A colleague replies to your message with one line and no punctuation. A friend does not react to news you were excited to share. A partner cancels plans for the second time. Something happens, a feeling may show up fast, and an interpretation can follow before you have checked what is known and unknown.
Separating an observable event from an interpretation can reduce the chance of acting on an unverified assumption. The event may itself be hurtful, and a repeated pattern can matter even when another person’s intent is unknown. The workflow below does not explain away the event; it records what happened, what you felt, what you inferred, and what remains unknown before you respond.
The focus here is the ordinary version of that moment: everyday friction, not a crisis. A model can help you slow down and separate the feeling from the story, in the same way a good friend might ask “wait, what actually happened, versus what do you think it means?” It cannot tell you what someone else intended, and it should never be asked to.
This workflow is for ordinary emotional friction — an ambiguous message, a disappointment, a minor conflict. It is not for abuse, coercion, self-harm, crisis, or trauma. If any of those apply, close the chat and contact a person: a trusted friend, a manager or HR if the situation is at work, or a professional. If you or someone else may be in danger, contact local emergency services now. For suicidal thoughts, the International Association for Suicide Prevention’s crisis-resource finder lists services by country — use that or your local emergency number, not a chatbot.
Why feelings and interpretations need different columns
The feeling is real information about your experience. It does not by itself establish another person’s intent, and it should not be dismissed merely because the interpretation is uncertain.
The story is the causal explanation you build on top of the feeling, often about someone else’s intent: they’re annoyed with me, they don’t care, this always happens. Interpretations can combine current facts with expectations and past experience. A language model also fills gaps from patterns in text, so its “most likely explanation” is a plausible completion, not a verified fact.
The four-column separation
Before responding to whatever triggered the reaction, write down four things:
- Facts — only what a camera or transcript would show. “She replied ‘ok’ with no other text” is a fact. “She’s annoyed with me” is not.
- Feelings — name them plainly. Hurt, embarrassed, anxious, angry, dismissed. One or two words each; resist the urge to justify them yet.
- The story — the explanation you have already half-decided is true. Write it exactly as it appears in your head, including the unfair parts.
- Missing information — what you do not actually know that the story assumes. Did they see the message at a bad time? Are they dealing with something unrelated? Is one-word terse actually their normal style?
Doing this on paper or in a private note is often enough. A model becomes useful at the next step: generating alternative stories you would not have generated alone, so your first draft is not the only one on the table.
A prompt that widens the story instead of confirming it
Weak prompt:
My coworker replied "ok" to my message and I think she's mad at me. What should I do?
This prompt hands the model your conclusion and your framing. It will likely produce sympathetic validation, a few “you might be overthinking it” caveats, and a suggested message — none of which tested whether the story is right.
Stronger prompt:
Here is what happened, only the facts: [paste facts].
Here is the story I've already told myself: [paste story].
Generate three alternative explanations that are at least as plausible as mine, including at least one that has nothing to do with me. For each, name what additional information would confirm or rule it out. Do not tell me which one is true — you don't have enough information, and neither do I yet.
Illustrative example output:
- She’s dealing with something unrelated right now. Would be confirmed by: she mentions a busy day later, or her other messages today are also short.
- “Ok” is just her normal texting style. Would be confirmed by: checking past messages from her — are one-word replies a pattern, not a one-off?
- She read it on her phone between meetings and meant to follow up. Would be confirmed by: a longer reply arriving later, or her bringing it up in person.
Your original story stays on the table as one more hypothesis to test — for example, that she is annoyed about something specific — usually with a direct, calm question rather than more guessing.
The output is not a verdict. It is a wider set of hypotheses than the one you started with, each attached to a way to check it. That is the actual value: it interrupts the single-story trap, without pretending to know what happened.
Keep the facts section free of other people’s private details you would not want stored by a third-party tool — full names, health information, or anything that identifies someone who has not consented to being discussed. Describe the situation in enough detail to be useful without turning it into a data-retention problem for someone else.
A pre-committed pause
For an ordinary, non-urgent interaction, choose a pause before replying when you notice that you are reacting to an unverified story. There is no evidence-based universal 24-hour rule: use the real deadline, safety context, and effect on the other person. Do not use delay to avoid a necessary conversation indefinitely.
Instead:
- Complete the four-column separation.
- Generate alternative stories.
- Pick the smallest next step that tests reality rather than assumes it — usually a short, calm, specific question, not a long explanatory message defending your feelings or accusing them of theirs.
- Re-read when you are able to consider more than one explanation. If the issue still matters, raise the observable event and a direct question with the person involved.
This is not a rule against feeling things quickly, and it is not a requirement to delay a clear safety boundary. For ordinary, non-dangerous ambiguity, it is a reminder to distinguish an interpretation from verified information before acting on it.
A second scenario, and where the workflow bends
Not every situation fits the same shape. Consider a partner canceling plans for the second time in a month.
Facts: plans confirmed Tuesday, canceled Thursday morning by text, second cancellation in four weeks, no reason given in the text.
Feelings: disappointed, a little dismissed, wary.
The story: “They don’t prioritize time with me the way I prioritize time with them.”
Missing information: what actually happened Thursday morning; whether the first cancellation had a clear reason that this one lacks; whether this is a pattern or a rough month.
Run the widening prompt, and the alternative explanations will likely include something ordinary (a work emergency, poor communication habits rather than low priority, difficulty asking for space) and something that does deserve attention (an actual pattern of deprioritizing the relationship). The workflow does not resolve which one is true — that requires information only a direct conversation can provide.
This is the point where a one-off tool stops being enough. If the four columns keep producing the same story about the same person across multiple incidents, that recurrence is itself the signal, not something a fresh round of alternative-story generation should paper over. At that point the honest move is a direct conversation — “I’ve noticed this has happened twice this month, can we talk about it” — not a third or fourth pass through the worksheet hoping a new angle appears. The weekly review covers how to track a recurring pattern like this across time with the same evidence discipline used here for a single incident.
What this cannot do
A model can offer vocabulary for a feeling you are struggling to name, and it can generate alternative interpretations you would not have thought of alone. It cannot determine the other person’s intent, cannot verify what is true about a relationship, and should never be asked to adjudicate whether your feelings are “valid” — that is a question for you and the people in your life, not for a chat window.
If the same story keeps recurring across many situations with the same person, or if the feeling is disproportionate to the trigger in a way that concerns you, that pattern is worth bringing to a therapist or counselor, not to more rounds of prompting. A model that is patient, always available, and never pushes back is a poor substitute for a person who can actually challenge you. When reflection becomes a reassurance loop covers the warning signs of leaning on a chatbot for that role.
Try it once, low-stakes
Pick a minor irritation from this week — not the biggest one, the smallest one worth examining. Run the four columns. Generate three alternative stories. Notice how many of them you would not have considered without the prompt forcing the count. That gap is the whole point of the exercise: not a better feeling, but a wider set of explanations before you act on the first one.
The worksheet below has the four columns plus the alternative-story prompt and a pause field as a single page you can reuse.
