A Claude Skill that tells you what a message is actually saying, and shows its work.
You got the email. You've read it four times. Something about it is off and you can't say what, so you send it to a friend and ask "am I reading too much into this?"
That question has a real answer, and it's findable. Careful messages leave marks. A standard element that's missing. Hedges clustered in one place and absent everywhere else. A question you asked that quietly didn't get answered. One flat sentence in a warm paragraph. These are things you can point at and quote, which means they can be checked.
subtext reads incoming messages against that evidence and sorts every finding into what's stated, what the structure supports, and what's a guess. The third category gets labeled and kept in its own box.
Then it writes the plain version. Your message with the management removed.
Anything that reads hidden meaning into text can turn into a machine that makes anxious people more anxious. Two rules stop that.
No finding without a quote. Every claim names the exact feature that produced it, and states what that feature does not prove. Most bad subtext reading is a real observation that grew a cause on the way to the conclusion.
The innocent reading is mandatory. Before anything ships, it argues the boring explanation as hard as it can. They were busy. It's a template. Legal wrote it. That's how they write to everybody. Every finding has to survive that or it gets demoted to speculation.
Which means sometimes the answer is that the email is fine and you're tired. It will say that.
git clone https://github.com/MichaelDoesMarketing/subtext.git ~/.claude/skills/subtextOr grab the packaged .skill file from releases and install without cloning.
"Got this rejection after four rounds. Is there anything in it or am I reading too much into it?"
"Here's how my manager wrote to me in March, and here's yesterday. What changed?"
"Client hasn't said no but hasn't said yes. Here's the last three emails."
Four modes. Read it for the full analysis. Is this a no? when you want the verdict and not the essay. What changed for comparing messages from the same person over time, which is the strongest evidence there is. Check my paranoia when you're already convinced something's wrong and you want the innocent case argued at full strength before you act.
The test that matters isn't running this on something obviously shady. It's running it on something everybody agreed was honest, and seeing whether it manufactures a problem.
So: Brian Chesky's May 2020 letter laying off 1,900 people at Airbnb. Widely held up as the best-handled layoff communication in tech. Source.
THE SHORT ANSWER
Unusually direct for the genre. One sentence in it isn't, and it's the one that decided who got the letter.
IN THE STRUCTURE
Specificity gradient. Every number that determines an outcome is exact. "14 weeks of base pay, plus one additional week for every year at Airbnb." "12 months of health insurance through COBRA." "Around 25% of our company."
The criteria that determined who received the letter is the vaguest phrase in the document: decisions were made "based on critical skills, and how well those skills matched our future business needs."
Supports: the compensation was settled and defensible, and the selection basis was not something the company was going to state in writing. Does not support: that the selection was arbitrary or unfair. Vagueness about criteria is not evidence about criteria.
Agency, and this one cuts the other way. "We did a comprehensive review of every team member and made decisions." Active voice, named actor, in a genre that runs on "a decision was made." He owns it grammatically where almost nobody does. That's a real finding in the positive direction and it's reported as one.
Register. The letter is personal throughout, opening with "This is my seventh time talking to you from my house." The one institutionally-voiced sentence in the whole thing is the selection criteria sentence. Three separate techniques converge on the same sentence, which is what turns an observation into a finding.
THE INNOCENT READING
Layoff letters are lawyered. Selection criteria are kept non-specific in essentially every one ever written, because specificity creates discrimination exposure. The vagueness is genre-standard, not concealment.
Survived, demoted: the finding holds as a seam, not a secret. That sentence is where legal review becomes visible in a letter that was otherwise written by a person. The suspicious version of this finding does not survive and is dropped.
THE PLAIN VERSION
Mostly the same letter. Which is the actual result.
The skill's headline conclusion on the most famous "good" layoff email is that it's good, with one seam. That's the demo, and it's better than a takedown, because the thing you need to trust about a tool like this is that it doesn't find something every single time.
SKILL.md the workflow, modes, evidence tiers, rules of engagement
references/techniques.md 12 readable features and what each one actually proves
references/genres.md standard shapes for rejections, reviews, layoffs, passes, notices
references/evidence.md the three tiers, the innocent reading, confidence calibration
references/paranoia.md the inverse mode, and when to tell someone the text is clean
assets/read-template.md output template
assets/baseline.md what to collect about a sender before reading them
It reads messages you received. It won't analyze private communications between other people to give you leverage over them, and it won't build a case against someone from their correspondence.
It reads text, not people. "This sentence avoids naming who decided" is a finding. "He's a coward" is not, and it isn't checkable either.
Anything with legal consequence gets read as structure and then handed to an actual lawyer.
battle-tested writes the message. versus picks the fight. subtext reads what came back.
MIT. Use it, fork it, ship it.