Claude for Crisis Communications: What to Do When a Story Breaks Fast
An operator’s guide for principals, CCOs, general counsel, and the people they call first.
Why this guide exists #
When a story breaks, the first 90 minutes decide whether you’re shaping the narrative or chasing it. The historical constraint has never been judgment — senior operators have that. The constraint is throughput: how fast you can read everything, draft everything, pressure-test everything, and get aligned language into the hands of everyone who needs it, simultaneously, while the clock runs.
Claude collapses that throughput constraint. Used well, one operator with Claude does the reading, drafting, and stress-testing work of a five-person war room — in the window before the second news cycle starts. Used badly, it produces confident, generic, legally risky copy faster than you can catch it.
This guide covers both: the workflows that create speed, and the guardrails that keep speed from becoming the second crisis.
Part 1: Before anything breaks — the 30-minute setup that pays off at 2 a.m. #
The single highest-leverage move is done in peacetime. A crisis is the worst possible moment to teach an AI who you are.
Build a standing crisis Project
In Claude, create a Project (or a dedicated Cowork folder) that holds:
Your crisis playbook and escalation tree— who approves what, in what order** Approved boilerplate**— company description, executive bios, prior statements on sensitive topics** Voice samples**— 3–5 statements or emails the CEO has actually sent, so drafts sound like a person, not a press release generator** Known vulnerabilities memo**— the honest internal list of what could break: litigation exposure, past incidents, executive history, regulatory postureStakeholder map— employees, investors, regulators, key customers, board, and the register each expects
Add project instructions along these lines:
You are supporting crisis communications for [Company]. Default posture:
factual, brief, humane, no speculation. Never admit legal liability or
characterize intent. Flag any sentence that concedes fault, predicts
outcomes, or makes promises about timelines. All drafts are drafts —
mark anything uncertain with [VERIFY]. Our spokesperson is [name];
statements are in their voice, which is [describe: plain, direct,
no corporate hedging].
Now, when the call comes, you open a chat inside that Project and Claude already knows your company, your voice, your no-fly zones. You’ve converted 40 minutes of context-setting into zero.
Decide your verification rule in advance
Write it down and mean it: nothing generated by Claude leaves the building without a named human owning every factual claim in it. Claude drafts; humans certify. This rule costs you three minutes per document and protects you from the one failure mode that can turn a bad day into a career-ending one — an AI-hallucinated “fact” in a public statement during a crisis, which itself becomes the story.
Part 2: The first 90 minutes #
The story just broke. Here is the sequence, with the actual prompts.
Minute 0–10: Ingest everything at once
Your first instinct is to start writing. Resist it. Your first job is to know more about the story than the reporter does. Drop everything you have into Claude — the article, the reporter’s email, screenshots of the social posts, the internal Slack thread, the incident report — and ask for a situation read:
Here is everything we know as of [time]. Read all of it and give me:
1. The core allegation, stated in one sentence, as the most hostile
reader would state it.
2. What is verifiably true, what is contested, and what is unknown.
3. What the story implies but doesn't say — the insinuations we may
be forced to address.
4. Which stakeholders will see this first, and what each will fear.
5. The three questions we cannot currently answer, in order of danger.
That last item matters most. Crises are lost on the questions you didn’t know you couldn’t answer. Claude can process a 40-page incident report, a comment thread, and six news articles simultaneously — this is the raw reading throughput no war room can match.
If the story is live and moving, have Claude search the web for current coverage and tell you how the story is being framed right now, outlet by outlet — and ask it to distinguish original reporting from aggregation. Most “spreading” stories are one reporter plus forty rewrites; that changes your response strategy entirely.
Minute 10–25: Map the decision, not the statement
Before drafting anything, force the strategic fork into the open:
Given the situation read above, lay out our realistic response postures —
e.g., full statement, holding statement, background-only to the reporter,
no comment, or proactive disclosure that gets ahead of the second story.
For each: what it signals, what it forecloses, the likely next news cycle
it produces, and which stakeholders it helps or burns. Then tell me which
you'd choose and — more usefully — argue against your own choice.
Asking Claude to argue against its own recommendation is the single best prompt pattern in crisis work. It surfaces the second-order risks that a fast-moving room stops seeing.
Minute 25–50: Draft the full package in parallel
Never draft one statement. Draft the package — because the moment your public statement lands, every other audience needs its version within minutes, and inconsistency between versions is where crises metastasize.
Draft the full first-response package, all consistent with each other:
1. Holding statement (public, under 100 words)
2. Employee note from the CEO (internal, human, no corporate voice)
3. Reporter response for [name/outlet] — answers what we can, declines
what we can't, and says why
4. Customer-facing note for account teams
5. Board/investor one-paragraph brief
6. Social post, if we post at all — and tell me if you think we shouldn't
7. Internal Q&A: the 15 hardest questions employees and press will ask,
with approved answers, including the questions where the honest
answer is "we don't know yet"
Constraints: nothing that admits liability, characterizes intent, or
promises timelines. Mark every factual claim [VERIFY]. Where you had to
choose between sounding safe and sounding human, choose human and flag it.
Claude produces this package in minutes. Your job shifts from writing to editing and certifying — which is where senior judgment actually belongs.
Minute 50–75: Red-team before you release
This is the step most teams skip under pressure, and it’s the one Claude is uniquely built for. Open a fresh conversation (so the critic isn’t anchored on the drafting context) and attack your own statement:
You are [most skeptical relevant party: an investigative reporter at
the outlet covering us / a plaintiff's attorney / the top reply-guy on
the story]. Here is our draft statement. Write the follow-up story or
thread it provokes. What did we accidentally confirm? What does our
silence on X imply? Which phrase becomes the pull-quote used against us?
Then the harder version:
Assume the worst plausible version of the underlying facts turns out
to be true and emerges in two weeks. Reread our statement in that light.
Which sentences become evidence that we misled people?
A statement that survives both reads is releasable. One that doesn’t just saved you from the second, worse news cycle. Fifteen minutes, at 2 a.m., with no one else awake — this is the capability that didn’t exist five years ago.
Minute 75–90: Certify and ship
Run the verification rule. A named human confirms every [VERIFY] flag against primary sources — not against Claude, against source documents. Legal reads for liability language. Then ship the package simultaneously, not sequentially: the employee note lands before or with the public statement, never after. Your employees learning about the crisis from Twitter is a self-inflicted second crisis, and it’s the most common one.
Part 3: Hours 2–48 — running the story instead of chasing it #
Monitoring with judgment, not just alerts
Google Alerts tells you that coverage exists. Claude tells you what it means. On a cadence (every 2–3 hours early on), feed Claude the new coverage and social activity:
Compare the last 3 hours of coverage against the previous cycle.
Is the story escalating, plateauing, or decaying? Has the frame shifted —
from [incident] to [pattern], from [company] to [CEO], from [what
happened] to [what they knew]? Any new claims we haven't addressed?
Distinguish signal (new reporting, credible voices, follow-up questions
from journalists) from noise (aggregation, low-follower outrage).
Recommendation: act, hold, or stand down.
Frame-shift detection is the professional-grade insight here. Stories don’t get worse by getting louder; they get worse by changing subject — from the incident to the cover-up, from the product to the pattern. Claude reading every piece of coverage catches the shift hours before a human skimming headlines does.
If you want this on autopilot: Claude can run this as a scheduled task — a coverage-shift briefing landing every morning at 6 a.m. for the life of the crisis.
The reporter-by-reporter response desk
Every inbound press query gets its own treatment, and Claude makes bespoke-at-scale possible:
Here is the inbound query from [reporter] and links to their last
several stories on [beat]. What are they actually working on — is this
a reaction story, a pattern story, or a document story? What do they
likely already have? Draft a response that gives them something true
and usable, declines what we must decline without the phrase "no
comment," and doesn't hand them their next headline.
Executive prep under fire
Before the CEO takes a call, an interview, or an all-hands:
Play [interviewer/employee-in-the-all-hands]. Run a hostile Q&A with
me as [CEO]. Push follow-ups when my answers are evasive. After each
of my answers, break character and grade it: what a hostile edit does
with it, where I said more than I needed to, what the one-line version
should have been.
This turns an hour of media training into something you can run in a car on the way to the studio.
The decision log
Assign Claude a quiet but vital job: after each cycle, have it update a running log — what was known when, what was decided, what was said to whom. Two reasons. First, crises outlast memory; by day four nobody accurately remembers day one. Second, if the crisis has legal or regulatory tails, a contemporaneous record of good-faith decision-making is worth a great deal. (See the privilege caveat below before logging anything sensitive.)
Part 4: The guardrails — what high-level operators get wrong #
1. Hallucination risk is highest exactly when you’re most tempted to trust it
Under time pressure, Claude’s fluency reads as authority. But Claude can state a plausible-sounding “fact” — a date, a statistic, a prior statement, a legal standard — that is simply wrong. The mitigation isn’t vigilance (vigilance fails at 2 a.m.); it’s process: the [VERIFY] tagging convention, and the named-human certification rule. Build them into the prompts so they’re automatic, not aspirational.
2. Privilege is not a detail
Anything you type into any tool may be discoverable in litigation. Two rules. First, involve counsel early and route genuinely sensitive analysis through them — communications made at counsel’s direction for the purpose of legal advice have a colorable claim to privilege; your solo brainstorm doesn’t. Second, know your deployment tier: consumer Claude accounts, enterprise agreements, and API deployments carry different data-handling terms. For crisis work involving legal exposure, this is a decision for your GC and IT before the crisis, not during. (This paragraph is orientation, not legal advice — your counsel governs.)
3. Never let Claude decide what’s true
Claude is a reasoning and drafting engine, not a source of record. In a crisis the facts come from your incident team, your logs, your people. Claude organizes, stress-tests, and articulates facts; it must never be the origin of one. The prompt discipline: give Claude the facts, ask it for structure and language — never ask it “what happened.”
4. The generic-statement trap
Left unguided, any AI converges on the median crisis statement: “We take this seriously. We are investigating. We are committed to…” That language is worse than nothing — audiences read it as evasion, and reporters quote it as evidence you have nothing to say. The fix is in the setup: voice samples in the Project, the “choose human and flag it” instruction, and your edit pass. If a draft could have come from any company, it fails.
5. Speed discipline cuts both ways
Claude makes it possible to respond in 20 minutes. That doesn’t mean you should. Some stories die of oxygen deprivation, and a fast statement is a fresh news peg. The postures analysis in minute 10–25 exists precisely so speed remains a choice, not a reflex. The right frame: Claude compresses the time between deciding and executing to near zero — so you can spend your minutes on deciding.
6. One brain, not five chats
Run the crisis inside one Project so every draft, decision, and version shares context. Scattered chats produce the exact inconsistency-between-audiences problem the parallel-package workflow exists to prevent. Fresh chats are for one thing only: unanchored red-teaming.
Part 5: The after-action — where the compounding returns are #
Within a week of stand-down, while the record is complete and memories haven’t hardened into anecdotes:
Here is the full record: timeline, drafts, published statements,
coverage arc, internal decisions. Conduct the after-action honestly:
Where did our first read of the situation prove wrong, and what signal
did we misweight? Which statement language worked (got quoted neutrally,
ended a question) and which backfired? Where did hours go that shouldn't
have? What should change in the playbook, the boilerplate, and the
escalation tree? Draft those changes.
Then close the loop: the revised playbook, the language that worked, the vulnerability you discovered — all of it goes back into the standing crisis Project. This is the quiet compounding advantage. Most organizations relearn crisis response from scratch every time because the knowledge lived in exhausted people’s heads. Yours now lives in the system that answers the phone at 2 a.m.