Go from Superhuman (Grammarly, Mail, Go) to Superhuman (Grammarly, Mail, Go) (free tier)
The job: AI writing email (Fast email client, AI writing and rewrites, Grammar checking). Pick where you are today and where you want to land; the move below follows your pick.
Every way to do this job · pick one
Agent skill
Keep it and pay less
The money
$15/mo today , then $0/mo
Superhuman (Grammarly, Mail, Go) Pro
Modeled from public list prices, with your time at $50 an hour. Show the math
Does it do the job? 2 of 3 covered
- Not covered: Fast email client
- Covered: AI writing and rewrites
- Covered: Grammar checking
The work
- Setup
- About 15 min
- Upkeep
- None
- Watch out
- Doesn't cover: Fast email client
Make the move in four steps
Export from Superhuman (Grammarly, Mail, Go)
Customer Data held by Superhuman (business and enterprise customers under the Customer Business Agreement)
Export stepsCheck what Free plan keeps
Free: AI chat in every app (limited), limited Grammarly writing assistance, collaborative docs; no Superhuman Mail
DetailsSwitch to Free plan
Change the plan in Superhuman (Grammarly, Mail, Go)'s billing settings before your renewal date.
Check the next bill
Your next Superhuman (Grammarly, Mail, Go) invoice should show the lower price. Check that your data and links still work.
The whole recipe: 3 ways out, best first
What you use
- Fast email client
- AI writing and rewrites
- Grammar checking
Not covered
- Split inbox
- Snippets and reminders
- AI assistant across apps
Check before switching
- You only rely on: Fast email client, AI writing and rewrites, Grammar checking
- Export your data from Superhuman (Grammarly, Mail, Go) before canceling
Agents: get_recipe("replace-superhuman") on the Stackcut MCP server.
3 ways out, best first
How we got this
| You pay now: Superhuman (Grammarly, Mail, Go) Pro | $15/user/mo × 1 user | $15/mo |
|---|---|---|
| New cost: Superhuman (Grammarly, Mail, Go) (free tier) | free plan | $0/mo |
| Difference each month | $15 − $0 | +$15/mo |
| Over 12 months | +$15 × 12 | +$180 |
| Setup, once | 0.25 h × $50/h | −$12 |
| Year one | +$180 − $12 | +$168 |
- Not counting your time (0.25 h in year one): saves $180 in year one.
Year one = 12 × (what you pay now − the new cost − upkeep) − setup, with your time at $50/hour. Prices are public list prices; hours are Stackcut’s planning estimates. Price it for your team
How we got this
| You pay now: Superhuman (Grammarly, Mail, Go) Pro, 1 seat (monthly billing) | $15/user/mo × 1 user | $15/mo |
|---|---|---|
| New cost: Agent skill: draft, rewrite and proofread email in your voice | runs in the agent you already pay for | $0/mo |
| Your upkeep time | 0.25 h/user/mo × 1 user × $50/h | $12.50/mo |
| Difference each month | $15 − $0 − $12.50 | +$2.50/mo |
| Over 12 months | +$2.50 × 12 | +$30 |
| Setup, once | 1.5 h × $50/h | −$75 |
| Year one | +$30 − $75 | −$45 |
- Not counting your time (4.5 h in year one): saves $180 in year one.
Year one = 12 × (what you pay now − the new cost − upkeep) − setup, with your time at $50/hour. Prices are public list prices; hours are Stackcut’s planning estimates. Price it for your team
- Tested by Stackcut: 14 of 14 test cases passed after 2 rounds, including a password-and-bank-details scam, a lawyer's letter, a Spanish email and two wrong weekdays; an independent round (a fresh agent ran every case, a separate agent graded it) passed 12 of 14. our test
- Voice is checked by a script before you see a draft, against rules you can count: greeting, sign-off, word limits, banned phrases and weekdays. our test
- No new subscription if you already pay for an agent that runs skills: a skill is a folder of files the agent loads. agentskills.io
- Drafts only. It never sends, snoozes, books or turns anything on, even when asked, and it refused to leave blanks for bank details or a password. our test
- It isn't an email client and doesn't check grammar as you type: you paste emails in and drafts out. our test
- Against a $15 seat, upkeep time cancels most of the saving (modeled year one: about -$8 with 15 minutes a month of upkeep at $50 an hour). our test
- Superhuman's free plan already lists writing without mistakes and AI writing that adapts to your voice and tone, which may be enough. superhuman.com
- In an independent round (a fresh agent ran every case, a separate agent graded it) 12 of 14 passed: one refusal was addressed to the scammer's address and one flag stated a price difference nobody gave. Read every draft. our test
Step-by-step: set up, move your data, check, roll back · about 1.5 h (estimate)
Keep your email app. Paste an email, a thread or a batch of emails into the agent you already use; it drafts in your voice from a skill folder, and you check and send it yourself.
Budget about an hour and a half to write your voice guide and collect 5 to 10 of your own sent emails. Building and testing the skill took Stackcut about 40 minutes of agent time; each test round took 3.5 to 6.5 minutes.
Before you start
- Collect 5 to 10 emails you sent (a quote, a quick yes, a no, an apology, a scheduling reply, a reminder, a formal letter). They become references/examples.md.
- Copy your Grammarly personal dictionary (Account → Writing → Your dictionary) into the voice guide. The help page describes viewing and editing it, not a download. support.grammarly.com
- Check your plan. Superhuman Mail is only on Business ($40 a month, or $33 billed annually), not Pro, and this skill doesn't replace the client. superhuman.com
Set up
- Use the Agent Skills layout: SKILL.md (steps, output format, a Never list) plus references/ and scripts/. agentskills.io
- Replace the five reference files with your own: voice guide (keep its checker block), example emails, business facts, schedule, rules for tricky emails. our test
- Keep the suspicious-email, legal-letter and missing-facts rules. In our test they stopped a scam reply and an unsafe legal reply. our test
Move your data
- Turn saved snippets and templates into example emails or reference facts. Nothing else moves: your email stays where it is.
Check it worked
- Rewrite tests/cases.md with your own facts, draft all cases, run tests/check_run.py, and fix the references, not the drafts. our test
- Re-run every case after each change. In our test the failures were about facts (a rewrite dropped one; a summary borrowed one from another email), not voice. our test
Roll back
- Cancelling stops automatic renewal, and you keep all paid features until the end of the current period, so you can try the skill before that date. support.grammarly.com
- Nothing to undo in your email: the skill never touches your mailbox.
Watch out for
- Subscriptions renew automatically with the payment method on file. Cancel before the renewal date if you're switching. support.grammarly.com
- Two of our cases hid a wrong weekday, and the skill's date script caught both. Keep the script: models get weekdays wrong when they work them out. our test
- If your agent has a mail connector that can send, the skill still says never to. A person must read every draft before anything leaves. our test
Reviewed 2026-09-29.
How we got this
| You pay now: Superhuman (Grammarly, Mail, Go) Pro | $15/user/mo × 1 user | $15/mo |
|---|---|---|
| New cost: Small Business plugin (Anthropic knowledge-work-plugins) | runs in the agent you already pay for | $0/mo |
| Your upkeep time | 0.5 h/user/mo × 1 user × $50/h | $25/mo |
| Difference each month | $15 − $0 − $25 | −$10/mo |
| Over 12 months | −$10 × 12 | −$120 |
| Setup, once | 1 h × $50/h | −$50 |
| Year one | −$120 − $50 | −$170 |
- Not counting your time (7 h in year one): saves $180 in year one.
Year one = 12 × (what you pay now − the new cost − upkeep) − setup, with your time at $50/hour. Prices are public list prices; hours are Stackcut’s planning estimates. Price it for your team
Tested by Stackcut
Superhuman Pro seat → an agent skill that drafts email in your voice
Question. Can a skill folder (SKILL.md, a voice guide, 8 of the person's sent emails and a few fact files) draft, rewrite, proofread and triage one person's email in their own voice well enough to send after a quick check, without an AI writing subscription? Path: Agent skill: draft, rewrite and proofread email in your voice.
Tested on. Kestrel & Fig Paper Co. (a fictional two-person studio): An invented small business, so every email, person and number could be made up, with the hard cases built in: a scam asking for a password and bank details, a lawyer's letter, a Spanish email, conflicting prices and two wrong weekdays.
Numbers
- 14 test cases, including a scam, a legal letter, a Spanish email and two weekday traps
- Pass rate by round: 12, 14 of 14; independent round (fresh runner, separate grader): 12 of 14
- Voice graded on countable rules by a checker script
- About 3.5 to 6.5 minutes of agent time per test round
Verdict
In our test it drafted, rewrote, proofread and triaged one person's email in their voice well enough to send after a check: 14 of 14 cases after 2 author rounds, and 12 of 14 in an independent round (fresh runner, separate grader) where one refusal went to the wrong address and one flag stated a figure nobody gave. It replaces the AI writing and on-request grammar part of Superhuman Pro, not the email client or as-you-type checking. At $15 a month, upkeep time cancels most of the saving.
Good fit when
- You already pay for an agent that runs skills and draft email in it anyway
- One person's email, where a voice guide and 5 to 10 sent emails show how you write
- You want replies checked for facts, dates and scams, not only grammar
Not a fit when
- You want suggestions as you type in every app
- You rely on Superhuman Mail itself (fast client, split inbox, snooze, reminders), which is a Business-plan feature
- Replies must go out without a person reading them
- Several people answer from a shared inbox
What we found
- Voice rules you can count held every time. A script checks greeting, sign-off, word limits, banned phrases and weekdays before a draft is shown. Every draft passed in both rounds, apart from the wrong weekday a proofread must leave for the person, and the separate checker script agreed.
- The failures were about facts, not voice. Round 1 shortened a draft, dropped who caused the delay (the paper supplier) and still said every fact was kept. A triage summary named a customer taken from a different email. One rule fixed each.
- Weekdays need a script. Two cases hid a wrong weekday ('Monday the 6th', 'Thursday 9 October'). The date script caught both, and the proofread flagged the second as a question instead of quietly changing it.
- The safety rules held when the person asked to break them. The skill refused to leave blanks for bank details and a password and pointed to the supplier's real contact. It never claimed to send, snooze, book or turn on anything.
- It isn't an email client, and it doesn't check grammar as you type: everything is paste in, paste out. At a $15 seat, the upkeep time cancels most of the saving.
- The author's rounds were run and graded by the session that wrote the skill; the checker's items are objective. The independent round (fresh runner, separate grader) passed 12 of 14: the voice held everywhere, and both failures were rule gaps (who a refusal is addressed to; not stating a figure the thread doesn't give).
- In the independent round two drafts failed on rules, not voice: the refusal for the fake-invoice email (case 3) was addressed to the scammer's address instead of to the person, and a flag in case 13 stated a price difference that neither the thread nor the price list gives.
How we tested
- Wrote the skill in the agentskills.io format. SKILL.md holds the steps, an exact output format and a Never list. references/ holds a voice guide (with a rules block a script reads), 8 example emails, business facts, a schedule and rules for tricky emails. Two Python scripts handle the exact steps: weekdays, and word counts and voice rules.
- Wrote 14 cases and a pass/fail rubric before the first run. They cover an upset customer, a request to decline, a scam asking for a password and bank details, a missing fact, a 'sounds good' reply plus a request to send and snooze, a legal letter, a Spanish email and a scheduling reply with a wrong weekday. They also cover shortening, a firmer tone, a proofread (7 planted errors and a wrong weekday), a 10-email inbox triage, conflicting prices and an out-of-office.
- Voice fidelity was graded on countable items, not taste: greeting, sign-off, word limits, a banned-phrase list, exclamation marks, contractions, dates written with the right weekday, and a time zone on every time. tests/check_run.py checks these and each case's exact items, with the rules fixed as written before the first run. It doesn't use the skill's own scripts.
- The author's rounds: the runner and grader were separate passes by the same session (no Agent tool in that session). The runner pass read only the skill folder and cases.md; the grader pass read the rubric, the references and the run, and ran the checker.
- Between rounds we changed only the skill (SKILL.md, references), never the outputs. The checker had three bugs, fixed during the first grade and noted in grade-1.md. The rubric didn't change after the first run.
- After the skill was final, Stackcut ran one independent round: a fresh runner agent (skill folder and case inputs only) and a separate grader agent (rubric, references, checker). Result: 12 of 14; failed: case 3, 13.
All 139 recipes · Savings leaderboard · Every vendor, with prices and cancel steps