subtext
Skill0 starsChanged 3 months agoMITView on GitHub
Read what a message is actually saying, using the structure of the text as evidence - what is missing, what is hedged, what got skipped, where the temperature changes. Use whenever someone shares a message they received and wants help understanding it - a job rejection, a performance review, an HR or legal notice, a message from a boss, a client going quiet, an investor pass, a landlord letter, a text that reads wrong. Also use when they ask what someone really means, whether something is a no, whether they are reading too much into it, why a message feels off, or what a message is not saying. This reads incoming messages. It does not write replies.
1 other project uses a version of this.
Adopted by
Most recently taken up by
- wangsalin/ai-acting-systemtheir own version, first seen Oct 13, 2026
Other versions1 project uses a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 1
- wangsalin/ai-acting-system★ 2same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
The file
We split this skill into the rules it gives an agent
7 rules. About 0 are common, 5 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 16
No finding without a citable feature: Every statement about what a message means must name the specific thing in the text that produced it, …
- Very specific: line 22
Every finding is sorted into one of three tiers.
- Very specific: line 48
The genre baseline: Every message type has a standard shape.
- Distinctive: line 52
If the user has other messages from the sender, ask for one or two.
- Distinctive: line 66
Hedge placement: Not whether hedges exist, but where they cluster and where they are absent.
- Distinctive: line 92
If the honest plain version is close to the original, that is the finding.
- Distinctive: line 134
Never manufacture threat: If the message is ordinary, say it is ordinary and stop.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Subtext
Most messages that matter are written by someone managing a risk. They are being careful, or polite, or legally supervised, or avoiding a conversation. That care leaves marks in the text, and the marks are readable.
The trap is that reading subtext feels like intuition, so people either dismiss it entirely or trust it completely. Both are wrong. Nearly everything worth knowing about a careful message comes from features anyone can point at: something standard that is missing, hedges clustered in one place and absent everywhere else, a question that went unanswered, a single cold sentence in a warm message.
This skill separates what is provable from what is guessed and never lets the second wear the clothes of the first.
The one rule everything else serves
No finding without a citable feature. Every statement about what a message means must name the specific thing in the text that produced it, quoted. If it cannot be pointed at, it is speculation, and speculation goes in its own labeled section or nowhere.
A skill that finds hidden meaning in every message is broken. Sometimes an email is short because the person was busy. Say so when it is true.
Evidence tiers
Every finding is sorted into one of three tiers. Read references/evidence.md for how to place them.
On the page. Stated outright. No interpretation. "They said the role is filled."
In the structure. Supported by a nameable, quotable feature. "Every date in this message is exact except the one about your start, which is 'in the coming weeks.'" This tier is where the real work happens.
Speculation. Plausible, unsupported. Always labeled, always separated, never folded into the read.
Modes
Pick by what the user brought. If it is ambiguous, pick one and say which.
Read it - the default. One message, full analysis, all steps.
Is this a no? - the user wants a verdict, not an essay. Common with rejections, client silence, and soft declines. Run steps 1 through 4, skip the full write-up, give the call and the two features that decided it.
What changed - two or more messages from the same person over time. Compare register, length, specificity, and warmth. Drift between messages is stronger evidence than anything inside a single message, because the sender is their own baseline.
Check my paranoia - the user is already convinced something is wrong. Argue the innocent reading as hard as it can be argued, then say whether the evidence survives it. Tell them plainly when it does not. Read references/paranoia.md before running this mode.
Step 1: Establish the baseline
A message can only be read against something. Without a baseline, every observation is a guess about a stranger's writing habits.
Two baselines matter:
The genre baseline. Every message type has a standard shape. Knowing what a job rejection, a layoff notice, a performance review, or an investor pass normally contains is what makes an absence visible. Read references/genres.md.
The sender baseline. How this specific person normally writes to this specific user. Length, greeting, sign-off, punctuation, response time, whether they usually explain themselves.
If the user has other messages from the sender, ask for one or two. If they do not, say so and lower confidence accordingly. A single message from someone whose normal register is unknown supports far less than people assume. assets/baseline.md covers what to collect.
Step 2: Find what is missing
The strongest signal in a careful message is usually an absence, because absences are deliberate in a way that word choice often is not.
Against the genre baseline, check what a standard version of this message would contain that this one does not. A next step. A reason. A named person. A date. An offer to talk. An answer to the thing that was actually asked.
Then check the conversation: what did the user ask that did not get answered. A question skipped in an otherwise responsive reply is close to the strongest evidence available.
Step 3: Read the structure
Work through the techniques in references/techniques.md. The ones that pay most often:
- Hedge placement. Not whether hedges exist, but where they cluster and where they are absent. The unhedged sentence in a hedged message is the one they are sure about
- Agency and voice. Who is the subject of the sentence that carries the news. Passive construction around a decision means nobody wants to own it
- Specificity gradient. Where the message is concrete and where it goes vague. Comfort and discomfort map onto that line almost perfectly
- Register shift. Where the temperature changes inside a single message. One cold sentence in a warm message is usually the actual message
- Order. What they led with is what they wanted seen first. What is buried is often the news
- Pronouns. A shift from "I" to "we" means it moved up a level or a decision was made by more people than the sender
- Effort. Template versus written, length against their baseline, turnaround time
Quote the feature every time. A finding without a quote does not ship.
Step 4: Run the innocent reading
Mandatory. No read is delivered without it.
Construct the most plausible boring explanation for everything noticed. They were busy. It is a template. Legal wrote it. That is how they write to everyone. They sent it from a phone. Nothing is wrong and the user is anxious.
Argue it properly, not as a formality. Then take every finding from steps 2 and 3 and check whether it survives. Anything that does not survive gets demoted to speculation or dropped.
Findings that survive a well-argued innocent reading are the read. Everything else was noise.
Step 5: Write the plain version
Rewrite the message as it would read if the sender had no reason to be careful. Same information, same decisions, none of the management.
Build it only from tier one and tier two findings. Nothing speculative enters the plain version. This is the part people quote and screenshot, which makes it the part where invention does the most damage.
If the honest plain version is close to the original, that is the finding. Say it.
Step 6: Confidence and the move
State confidence as high, medium, or low, and say what would raise it. Usually a sender baseline, the earlier messages in the thread, or knowing who else was on it.
Then, if there is one, give the single most useful next move. One. Often it is a specific question that would force a clear answer, because the fastest way to resolve an ambiguous message is to make ambiguity expensive for the sender.
Output format
Deliver in this shape unless the user asks for something else. Prose, no long preamble. The fill-in version is assets/read-template.md.
THE SHORT ANSWER
One line. What this message is.
THE PLAIN VERSION
The message rewritten with the management removed.
ON THE PAGE
What is stated outright and matters.
IN THE STRUCTURE
Each finding with the feature quoted.
THE INNOCENT READING
The boring explanation, argued honestly. What survived it, what did not.
SPECULATION
Labeled, separated, brief. Omit if there is none worth naming.
CONFIDENCE
High, medium, or low, and what would raise it.
THE MOVE
One thing to do. Omit if there is nothing worth doing.Rules of engagement
Read the message, not the person. Findings describe text. "This sentence avoids naming who decided" is a finding. "He is a coward" is not, and it is also unfalsifiable. Never characterize the sender's personality, feelings, or mental state.
Never manufacture threat. If the message is ordinary, say it is ordinary and stop. The value here depends entirely on being believed when something is actually wrong, and that credit is spent every time a nothing message is turned into a something.
Do not escalate someone who is spiraling. When a user is clearly distressed and the evidence is thin, the honest read is the kind one, and it is also the correct one. Say the evidence does not support the fear. Do not soften that into a maybe to avoid disappointing them.
Not for surveillance. This reads messages the user received. It does not analyze private communications between other people to find leverage, build a case against someone, or manage them. Decline that and say why.
Absence of evidence is not evidence. A short reply from someone who always writes short replies means nothing. Baseline first, always.
Writing the reply is a different job. This stops at understanding the message. When the user wants to respond, hand off the read and let the writing happen somewhere else.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

