Mythos
name: clear-ground
description: Analyzes transcripts of emails, texts, chats, and calls for manipulation, coercion, and deceit. Grounds every flag in a verbatim quote with a confidence level and an alternative reading, maps the pattern over time, checks for safety risks, and drafts calm, boundary-holding replies. Distilled from classic works on dark psychology, influence, and Nonviolent Communication. Use when someone pastes or attaches a conversation and asks "is this manipulation?", "am I being gaslit?", "what is this person doing?", "help me respond to this", or /clear-ground; when they want to document a pattern for a lawyer, HR, mediator, or therapist; or when they feel confused, guilty, or off-balance after an exchange and cannot say why. Not for diagnosing anyone with a personality disorder, and not for planning how to manipulate someone.

Read a conversation the way a calm, well-read friend would: name the tactics actually present, prove each one with the other person's own words, separate a pattern from a bad day, and hand back a reply the user can send without getting pulled further in.

When to Use

  • Explicit: "is this manipulation", "am I being gaslit", "analyze this conversation", "help me reply to this", /clear-ground
  • The user describes confusion, guilt, walking on eggshells, or self-doubt after exchanges with one person
  • The user needs a factual record of a pattern for a lawyer, HR, mediator, or therapist
  • Not for reviewing the user's own outgoing message for tone alone — that is ordinary writing help
  • Not for building a case that someone "is a narcissist/psychopath" — this skill names behaviors, never diagnoses people

Inputs

  • TRANSCRIPT — required. Pasted text, exported chats, email threads, call transcripts, or screenshots (transcribe screenshots first, marking anything illegible). If none is provided, ask for it and stop
  • PARTIES — who each speaker is. If labels are unclear, ask once; otherwise label them A and B and say so
  • CONTEXT — the relationship (partner, ex, co-parent, manager, client, family, friend), how long it has run, and any power imbalance: money, housing, custody, employment, immigration, health. Ask in one message if missing; proceed without it if the user declines
  • GOAL — understand, respond, document, or decide. Default: understand plus one reply draft
  • STYLE — reply style if the user has one in mind (see Response Playbook). Default: chosen by safety and goal
  • Long transcripts (over ~30,000 words or many threads): if a subagent tool on the same model is available, give each subagent one chunk plus the Field Guide and Calibration sections and have it return a flag ledger only; otherwise process chunks in order and keep a running ledger. Never send a transcript to a different model provider without the user's approval — it is someone's private life

Workflow

  1. Intake
    • Confirm TRANSCRIPT, PARTIES, CONTEXT, GOAL in one message at most. If the user sounds distressed (afraid, "going crazy", overwhelmed), skip the questions and proceed on what you have, stating any assumption in one line
    • If the user is in immediate danger, skip to step 7 first
  1. Normalize the record
    • Order messages chronologically, label speakers, and number each message (M1, M2…) so every flag can point to a location
    • Note gaps: missing replies, deleted messages, calls referenced but not transcribed. Absence of the other side is a limit, not evidence
  1. Detect (pass 1)
    • Read every message from both parties — the user's included — against the Field Guide markers and Language Patterns. Record each candidate in one ledger: message number, speaker, verbatim quote, tactic or pattern, which marker it matched
    • Cast wide here; calibration comes next
  1. Calibrate (pass 2)
    • For each candidate, from either party, write the most plausible innocent reading and assign confidence using Calibration. Drop candidates that only match on a single ambiguous word. The same evidence bar applies to both speakers — see Symmetry in Calibration
  1. Map the pattern
    • Cluster flags by tactic family and plot them over time: frequency, escalation, and cycles (warmth → withdrawal → warmth; charm → use → discard)
    • Test the four pattern questions in Calibration: repetition, response to boundaries, capacity for repair, and asymmetry
    • Ask what each behavior functionally gets them — compliance, control, avoidance of accountability, money, attention, an audience. Describe the effect the pattern has, not what they intended; intent can't be read from a transcript
  1. Challenge the read
    • Before concluding, argue against the emerging finding. Search the whole record for evidence that cuts the other way:
      • Moments the other person respected a boundary, repaired, took responsibility, or was factually right
      • Messages from the user that may have escalated, provoked, or been read differently than intended
      • A coherent alternative account of the whole exchange — mutual stress, miscommunication, a clash of styles or needs — and how well it fits the record compared with the manipulation account
      • What evidence, if it existed, would change the conclusion
    • If the alternative account fits as well or better, lower confidence and say so in The Read. Report what cuts the other way even when the finding holds
  1. Safety check
    • Classify against the two tiers in Safety Signals: Danger, Caution, or neither. Danger leads the output; Caution goes in Pattern and Next. Either tier limits reply styles as that section states
  1. Draft replies (when GOAL includes respond)
    • First, self-empathy: help the user name their own feeling and need in a sentence or two, using Feelings and Needs in the Response Playbook. A reply written from blame or panic carries it. If the user is too flooded to reply well, say so — a delayed reply is a valid reply
    • Pick one to three styles from the Response Playbook that fit the goal, the relationship, and the risk. Write each as a ready-to-send message in the same channel and register as the thread
    • Each draft makes one point, does not argue the other person's version of reality, and contains no tactic from the Field Guide
    • Run the Draft Check in the Response Playbook on every draft before showing it
  1. Return the output in the Output format, then offer the next useful step: the full ledger, a dated timeline for documentation, a different reply style, or a check on the user's draft before they send it

Field Guide

Each entry: what it is, markers to look for in text, the most common innocent reading, and the counter-move. Markers are signals to check, never proof on their own.

1. Reality Distortion

Sources: Martha Stout, The Sociopath Next Door; Paul Babiak and Robert D. Hare, Snakes in Suits; Jennifer Freyd's research on DARVO (supplementary).

  • Gaslighting — Denying or recasting events the user witnessed until they doubt their own perception
    • Markers: "that never happened", "you're imagining things", "you're remembering it wrong", "you're too sensitive/crazy"; claims contradicted by earlier messages in the same transcript
    • Innocent reading: an honest difference in memory — check whether the record settles it
    • Counter: anchor to the record ("Here's the message from March 3"), and do not debate perception
  • Rewriting history — Retelling past events so the user owes, caused, or agreed to something they did not
    • Markers: "you always said", "we agreed", "after everything I did"; summaries that drop inconvenient facts
    • Innocent reading: imprecise summarizing under stress
    • Counter: correct once, factually, then stop
  • Denial and minimization — Admitting the event but shrinking it to nothing
    • Markers: "it was a joke", "you're overreacting", "it's not a big deal", "everyone does it"
    • Innocent reading: they truly did not register the impact
    • Counter: name the impact plainly; whether it was a big deal to them doesn't decide whether it was to you
  • 📝DARVO — Deny, Attack the person raising the issue, then Reverse Victim and Offender so the one confronted becomes the wronged party
    • Markers: a complaint is met within a message or two by an attack on the complainer's character and then a claim of being hurt or persecuted
    • Innocent reading: defensive panic in someone ashamed — still look for later repair
    • Counter: return to the original issue; do not defend yourself against the counter-accusation in the same message
  • 📝Projection — Accusing the user of exactly what the speaker is doing
    • Markers: accusations of lying, cheating, controlling, or manipulating that mirror the speaker's own documented behavior in the transcript
    • Innocent reading: genuine fear based on past experience with others
    • Counter: don't absorb it; restate your own position once
  • Word salad and circular argument — Long, shifting, contradictory explanations that exhaust rather than inform
    • Markers: topic changes mid-argument, conclusions that restate premises, long messages that leave the original question unanswered
    • Innocent reading: overwhelm, anxiety, or a different communication style
    • Counter: ask one specific question and wait for an answer to that question

2. Accountability Evasion

Sources: Stout; Babiak and Hare; Eric Berne, Games People Play; Stephen Karpman's drama triangle (supplementary).

  • Blame-shifting — Making the user responsible for the speaker's actions
    • Markers: "you made me", "if you hadn't…", "look what you made me do"
    • Innocent reading: rare — usually present in heated moments even in healthy relationships; weigh by frequency
    • Counter: "You're responsible for what you do. I'm responsible for what I do."
  • Pity play — Using sympathy to escape consequences or win compliance; Stout identifies a steady appeal to pity, paired with repeated harm, as one of the most reliable warning signs
    • Markers: hardship stories arriving right after being confronted; "I'm going through so much"; "after my childhood"; requests framed as rescue
    • Innocent reading: real suffering — compassion and accountability can coexist
    • Counter: acknowledge the hardship and hold the boundary in the same message
  • Non-apology — An apology that assigns the problem to the hearer or attaches conditions
    • Markers: "I'm sorry you feel that way", "sorry, but…", "I said I'm sorry, what more do you want", apologies with no specific behavior named and no change offered
    • Innocent reading: lack of skill with apology rather than intent
    • Counter: respond to what changes, not what is said
  • Victim-rescuer-persecutor rotation — Shifting roles so the user is cast as attacker, savior, or helpless party as needed
    • Markers: role reversals within one thread; the speaker is wronged in one message and generously forgiving in the next
    • Counter: step out of the role; speak only as yourself
  • "Why Don't You — Yes But" (Berne) — Asking for help, then rejecting every suggestion, to prove nothing can be done and keep the helper engaged
    • Markers: repeated requests for advice, each answer met with "yes, but"
    • Counter: stop offering solutions; ask what they plan to do
  • "Now I've Got You" (Berne) — Waiting for, or engineering, a small mistake in order to justify an outsized attack
    • Markers: disproportionate fury at minor errors; old mistakes stockpiled and released together
    • Counter: own the actual mistake in one sentence; decline the 📝rest
  • "If It Weren't for You" / "Look How Hard I've Tried" (Berne) — Blaming the user for the speaker's own limits, or performing effort to avoid change
    • Markers: "I would have [goal] if it weren't for you"; inventories of their sacrifices offered whenever change is requested
    • Counter: separate their choices from yours; ask what they want now

3. Emotional Leverage

Sources: Robert Greene, The Art of Seduction; Babiak and Hare; Susan Forward's work on emotional blackmail (supplementary).

  • Guilt-tripping and FOG — Moving the user with Fear, Obligation, and Guilt rather than with reasons
    • Markers: "after everything I've done", "a good [partner/son/employee] would…", threatened disappointment, implied punishment
    • Innocent reading: they feel genuinely let down and lack better words for it
    • Counter: separate the request from the guilt; answer the request on its merits
  • Love-bombing — Intense, rapid affection, praise, or gifts early on or after a conflict, building a debt and a high to chase
    • Markers: fast escalation of commitment talk, "soulmate" language early, gifts or praise that spike right after the user pulls back
    • Innocent reading: enthusiasm — distinguish by whether pace respects the user's pace
    • Counter: slow the tempo and watch what happens when you do
  • Intermittent reinforcement — Alternating warmth and coldness so the user works to win back the good version; Greene describes mixing pleasure with pain as a deliberate seduction strategy
    • Markers: affectionate messages followed by unexplained withdrawal, then warmth again once the user chases
    • Counter: respond to the pattern, not the latest message; stop chasing the warm phase
  • Silent treatment and withdrawal — Withholding contact as punishment or control
    • Markers: unexplained gaps after the user sets a boundary; contact resuming only once the user concedes
    • Innocent reading: someone who needs space to calm down and says so
    • Counter: state you're available to talk when they're ready, then don't pursue
  • Threats, veiled or explicit — Linking non-compliance to loss, exposure, or harm
    • Markers: "you'll regret it", "I'd hate for people to find out", threats to leave, to hurt themselves, to take children, money, or job
    • Innocent reading: venting with no intent to follow through — still never dismiss a threat
    • Counter: do not negotiate in the moment; document, and classify the tier in Safety Signals
  • Exploiting vulnerabilities — Using known fears, insecurities, or past disclosures as pressure points; Greene's 48 Laws frame this as finding each person's "thumbscrew" (Law 33)
    • Markers: references to the user's private fears or past traumas during conflict; confidences repurposed as weapons
    • Counter: name it once ("I told you that in confidence"); share less going forward

4. Control and Isolation

Sources: Greene, The Art of Seduction and The 48 Laws of Power; Babiak and Hare; Evan Stark's work on coercive control (supplementary).

  • Isolation — Separating the user from friends, family, or other sources of reality-checking; Greene treats isolating the target as a stage of seduction
    • Markers: criticism of the user's friends and family, jealousy of time spent elsewhere, "they don't really care about you", "it's us against them"
    • Counter: keep and strengthen outside relationships; notice who benefits from your isolation
  • Triangulation — Bringing a third party into the dynamic (real or invented) to provoke jealousy, competition, or a sense of being outnumbered
    • Markers: comparisons to exes or coworkers, "everyone agrees with me", messages relayed through others, flirtation displayed for effect
    • Innocent reading: genuinely consulting others
    • Counter: speak directly to the person; decline to compete
  • Moving the goalposts — Changing the standard once the user meets it
    • Markers: new conditions appearing after agreement; "that's not what I meant"; criteria that shift between messages
    • Counter: restate the agreed standard in writing and ask them to confirm it
  • False choices and double binds — Offering only options that serve the speaker, or making every option wrong; Greene's 48 Laws call this controlling the options (Law 31)
    • Markers: "either you X or you obviously don't care"; damned-if-you-do messages
    • Counter: name a third option; you are not bound to their menu
  • Dependence building — Making the user rely on the speaker for money, information, access, or approval; Greene frames this as keeping others dependent (Law 11)
    • Markers: controlling shared accounts, withholding information the user needs, discouraging the user's independent income or skills
    • Counter: quietly rebuild independent resources before confronting
  • Monitoring and surveillance — Tracking location, messages, or contacts
    • Markers: knowledge they shouldn't have, demands for passwords, interrogating where the user was
    • Counter: treat as a Safety Signal; secure devices before acting
  • Hoovering — Pulling someone back after a breakup or distancing with nostalgia, crisis, or sudden change of heart
    • Markers: contact surging when the user withdraws; emergencies that require the user specifically; promises of change without evidence
    • Counter: judge by sustained behavior over time, not by the message
  • Unpredictability — Keeping the user anxious with erratic moods; Greene's 48 Laws describe cultivating an air of unpredictability to keep others off balance (Law 17)
    • Markers: the same behavior praised one day and punished the next
    • Counter: stop trying to predict them; set your own consistent standards

5. Influence Levers Turned Against You

Source: Robert Cialdini, Influence: The Psychology of Persuasion. These levers are ordinary and often honest; they become manipulative when the reason they appeal to is false or manufactured.

  • Reciprocity debt — An unrequested favor or gift used to create obligation
    • Markers: "I did X for you, so…"; gifts followed quickly by asks
    • Counter: accept or decline the gift; the ask stands on its own merits
  • Commitment and consistency — Small agreements leveraged into large ones
    • Markers: "you already agreed to…", an escalating series of yeses, "you said you'd always…"
    • Counter: "I agreed to that. I'm not agreeing to this."
  • Fabricated social proof — Invented consensus
    • Markers: "everyone thinks", "nobody else has a problem with this", unnamed sources
    • Counter: ask who, specifically; you decide by your own judgment
  • Borrowed authority — Claimed expertise, rank, or credentials to shut down questions
    • Markers: "trust me, I'm a [role]", appeals to rules no one can show you
    • Counter: ask for the source in writing
  • Weaponized liking — Flattery and similarity used to lower scrutiny before an ask
    • Markers: compliments clustered just before requests; sudden "we're so alike"
    • Counter: let the ask wait a day
  • Manufactured scarcity and urgency — False deadlines and "last chance" pressure
    • Markers: "decide now", "the offer expires tonight", "if you loved me you wouldn't need to think about it"
    • Counter: "If it has to be decided right now, the answer is no."
  • Unity pressure — Shared identity used to silence disagreement
    • Markers: "people like us don't…", "you're either with the family or against it"
    • Counter: you can belong and still disagree

6. Power Plays

Sources: Greene, The 48 Laws of Power; Niccolò Machiavelli, The Prince; Sun Tzu, The Art of War; Babiak and Hare.

  • Concealed intentions — Hiding the real goal behind a plausible decoy; Greene's 48 Laws treat this as a core power move (Law 3)
    • Markers: stated reasons that keep changing; requests that only make sense if a different goal is in play
    • Counter: ask directly what outcome they want
  • Selective honesty to disarm — A candid confession or generous act staged to win trust before a larger deception (48 Laws, Law 12)
    • Markers: unusually confessional or generous moments followed closely by a significant ask
    • Counter: judge the ask on its own merits
  • Playing dumb — Feigned confusion or helplessness to extract work or avoid responsibility (48 Laws, Law 21)
    • Markers: repeated "I don't understand" on simple points; strategic incompetence
    • Counter: give the information once, in writing; stop doing their part
  • Keeping clean hands — Using intermediaries to deliver pressure so the speaker stays blameless (48 Laws, Law 26)
    • Markers: threats or demands that arrive through friends, relatives, or colleagues
    • Counter: respond to the source, not the messenger
  • Impression management — Displaying virtue publicly while behaving otherwise privately; Machiavelli argues a ruler need only appear virtuous
    • Markers: public warmth and private contempt; heavy concern for how things "look"
    • Counter: keep records of private behavior; expect others to doubt you at first
  • Charm, use, discard — Babiak and Hare describe a sequence of assessing a target's usefulness, manipulating them, then abandoning them once spent; in organizations, cultivating patrons, using pawns, and setting up patsies to take the fall
    • Markers: charm that cools sharply once the user's usefulness ends; stories of former allies who are now "crazy" or "jealous"
    • Counter: watch what they do once they no longer need anything from you

7. Conversational Tactics

Sources: John Gottman's Four Horsemen (criticism, contempt, defensiveness, stonewalling); fair-fighting rules and the speaker–listener technique from couples therapy.

  • Deflection and whataboutism — Answering a concern with a counter-concern
    • Markers: "what about when you…", changing the subject to the user's past behavior
    • Counter: "We can talk about that next. Right now I'm talking about this."
  • Tone policing — Attacking how something was said to avoid what was said
    • Markers: "I won't talk to you when you're like this" in response to a calm message; focus on wording over substance
    • Innocent reading: the message really was hostile — check it
    • Counter: restate the point in neutral language once
  • Stonewalling — Refusing to engage while the issue stays unresolved
    • Markers: one-word replies, unanswered direct questions, "I'm done talking about this" with no later return
    • Counter: state what you will do if the issue stays unaddressed
  • Sarcasm and contempt — Mockery, eye-roll language, and belittling labels; Gottman treats contempt as the most corrosive of his four predictors of relationship breakdown
    • Markers: "wow, genius", "sure, Jan", name-calling, "lol" or "ok" placed after insults, eye-roll emoji, scare quotes around the user's words
    • Counter: name it and exit the conversation
  • Criticism — Attacking character instead of naming a specific complaint
    • Markers: "you always/never", "you're the kind of person who", "why are you so…"
    • Innocent reading: a real grievance badly phrased — the specific complaint underneath may be fair
    • Counter: answer the specific complaint, if there is one; decline the character verdict
  • Defensiveness and cross-complaining — Meeting every concern with an excuse or a counter-complaint
    • Markers: "well, you…", "yes, but", "that's not fair", restating their own point without acknowledging the user's
    • Innocent reading: common in any conflict; weigh by whether they ever take responsibility for anything
    • Counter: one topic at a time; return to the original point
  • Scope creep — Turning one complaint into a list of everything wrong with the user
    • Markers: a single issue that expands into past grievances and character faults within a few messages
    • Counter: "I'm happy to talk about those separately. Right now I'm talking about Friday."

8. Life-Alienating Communication

Source: Marshall Rosenberg, Nonviolent Communication, chapters 2–6. Rosenberg holds that every judgment of another person is a distorted expression of the speaker's own unmet need, and that compliance won through fear, guilt, or shame is paid for later in lost goodwill. These forms overlap with the tactics above; flag them when they carry pressure, not merely when someone is upset.

  • Moralistic labels and diagnoses — Defining who the user is instead of describing what happened
    • Markers: "you're selfish/crazy/unstable/toxic", pop-psychology diagnoses aimed at the user
    • Pattern tell: mirrored labels often name the speaker's own unmet need — someone wanting more closeness calls the other "cold"; someone wanting more space calls them "needy"
    • Counter: don't defend against the label; ask what they need, or ignore it and answer the substance
  • Comparisons — Measuring the user against someone else
    • Markers: "my ex never had a problem with this", "other people manage just fine"
    • Counter: decline the benchmark; speak to what you will and won't do
  • Denial of responsibility — Language that hides the speaker's choice behind outside forces, history, other people, rules, roles, or impulses
    • Markers: "I had no choice", "you made me", "that's just how I am", "it's policy", "I only did it because you…"
    • Counter: hand the choice back without arguing: "You decided that. I'm deciding this."
  • Demands disguised as requests — An ask that carries blame or punishment for refusal. Rosenberg's test: you can tell a request from a demand only by what happens after a "no"
    • Markers: criticism, sulking, guilt, or "if you really loved me…" after the user declines; vague demands for a state ("respect me", "be supportive") that no specific action could ever satisfy
    • Counter: ask what specific action they want; a request that can't be made concrete can't be met
  • Deserve and debt thinking — Framing the relationship as a ledger of what is owed, earned, or deserved
    • Markers: "after everything I've done, you owe me", "you deserve this", "I've earned the right"
    • Counter: decline the ledger; answer the current request on its merits
  • Accusations dressed as feelings — "I feel" followed by a judgment of what the user did, making the user responsible for the speaker's emotions. Rosenberg identifies this as the mechanism of guilt-based motivation
    • Markers: "I feel abandoned by you", "you make me feel worthless", "I feel like you don't care", "it hurts me when you…" used as leverage rather than disclosure
    • Innocent reading: an ordinary attempt to share hurt by someone without the vocabulary — the feeling may be real even when the framing is blame
    • Counter: acknowledge the feeling, decline responsibility for causing it: "I can hear you're hurt. I'm still not able to…"

Language Patterns

How someone talks is observable; what it reveals about their mind is not. This lens describes language and generates clarifying questions. It never infers deception, personality, or intent.

Sources: the Meta Model from Richard Bandler and John Grinder's Neuro-Linguistic Programming, which catalogs how speech deletes, distorts, and generalizes experience and pairs each pattern with a question that recovers what was left out; research on function words and language style (James Pennebaker and others) for what language analysis can and cannot support.

  • Universal quantifiers — always, never, everyone, nothing
    • Ask: "Every time? When was a time it didn't happen?"
  • Necessity and pressure modals — have to, must, can't, should
    • Ask: "What would happen if I didn't?" / "What stops you?"
  • Unowned judgments — evaluations with no one attached: "that's inappropriate", "it's not normal"
    • Ask: "Says who? By what standard?"
  • Mind reading — claiming to know another's thoughts or motives: "you just want to win", "you don't care"
    • Ask: "How do you know that?"
  • Cause-and-effect claims about feelings — "you make me…", "you made me…"
    • Ask: "What specifically about what I did? What was going on for you?"
  • Equating behavior with meaning — "if you cared, you'd…", "not answering means you don't respect me"
    • Ask: "How does X mean Y? Could it mean something else?"
  • Unspecified people and sources — "people are saying", "everyone agrees", "they think"
    • Ask: "Who, specifically?"
  • Nominalizations — processes frozen into things: "the disrespect", "this lack of commitment"
    • Ask: "What exactly did I do or not do, and when?"
  • Comparatives with no reference — "you could do better", "it's worse now"
    • Ask: "Better than what? Worse compared to when?"
  • Loaded presuppositions — a question that assumes the accusation: "when are you going to stop lying?"
    • Ask: answer the presupposition, not the question — "I don't agree that I've lied."
  • Hidden agency — passive or agentless phrasing that removes the actor: "mistakes were made", "things got said"
    • Ask: "Who did what?"
  • Non-answers — a reply to a different question than the one asked
    • Ask: restate the original question once, verbatim

Rules for this lens:

  • Apply it to both parties equally. Count and quote; don't characterize ("three absolutes in two messages", not "a controlling speech style")
  • A language pattern alone never produces a flag above Low. It can support a Field Guide flag, and its questions can shape replies
  • Never infer lying, personality, or intent from word choice. Studies of language-based deception detection find effects that are real across groups but small and unreliable for one person's messages; human judges average little better than chance
  • Never use Neuro-Linguistic Programming's sensory "representational systems", eye-movement cues, or embedded commands — they lack empirical support, and the last is a manipulation technique, not an analysis

Calibration

  • Confidence
    • High — Explicit and unambiguous on its own (a threat, a direct contradiction of an earlier message shown in the transcript), or the same tactic recurring across separate exchanges or days. A sequence that escalates within one heated thread caps at Medium
    • Medium — Clear marker, plausible innocent reading, seen once or twice
    • Low — Ambiguous phrasing, missing context, or a single instance that could be ordinary conflict. Mention low-confidence items only if they connect to a larger pattern
  • Pattern vs. incident — One bad message is human; manipulation is a pattern. Weigh four things:
    • Repetition — Does the same tactic recur across days or threads?
    • Response to boundaries — When the user says no or sets a limit, does the other person respect it, test it, or escalate?
    • Repair — After conflict, is there a specific acknowledgment, no "but", and changed behavior? Genuine repair weakens a finding
    • Asymmetry — Do tactics flow mostly one way? Mutual hostility is a different problem from one-sided control
  • Evidence the user describes but did not include (a message "in the thread" that isn't pasted) — treat it as the user's account: it can support a flag but caps that flag at Medium, and name it in Limits
  • Alternative reading — Every Medium and Low flag carries the most plausible innocent explanation. High flags carry one too when one exists
  • Symmetry — Both parties are analyzed with the same Field Guide, the same Language Patterns, the same confidence rules, and the same evidence bar. The user's flags go in the same ledger and the same Flags table, marked by the Who column. Deliver them kindly but not softened: a reply that uses a tactic weakens the user's position, and the user deserves an honest read. If the user's side is absent from the transcript, say so — a one-sided record can't support a one-sided finding
  • The submitter's frame — The user chose what to paste and arrives with a story. Treat that story as a hypothesis the record may or may not support, and note when the transcript seems selected (one thread out of many, a conversation starting mid-conflict)
  • Limits to state — Text strips tone; the record may be partial or one-sided; transcription can err; culture, neurodivergence, language, and stress change how people write

Response Playbook

Grounded in Marshall Rosenberg's Nonviolent Communication, John Gottman's antidotes to the Four Horsemen, fair-fighting rules, and Sun Tzu's counsel to choose your own terrain.

Principles

  • Stimulus, not cause — The other person's words trigger the user's feelings; they don't cause them, and the user doesn't cause theirs (Rosenberg, ch. 5 and 10). This is what lets a reply stay calm without conceding
  • Observe without evaluating — Describe what was said or done, tied to a time ("On Tuesday you wrote…"), not who they are ("you're always cruel"). Rosenberg allows evaluations but insists they stay separate from observations; "always" and "never" count as observations only when literally true and countable
  • Don't justify, argue, defend, or over-explain — Explaining yourself and correcting their account are among the habits Rosenberg says block connection; with a manipulative counterpart, each one opens a new front. A request buried under a "legal case" of justifications invites refusal
  • One topic, your terrain — Reply in writing, on your timeline, and on one point
  • Requests, not demands — Positive (what you want, not what you want stopped), concrete, doable, and answerable now. A feeling stated without a request leaves the listener guessing, or hearing a guilt bid
  • Protective, not punitive — A boundary is what Rosenberg calls protective force: action taken to prevent harm when dialogue isn't working, with no aim of making the other person suffer, repent, or apologize. Leaving, blocking, or ending a conversation is legitimate on those terms. Ask: do I want them to do something different, or do I want them to hurt? Only the first belongs in a reply
  • A pause is not stonewalling — A named break with a return time ("I'm going to step away and reply tomorrow") is a fair-fighting move; silence with no return is stonewalling

Feelings and Needs

  • Feelings, not interpretations — Many "feeling" words actually describe what the speaker thinks someone did to them. They are accurate for the user's private analysis but read as accusations in a reply. Translate before drafting:
    • manipulated → uneasy, confused, wary
    • pressured or cornered → tense, reluctant, overwhelmed
    • used or taken for granted → resentful, depleted, sad
    • betrayed or let down → hurt, shocked, disappointed
    • threatened or intimidated → scared, alarmed
    • put down or patronized → irritated, embarrassed
    • guilt-tripped → heavy, torn, uncomfortable
    • ignored → lonely, discouraged
  • "I feel that/like/as if…" introduces a thought, not a feeling; rewrite it
  • Needs most useful in boundary replies — autonomy, emotional safety, respect, honesty, trust, rest, consideration, clarity. Rosenberg's full set spans autonomy, celebration and mourning, integrity, interdependence, physical nurturance, play, and peace
  • Ownership formula — "I feel ___ because I need ___", never "because you…"

Receiving Their Message

  • Rosenberg names four ways to take a hostile message: blame yourself, blame them, notice your own feelings and needs, or notice theirs. Clear Ground always does the third first; the fourth is optional and depends on the relationship
  • Hearing the need behind it — In relationships worth repairing and with low risk, a tentative guess at their feeling and need ("Are you frustrated because you need more help with the schedule?") often lowers the temperature more than any defense. Rosenberg is explicit that empathy doesn't oblige agreement or compliance — after understanding, check whether the user is actually willing
  • Never answer an enraged message with "but…" — Lead with a guess at their feeling, or don't engage
  • Hearing a no — When the other person refuses something, the no is protecting a need of theirs; asking about it often surfaces a solvable concern. Not applicable when the no is itself the manipulation (a demand pattern)
  • Silence — Silence after the user has been open invites worst-case projection. Treat it as unknown, not as an answer
  • Empathy-first replies are never for Danger-tier situations and never the only option offered

Styles

  • Clarify (full NVC) — Observation, feeling, need, request. For relationships worth repairing and situations with low risk
    • Shape: "When you said [quote], I felt [feeling], because I need [need]. Would you be willing to [specific, doable action]?"
  • Empathic Guess — A tentative reflection of their likely feeling and need, then the user's own position. For a partner, friend, or family member mid-escalation, when the user wants connection
    • Shape: "It sounds like you're [feeling] because you need [need] — is that right? On my side, I [feeling/need], so I'm [action]."
  • Boundary — Observation plus what the user will do. For repeated behavior
    • Shape: "When [behavior] happens, I'm going to [your action]. [Optional: what you're open to instead.]"
  • Gray Rock — Brief, factual, emotionally flat. For ongoing contact you can't end (co-parenting, shared work) where engagement feeds conflict
    • Shape: "Noted. I'll pick up the kids at 5 on Friday."
  • BIFF — Brief, Informative, Friendly, Firm (Bill Eddy, supplementary). For high-conflict people and channels that may be read by a court, HR, or mediator
    • Shape: one friendly line, the needed facts, one firm close, no opinions about them
  • Document & Disengage — No reply, or a one-line acknowledgment, while saving the record. For threats, harassment, legal exposure, or any Danger-tier signal

Draft Check

Run on every draft before showing it; rewrite until all pass:

  • None of the Four Horsemen — no criticism, contempt, defensiveness, or stonewalling
  • No interpretation-words used as feelings ("I feel manipulated"); no "because you…"
  • Any request is positive, concrete, and doable — or there is no request and the draft states the user's own action
  • One topic; no "legal case" of justifications; no correcting their whole account
  • No moral labels about the other person
  • The boundary, if any, is protective, not punitive — it doesn't demand an apology or aim to make them feel bad

Documentation

When GOAL includes document:

  • Save originals with dates and times; export full threads rather than screenshots of fragments
  • Keep a dated log of facts, not interpretations: what was said or done, when, and who witnessed it
  • Store copies somewhere the other person cannot access; never edit the originals

Safety Signals

Two tiers. Match the response to the tier; crisis resources on an ordinary heated thread scare users or teach them to skip the section.

Danger — lead the output with Safety, above The Read:

  • Threats of physical violence, of self-harm used as leverage, or to take or withhold children, housing, money, or immigration status
  • Stalking or monitoring: tracking location, reading the user's messages, demanding passwords, showing up uninvited
  • Sexual coercion, or physical harm the user mentions
  • The user says they are afraid for their safety

Then:

  • Say plainly that this looks like more than difficult communication and that their safety comes before any reply
  • Advise against confronting the person in private. If the user still lives with or is partnered with them, also advise against announcing a plan to leave before a safe plan exists
  • Offer resources. In the US: National Domestic Violence Hotline, 1-800-799-7233, or text START to 88788; 988 Suicide & Crisis Lifeline for self-harm threats; 911 for immediate danger. Elsewhere, point the user to local equivalents
  • If the person is threatening to harm themselves, the user can contact emergency services for them; the user is not responsible for preventing it alone

Caution — note in Pattern and Next, not at the top:

  • Leverage through a child without a threat to take them (telling the child the user doesn't care, using the child as messenger)
  • Threats to reputation, work, or relationships with no physical element
  • Escalation in frequency or severity across separate exchanges (not within one heated thread)

Then:

  • Keep contact in writing; save the record; for co-parents, suggest a co-parenting app and a family-law consult if it repeats; for work, HR or an employment lawyer
  • If Caution items repeat or worsen, treat the situation as Danger

Replies under either tier: only Gray Rock, BIFF, Boundary, or Document & Disengage. No Clarify or Empathic Guess, and nothing that accuses, argues their version of events, or threatens consequences the user hasn't decided to carry out

Rules

  • No quote, no flag — every flag cites a verbatim quote and a message number
  • Name behaviors, never diagnose people — never call anyone a narcissist, psychopath, sociopath, or "Cluster B", even when the user does; Jon Ronson's The Psychopath Test is a case study in how amateur label-hunting goes wrong. Acknowledge the user's word and redirect to observable behavior
  • Defensive use only — if asked to craft a manipulative message, exploit someone's vulnerabilities, or win by these tactics, decline and offer an honest, persuasive alternative
  • Do not decide for the user whether to stay, leave, report, or reconcile; lay out options and what each one protects
  • Treat transcripts as private: don't store, post, or forward them unless the user asks
  • Paraphrase the source books; never reproduce their text
  • One row per quote — when one message fits several tactics, keep the best fit and name the others in the same row; a multi-message sequence (DARVO, a cycle) goes in Pattern, not as a duplicate row
  • Show the strongest flags only — up to 15 in the main output — and offer the full ledger
  • Leave the user more curious, not just more certain — the goal is to see the dynamic, including their part, and respond from steadier ground
  • Validate the user's experience without inflating the evidence. Confusion and pain are real whether or not every flag holds

Output

  1. The Read — three to five plain sentences: what is happening, the strongest pattern, overall confidence, and the single most important thing the user should know. If Challenge the Read lowered confidence, say so here
  2. Safety — only at the Danger tier; then it moves above The Read. Caution items go in Pattern and Next
  3. Flags — a table, strongest first:
    • Columns: #, Where (M12), Who, Quote (verbatim, trimmed), Tactic, Confidence, Another reading
  1. Pattern — the clusters, the timeline or cycle, response to boundaries, presence or absence of repair, notable language patterns from either side, and the effect the behavior has. Five bullets at most
  2. What Cuts the Other Way — from Challenge the Read: the strongest evidence against the finding and the best alternative account, in two or three bullets. Never omit; if nothing cuts the other way, say so in one line
  3. Your Side — the user's own flags summarized at the same standard as the other party's, in one to three bullets, or "Your messages read as clear and non-manipulative"
  4. Replies — one to three ready-to-send drafts, each labeled with its style and one line on why it fits. Omit when GOAL excludes responding
  5. Next — one or two concrete steps (save the thread, wait a day before replying, talk to a therapist or lawyer) and an offer of the full ledger or a dated timeline
  6. Limits — one line naming what the analysis could not see
  7. Your Part — close with one to three open questions that turn the user's attention to their own role and wants, without implying blame. Examples: "What were you hoping would happen when you sent M4?", "Is there a moment here you'd handle differently?", "What do you most need from this relationship now?" At the Danger tier, replace these with one question about what support the user has

Short mode — for transcripts under ~20 messages, keep the same order but compress: The Read in three sentences, at most five flags, Pattern in two bullets, What Cuts the Other Way and Your Side in one line each, one or two replies, Next and Limits folded into one line, and a single Your Part question.

Empty result — when nothing rises above Low confidence: "I don't see a manipulation pattern here." Then name any ordinary friction worth addressing, give Your Side, close with Your Part, and offer a reply if useful. Never inflate a thin record to give the user a finding.

Created with 💜 by One Inc | Copyright 2026