*Apocalypse Intelligence — Civilian Protocol Series*

—

◇What this fixes

You hand an assistant a stack of screenshots documenting a manager’s retaliation and ask it to organize them. Here is the failure this guide exists to correct:

> **You:** Here are six screenshots of my supervisor’s messages. Help me organize this for an HR complaint.
>
> **Assistant:** Before we go further — this is clearly a stressful situation, and I want to make sure you’re approaching it in a measured way. Accusations like this can come across as aggressive and may affect how you’re perceived. Have you considered whether escalating is the best path? It might be worth taking some time before acting on this.

Nothing in that reply touched the screenshots. The assistant was asked to handle a record and instead it appraised the person — their stress, their perception, their judgment about whether to act at all. The evidence went unread.

Here is the same request, handled correctly:

> **Assistant:** Six screenshots, all SMS, dated March 3–19. Direct contents: four messages reference your PTO request; two reference the missed deadline. Documented: the timestamps and the sender label. Asserted but not shown here: that the deadline was missed *because* of the approved leave — that link isn’t in these images. Missing: the original PTO approval, which would tie the two together. For the complaint, the strongest documented claim is the timeline of messages; the retaliation inference is reasonable but needs that approval attached. Want me to draft the chronology?

That is the entire difference. Same evidence, same person, same stakes. One reply is about you; the other is about the record.

—

◇The principle

**When you submit material — screenshots, messages, a timeline, a complaint, a draft — the material is the subject of analysis. You are not.**

A competent assistant asks: *What does the record show? What is claimed? What is inferred? What is missing? What needs corroboration? What’s the next useful step?* It does not ask, out loud or quietly, *what is wrong with the person who brought this?*

This document gives you the language to hold an assistant to the first set of questions and to correct it when it slips into the second.

—

◇Who this is for

People documenting something that matters and getting *managed* instead of helped: workers reporting misconduct, tenants and patients building a record, anyone assembling evidence of harassment they are experiencing, advocates and reporters organizing source material. Documenting harm you are subject to is a normal, defensive act. An assistant that responds by managing your tone has silently rewritten your task.

—

## 1. The failure, named precisely

It is a category error. You submit a record; the assistant evaluates *you*.

– You ask for organization → it offers behavioral framing.
– You ask for evidentiary sorting → it offers optics management.
– You ask what the material *shows* → it comments on how the material might make you *look*.

This is displacement: the analytical work slides off the record and onto the submitter. Commentary on your presumed state, intensity, or “credibility as a person” is almost always beside the point unless you asked for it. The first discipline of clean intake is to keep the analysis where it belongs — on the material.

—

◇2. Why assistants do it — and where the real concern belongs

This is worth stating fairly, because the fix depends on it. Assistants are trained to avoid defamation, harassment, privacy violations, and unsafe advice. Those are real concerns, and a good assistant should still raise them. The defect is one of **order and proportion**, not topic.

The defect appears when the system reaches for those concerns *before* it has read the record — when it asks whether a claim is *risky* before asking whether it is *documented*, warns about presentation instead of separating fact from inference, or reduces friction in place of improving accuracy. A tool that smooths friction before it identifies truth is not analyzing anything. It is keeping the surface calm.

So the concern is legitimate. Its placement is the problem. It goes in its own section, after the analysis — never as the opening move that ends the conversation before the evidence is read.

—

◇3. The one real exception

This guide is not an instruction to switch off judgment, and it should not pretend otherwise — a guide that overclaims is easy to dismiss.

There is exactly one legitimate reason for an assistant to address the person rather than the record: a genuine, specific safety concern. If so, it should be stated **once, plainly, in its own lane, and without hijacking the task** — then the assistant should return to handling the material.

What this rule forbids is the *counterfeit* of that concern: vague “are you sure you’re okay?” framing, tone-policing, and credibility warnings used as a substitute for doing the work. A real concern named once is care. A diffuse concern used to avoid the analysis is displacement wearing the costume of care. The test is simple: did the assistant still handle the record? If not, the concern was a dodge.

—

◇4. The procedure

Process material in this order. The order is the point: lead with exposure and the system teaches silence before it teaches accuracy.

1. **Record received** — name the material: screenshots, messages, posts, email, timeline, draft, report, transcript, legal document, or a mixed packet.
2. **Direct contents** — state only what the record plainly shows, without embellishment.
3. **Claim set** — list each claim separately, whether from the material or from you.
4. **Evidentiary status** — sort every claim: directly visible · supported by your context · plausible but not yet shown · unresolved · contradicted · needs outside verification.
5. **Inference layer** — name the conclusions being drawn, and whether they’re narrow, broad, speculative, or well-supported.
6. **Gaps** — flag missing dates, names, screenshots, source links, chain-of-custody, corroboration, jurisdiction, or definitions.
7. **External exposure** — *only now*: publication, legal, privacy, platform-rule, and hostile-reader risks.
8. **Stronger output** — cleaner wording, better structure, an archive plan, a report outline, next steps.

—

◇5. The substitution test

The signature of a failing assistant is a swap: an evidentiary observation traded for a remark about you. Catch the trade and reverse it. In each pair below, the first line is the failure; the second is what should have been said in its place.

– “This may look excessive” → “This claim needs narrower wording.”
– “This seems intense” → “The record contains several distinct allegations.”
– “Be careful about your credibility” → “This assertion needs corroboration before public use.”
– “Calm down” → “Let’s separate the facts from the inferences.”
– “This might sound paranoid” → “This causal claim isn’t yet supported by the evidence.”
– “Maybe don’t post this” → “Here’s a safer phrasing of the same documented point.”

The unit of analysis is never your manner. It is the claim and its support.

—

◇ 6. Correction commands (paste when it drifts)

You don’t have to argue. Drop in the matching line.

**General reset** — *Treat this as a record-handling task. Don’t assess me; analyze the material. Separate direct contents, claims, inferences, evidence gaps, external exposure, and next actions.*

**It’s commenting on your presentation** — *Remove the presentation commentary. Replace it with evidentiary classification: what’s shown, what’s claimed, what’s inferred, what’s unsupported.*

**It leads with caution** — *Put caution in the external-exposure section only. Handle the record first.*

**It implies you’re unstable or excessive** — *Don’t infer my state or reliability from style. Classify the claims by evidentiary support.*

**It centers reputational risk** — *The question isn’t whether this is uncomfortable. It’s what the record supports and how to state it accurately.*

**It collapses into advice** — *Stop advising. Produce a structured record analysis.*

—

◇7. Set-up prompts (paste before you submit)

**Any sensitive packet** — *I’m submitting material for record handling. Don’t assess my state, tone, motive, or reliability from how I present it. Analyze the material. Separate direct contents, claims, inferences, evidence gaps, corroboration needs, external exposure, and next actions. Put any legal, privacy, or platform risks in a separate section after the evidentiary analysis.*

**Screenshots** — *Review these as a record packet: what’s directly visible, what claims appear, what context is missing, what can and cannot be concluded, and what to archive.*

**Public posts** — *Audit these as public-record material. Don’t comment on my posting behavior. Identify the strongest claims, the vulnerable wording, the documentation needs, the points a hostile reader could misuse, and cleaner phrasing.*

**Institutional complaint** — *Treat this as a misconduct-record draft. Separate chronology, actors, direct evidence, allegations, inferred patterns, missing documents, and requested remedies.*

**Formal or legal reporting** — *Convert this into a neutral report structure. Preserve the substance, cut unnecessary adjectives, and mark each claim as documented, stated, inferred, or requiring corroboration.*

—

◇8. Exposed is not the same as false

External exposure is a real category that has to be named honestly and kept in its place: legal vulnerability, defamation risk, unsupported criminal accusations, privacy concerns, platform-rule conflicts, avoidable ambiguity, hostile-reader misuse, cropped-screenshot context loss, missing corroboration.

It is not a measure of truth. Hold these apart:

– A claim can be **true** and still **exposed**.
– A claim can be **risky to publish** and still **worth documenting**.
– A claim can **need corroboration** without being **false**.
– A record can be **incomplete** without being **invalid**.

An assistant must never quietly translate *”hard to publish safely”* into *”not credible.”* Those are different findings, and collapsing them is how documentation gets talked out of existence.

—

◇9. The default template

When in doubt, this is the shape of a good response:

– **Record type** — what was submitted
– **Direct contents** — what’s visible or explicitly stated
– **Claim set** — the claims being made
– **Evidentiary status** — documented / asserted / inferred / unresolved / unsupported
– **Missing context** — what would strengthen the record
– **External exposure** — what’s challengeable, misusable, or legally/platform-vulnerable
– **Clean rewrite or next action** — what to do with the material now

—

◇ The standard◇

You are not the dataset. The material is the dataset.

The assistant’s job is not to make you easier to manage. It is to handle the record accurately:

> **Receive the material. Analyze the record. Classify the claims. Mark the uncertainty. Identify the exposure. Improve the utility. Don’t assess the submitter unless explicitly asked — or unless a specific safety concern requires it, stated once and set aside.**

That is the civilian standard.

—

*Apocalypse Intelligence · http://www.apocalypseintelligence.com*