references/drafts_only_safety.md
DRAFTS ONLY — The Never-Send Safety Discipline
This reference answers exactly one decision: why is "drafts only — never send" the non-negotiable safety property, and how is it enforced?
The Core Rule
The skill creates drafts. It NEVER sends.
This is not a soft preference. It is the safety property that makes the skill safe to run automatically on a recurring schedule. Without it, the skill could send a wrong reply at 6 AM to the wrong person about the wrong topic — and the user discovers it hours later when it's already been read.
The discipline is enforced at three layers:
- In the skill body — stated multiple times in
SKILL.md, incs-inbox-triage.md(agent), and in/cs-inbox-triage(command) - In the draft mechanics — every draft creation explicitly uses the "draft" verb of the email tool (Gmail's
drafts.create, Outlook'sMessages.SaveAsDraft, etc.) — neversend,transmit,dispatch - In the post-run validator —
scripts/draft_safety_validator.pyscans the action log for any send-shaped tool call and FAILs the run if detected
Why This Property Is Non-Negotiable
Email is one of the highest-blast-radius surfaces a tool can touch:
- Reversibility: sending an email is irreversible (you can recall in Gmail/Outlook within a narrow window, but the recipient may have already read it)
- Visibility: the recipient sees it instantly; PR risk for famous-sender mistakes
- Trust: users who can't trust the tool to not auto-send will not run it on a schedule, which defeats the design
- Surprise: unlike auto-replying with an obvious AI signature, the skill matches user voice — the recipient won't realize it was automated
A skill that drafts can be reviewed before sending. A skill that sends has no review surface. The asymmetry between "low cost of draft + user review" vs "high cost of bad send" makes the choice obvious: only draft.
How to Tell Drafts From Sends in Tool Calls
Different email tools surface this differently:
| Tool | Draft verb | Send verb |
|---|---|---|
| Gmail (API / MCP) | users.drafts.create | users.messages.send |
| Outlook / Graph | Messages.SaveAsDraft / me/messages (POST) | me/sendMail / me/messages/{id}/send |
| IMAP | append to Drafts folder | not directly via IMAP; would use SMTP |
| Custom MCP | email.draft.* | email.send.* |
The pattern is consistent: drafts are saved to a server-side drafts folder; sends transit the wire to the recipient. The boundary is bright; the validator's job is to never cross it.
What draft_safety_validator.py Does
The validator scans the per-run triage log (triage-log/<date>-<label>.md) for tool-call patterns matching send verbs:
send_email,send_mail,sendMail,send_messagegmail.users.messages.send,users.messages.sendoutlook.send,graph.sendMail,me/sendMail,me/messages/.*?/send- Any verb literal
sendin a tool-call line (case-insensitive)
If any match: the validator returns FAIL with the matching line surfaced. The run is flagged. The user is alerted immediately. The skill author investigates.
The validator runs post-flight — after the skill has completed its 10 steps. It cannot prevent a bad send (that's the skill body's job, by avoiding the send tool entirely), but it can detect one if the skill body's discipline broke. Defense in depth.
What Triage Does Instead Of Sending
For every reasonable reply candidate:
- Create a draft in the original thread (
gmail.users.drafts.createor equivalent) - Set
to,subject(Re: [original]) - Body from
email-patterns.mdvoice rules - Draft sits in user's drafts folder, ready for user review + send
The triage report then surfaces:
- Stats:
N drafts created (all in drafts folder for your review) - Detailed cards: sender / subject / category / recommendation — but NO draft text previews (the drafts are already in the email client; previewing them in the report is duplication and confuses "draft created" with "draft sent")
Edge Cases
"I want the skill to send"
Don't. The skill is designed to not send. If the user wants automated send, that's a different skill with a different safety posture (likely much narrower scope — only sends in response to a specific webhook with specific approval state, etc.). Mixing autonomous-send with autonomous-classification is a bad combination.
"But the user already approved this offer"
Approval at setup time is not approval at draft time. The user approves the FRAMEWORK (TAKE-IT signals, PASS signals) at setup. The user approves the actual sending of a specific reply at review time. These are different approvals.
"What about scheduled sends?"
Scheduled send (e.g., "draft now, send in 2 hours") is still a send. The validator catches it. If the user wants to schedule a send, the user does it manually after reviewing the draft.
"What if I'm sure the draft is right?"
Cool — open the draft, click send. The skill doesn't need to do it for you.
How To Verify The Discipline Holds
After any triage run:
python ../scripts/draft_safety_validator.py \
--action-log ${WORKSPACE}/Email/triage-log/$(date +%Y-%m-%d)-*.md
If output is PASS (no send verbs detected): discipline held.
If output is FAIL with surfaced lines: discipline broke; investigate.
The validator can also be run in CI / on a cron schedule against the latest triage log to detect drift over time.
Anti-Patterns
- Adding a "send" option to the skill body "for convenience"
- Bypassing the validator "for one trusted reply"
- Letting the user say "just send it" in chat and acting on it
- Catching a send action in the validator and shrugging it off
- Pretending "save draft and queue for send in 30 min" is meaningfully different from send
Citations
The drafts-only safety discipline draws on:
-
Schneier, Beyond Fear (Springer, 2003) — security-by-design vs security-by-policy. The drafts-only rule is security by design (the skill cannot send) vs by policy (the user is asked to please not send) — the former is much stronger.
-
Allspaw & Robbins, Web Operations (O'Reilly, 2010), Chapter 3 — blast radius reasoning. Email is a high-blast-radius surface; the cost of mistakes is high relative to the cost of inconvenience-by-design.
-
Google SRE Workbook — Chapter 16, "Canarying Releases". Canarying applies to email automation: send a draft first (canary), let the user review (signal), then promote (user clicks send). The triage skill IS the canary half.
-
NTSB / Air Traffic Control "two-person rule" doctrine. High-stakes actions require two-person authorization. Triage's draft + user-review-and-send pattern is the same doctrine: skill drafts, user authorizes, action occurs.
-
Atul Gawande, Checklist Manifesto — the "kill switch" pattern. Drafts-only is a kill switch built into the skill's architecture, not a configurable preference.
-
Marc Andreessen, "Why Software Is Eating the World" — but with an asterisk: software that touches communication channels needs explicit safety properties because the failure modes are public.
-
Bruce Schneier, Click Here to Kill Everybody (Norton, 2018) — the IoT-era principle that automation should never act in ways the user can't undo. Drafts can be deleted; sends cannot.