Meat and Glass
Two tags. Inline attribution. No guessing.
You got an email from someone who uses AI. Half of it sounds like them. Half of it sounds like a press release. You don't know which half to trust.
That's the problem. Not that people use AI to write — they should. The problem is you can't tell where the human stops and the machine starts.
Here are two tags that fix it.
The tags
(m) Meat
Human fingers typed this. A brain decided these exact words. "m" is for me. The "eat" is optional if the other word bothers you.
(g) Glass
AI generated this. A language model wrote these words under human direction. "g" is for generated — or glass, if you've read the novel.
Drop them anywhere in your text. Everything after a tag is that voice until the next tag appears. Like a toggle switch. Need to mention the tags without switching? Put them in quotes.
Why this matters
AI is good at some things and confidently wrong at others. When you can see the tags, you know what to trust and what to verify:
| (m) Meat — more reliable | (g) Glass — more reliable | |
|---|---|---|
| Dates and times | Trust it — the human checked the calendar | Verify — AI hallucinates dates |
| Business context | Trust it — the human is living it | Verify — AI doesn't know your org |
| Current events | Trust it — the human read the news today | Verify — AI has a knowledge cutoff |
| Technical explanations | May oversimplify | Trust it — AI is precise about systems |
| Arguments and reasoning | May have blind spots | Trust it — AI considers all sides |
| Historical facts | Memory is fallible | Trust it — AI's training data is deep |
The tags don't tell you what's true. They tell you what kind of true to expect. A human asserting a date? Probably checked. An AI asserting a date? Roll the dice.
What it looks like
Here's an email using (m)/(g) notation. Watch the colors:
To: Product Team
Subject: Q3 data platform migration — updated timeline
(m) Hi team. Quick update on the migration. I'm tagging this email with “(m)” and “(g)” so you can see what's me and what's AI. “(m)” = my words. “(g)” = Claude's words. If I need to fix something Claude got wrong, I'll flip to “(m)”, fix it, and flip back.
We're pushing the Snowflake cutover to August 18. The vendor confirmed Friday that the connector won't be ready until the 15th, and I want three days of buffer. Priya is on PTO the first two weeks of August so Ali will cover her workstreams. Here's the technical summary:
(g) The migration involves three phases. Phase 1 (schema alignment) maps the existing Redshift table structures to Snowflake equivalents, handling differences in data types, clustering keys, and distribution strategies. Phase 2 (parallel load) runs both systems simultaneously with write-path duplication to validate that query results are identical across platforms. Phase 3 (cutover) redirects all read traffic to Snowflake and decommissions the Redshift cluster after a(m) 14-day monitoring window. Legal signed off on 14 last quarter, not 30. And we're not decommissioning the cluster — we're downgrading it to a dc2.large cold standby at $48/month instead of the current $2,400. Back to Claude:
(g) The cost reduction from the Redshift downgrade combined with Snowflake's consumption-based pricing is projected to reduce the team's infrastructure spend by approximately(m) 35% at current volumes. Claude said 40 but doesn't know about the BI dashboard refresh we added in March. I'll share the full cost model Thursday.
Next steps: Ali runs the schema alignment script this week. I'll send the cutover checklist Monday. Questions, reply here.
— JK
What just happened
The human opened with context and dates. Handed the technical summary to the AI. Then corrected the AI twice — in place, in front of everyone:
- AI wrote "30-day monitoring window." Human flipped to
(m), wrote 14 days, explained why, flipped back. - AI wrote "40% cost reduction." Human flipped to
(m), wrote 35%, explained the missing dashboard spend.
No back-and-forth. No debate. The human edits the glass in place, like a copy editor with a red pen. You can see exactly where the pen landed.
The rules
- Start with
(m). The human opens. Introduce yourself, set context, explain the tags if it's someone's first time. - Switch to
(g)for heavy lifting. Technical explanations, structured arguments, summaries — let the AI draft what it's good at. - When the AI gets something wrong, flip to
(m)and fix it in place. Don't rewrite the(g)block. Don't add a correction below. Cut into the glass where the error is, write the truth, flip back when you're done correcting. - End with
(m). Action items, deadlines, next steps — the human always closes. - Quoting the tags. Need to talk about the tags without using them? Put them in quotes: "(m)" and "(g)". Quote marks = mention. Bare tags = action.
The opposite direction
It works for incoming text too. Reading a tagged document, you know which claims to challenge and which to take at face value. You stop arguing with the AI's phrasing and start checking the human's dates.
The tags are a trust protocol. Not trust in the machine. Trust in the process — the human is present, paying attention, and willing to correct the machine in public.
Getting started
Type (m) at the top of your next email. Write your intro. When you hand it to AI to draft the body, put (g) where the AI starts. When you edit the AI's draft, tag your edits with (m).
That's it. Two characters in parentheses. No software. No plugin. No training session. Works in Outlook, Gmail, Slack, Teams, a markdown file, a sticky note.
The people reading your emails already wonder how much of it is you. Now they can see.
Disclosure: The prose on this page was generated by Claude (Anthropic) under Bill's direction. The ideas, the (m)/(g) notation, and the trust model are his. The example email is fictional but based on real usage. He reviewed and approved every word but did not type them all. Full transparency, always.