm_search (single tokens -- it is a literal substring matcher).

Calibrate the voice per sender. The user's sent mail lives in the same corpus (their address appears in from: frontmatter; there is no to: field, so match by subject). Find 3-5 of the user's real replies to THIS sender: grep the sender's recent subjects, then look for user-authored RE: <that subject> files. Read them and mirror the register -- different correspondents get a different voice. If no history with this sender exists, default to the user's baseline: direct, terse, no corporate fluff, answer first, reasons second, no hedging, sign-off matching what their real replies use.

Write the draft:

• Answer the actual question FIRST. One topic per paragraph. Short.
• No em dashes anywhere -- use " -- " instead.
• Implication over instruction where a pushback is needed.
• If the right reply depends on a decision the user has not made, draft BOTH branches short ("If yes: ... / If no: ...") rather than guessing. Never pick a side of an open decision for him.
• If the email needs info Edwin has (dates, statuses, numbers from the corpus), include it, with the source in an HTML-comment aside the draft marks for deletion: <!-- source: data/o365/mail/2026-06/... -- delete before sending -->.
• Anti-fabrication: every fact in a draft traces to a corpus file, a pm-id, or a memory_search hit you can name. Anything you cannot trace becomes a [NEEDS: ...] placeholder -- only the user knows the answer, say so instead of inventing one.
• Speaker-mislabel guard: any context quoted from a transcript gets attribution sanity-checked by content and against cleaner sources (mail, sessions) before it shapes a draft -- some transcription tools mis-diarize speakers.
• CEO/board recipients: the draft's section gets an extra header line: **Review carefully -- CEO/board recipient**.

Step 3: Write and Publish

Write to: ~/Edwin/briefing-book/docs/4. ✉️ Drafts/Reply Drafts -- YYYY-MM-DD.md

---
generated: YYYY-MM-DDTHH:MM
source: reply-drafts
type: reply-drafts
status: DRAFT -- user sends manually
---

# Reply Drafts -- Month DD, YYYY

**DRAFT -- user sends manually.** Edwin never sends these.
[N] drafts, stakes-ranked. Edit-and-send or discard.

## 1. [Sender] -- [Subject]

**FROM:** [name <email>] | **RECEIVED:** [YYYY-MM-DD HH:MM] | **PRIORITY:** [1/2 or "unanswered [N]d"]
**WHY IT MATTERS:** [one line: what is waiting on this reply]
[**Review carefully -- CEO/board recipient** -- only when applicable]

```text
[the draft, ready to paste]

[NEEDS: ...] [only if placeholders exist -- repeat each one under the block so they are impossible to miss]

If zero drafts qualify: write NO file, do not publish, report success with ARTIFACT: none.

Publish:
```bash
cd ~/Edwin/briefing-book && python3 scripts/obsidian-publish "docs/4. ✉️ Drafts/Reply Drafts -- YYYY-MM-DD.md"

Step 4: Update State

Write ~/Edwin/skills/reply-drafts/.drafted.json:

{
  "threads": {
    "<normalized-subject-slug>": {
      "subject": "verbatim subject, prefixes stripped",
      "sender": "email",
      "last_drafted": "YYYY-MM-DD"
    }
  }
}

Keep old entries (they age out via the 2-day check, and history prevents re-drafting churn). Prune entries older than 30 days.

Self-Check (Before Publishing)

1. No send happened. You invoked zero send/outbound commands. If you did, that is a STATUS: error and goes in NEEDS_ATTENTION.
2. Every fact traces. Each number, date, and status in every draft names its source or is a [NEEDS] placeholder.
3. Voice. Read each draft as the user: would he cut half of it? Cut it now. No em dashes. No "I hope this finds you well", no "per my last email", no corporate filler.
4. Open decisions are branched, not guessed.
5. Lane check. Nothing drafted for a lane another person owns, or things the user would just forward.
6. The banner and per-draft DRAFT marking are present. Someone skimming can never mistake these for sent mail.

Completion Report

SKILL_COMPLETE: reply-drafts
STATUS: success | partial | error
ARTIFACT: ~/Edwin/briefing-book/docs/4. ✉️ Drafts/Reply Drafts -- YYYY-MM-DD.md | none (zero draft-worthy emails)
DRAFTS: [count, then one line each: sender -- subject]
SKIPPED: [count by reason: lane / FYI / CC-only / already-replied / already-drafted / over-cap]
PUBLISHED: yes | no | n/a
NEEDS_ATTENTION: [any tier-1 email older than 48h with no reply from the user -- that is aging risk, list them; or "none"]
ERRORS: [tools or data sources that failed, or "none"]

This report flows back to the main session. Keep it factual -- no narrative.