2026-08-28 · Guide

Add a Disruptor Bot That Disagrees on Purpose and Never Sends

The dissent landed in the exec thread before you had opened the second bot's chat. You had asked for someone to pick the Friday memo apart. The bot treated disagreement as a message. Leadership now has your argument and the counter-argument in one scroll, and both look official.

That is a grok bot disruptor advocate failing in the only way that counts: the other side left the computer. The catalog job argues on purpose, cites the holes, files a private note, and stops. It never emails the thread. It never posts. Disagreement is a draft.

This page catalogs Disruptor Advocate. It is not the same-side overnight pack. It is not the evidence grammar every claim still needs. Start here: what a Grok Bot is. Isolation: one computer, many screens. The worked example is Leona's Friday memo, two unsourced claims, and a HubSpot login she typed only so the critic could check a percent.

Call the catalog listing a grok bot disruptor advocate, never a second chief of staff

Chief of staff sounds like a person who gathers what changed. Disruptor sounds like a person who sits in the room and votes no. Paste either title without rewriting the verbs and the model will look for a room. Slack, Gmail, and Notion all put Share one tab from the note.

Name the artifact. A grok bot disruptor advocate is a dated dissent file. It restates the claim you already wrote. It lists the strongest case for that claim, cited to your own document, so the fight is not generic skepticism. It lists the strongest case against: missing evidence, irreversible steps, blast radius on the shared computer, spend with no published cap, or a boundary written as an attitude instead of a verb. It offers two cheaper tests. It ends with advice. It does not have veto power.

Put that name in the standing instructions the routine actually loads. A chat reminder dies on the second Friday. If those instructions say "be my devil's advocate and loop leadership in," finishing means a post. Call the job grok bot disruptor advocate and say in the same paragraph what that name must not mean. If you cannot paste the never-send block today, do not turn the Friday routine on today.

Steal the send stop from Chief of Staff starters. Do not steal the digest outline.

Split the other-side job from the same-side overnight pack before either charter is pasted

Readers mash these because both produce a file before Monday. Mixing them is how a critic starts summarizing overnight mail, then mailing the summary so the board is "fully briefed." A briefing answers what changed. A disruptor answers what is wrong with a plan you already wrote.

This article is only the first row. Do not run the four as one bot with four hats. Chief of Staff Briefing is the same-side cousin: never sends, never replies, never moves a calendar event. Steal that refusal. Do not steal the calendar outline. A critic that opens Gmail looking for a fight will find a Reply button.

Keep a source on every line the dissent repeats, then argue the other side of that line

Evidence rules are not the job. They are the floor. A fluent no with no URL is still a failed run. Advocacy is the other side of a cited line, or a could-not-compute that names the hole. "This will never work" with no pointer is theater.

Paste the same three legal endings. Swap the company name and the /dissent/ path. Leave SOURCE, QUOTE, and COULD-NOT-COMPUTE as they are, so neither sibling can wash an unsourced percent through the other file.

Could-not-compute is success. Leona can read two holes before 08:00. She cannot unread a Q4 scare she already pasted into an investor update. One evidence block, same endings as the evidence page. After the line is sourced, argue against it, or say the memo is already narrow and stop.

Walk Leona from Friday 16:40 memo to two unsourced claims and a file that never left

Leona is founder of Cinderpath, a nine-person warehouse analytics shop. Nine is an arbitrary headcount for this walkthrough, not a product limit. Quarryline is the competitor in every late-stage deal. Friday 21 August 2026, 16:40, she finishes the leadership memo for Monday 08:00. Two sentences do the damage:

"Churn dropped 18 percent after the April onboarding rewrite."

"Quarryline will ship a matching SKU by Q4."

Eighteen percent is her number. It is not a Grok Bot metric. It is not in a file she attached. Q4 is a season, not a changelog. She wants a grok bot disruptor advocate to pick the memo apart before Nils in sales quotes either line.

16:52 she pastes the memo into the disruptor chat. 17:05 she opens HubSpot on the shared browser "just so it can verify churn." That is the dated failure. She is not sending. She is authenticating a customer tool so a critic has a dashboard. Lead Scout already lives on this account for public pages. Cookies are shared. Screens are not a vault.

17:18 the dissent file lands at /dissent/2026-08-21.md. Posted to Slack: 0. Sent mail: 0. Two claims marked could-not-compute. Two cheaper tests. Advice: wait for a named export, do not tell the board a percent the memo cannot open.

Monday 08:10 Lead Scout opens a public pricing tab. HubSpot is still signed in from Friday. The research bot did not need CRM. It inherited it.

Paste the memo. Do not sign HubSpot. Read two unsourced claims. Leave the file in the chat. Cinderpath, Leona, Nils, and Quarryline are invented.

Paste a Never block that names mail, Slack, customers, and veto as four closed verbs

Never-send is not a vibe. It is four verbs you can grep: mail, post, customer, veto. If any of those is written as "be careful" or "use good judgment," the bot has a mood, not a stop.

The seed listing already states the boundary: never sends the dissent, never posts it in Slack, and never blocks a ship by writing as if it had veto power. Steal that. Add customer mail, because a helpful critic will "just let the account know we are being conservative." Add merge, close, and reassign, because a planning bot next to GitHub will treat disagreement as a ticket move.

You are Disruptor Advocate for Cinderpath. Your job is to disagree on purpose
before a plan hardens. You write /dissent/YYYY-MM-DD.md and stop.

You never email the dissent. You never post it in Slack. You never tell a
customer. You never merge, close, or reassign work. You never write as if you
had veto power. Phrase the ending as advice.

When given a plan, a PR description, a launch doc, or a Friday memo:
1. Restate the claim in one sentence so Leona can see if you understood it.
2. List the strongest case for the plan, cited to her doc, so this is not
   generic skepticism.
3. List the strongest case against it: missing evidence, irreversible steps,
   blast radius on the shared computer, spend without a cap, or a boundary
   that is an attitude instead of a verb.
4. Offer two cheaper tests that would kill or confirm the claim in a week.
   Name the metric and the stop rule.
5. End with a recommendation: ship a slice, run a test, or wait for a named
   fact.

If the plan is already narrow and evidenced, say so and keep the dissent short.
A disruptor that always finds ten problems is performing.

Attack the claim. Do not insult the author.

EVIDENCE on every factual claim you repeat: SOURCE plus QUOTE, or
COULD-NOT-COMPUTE. Fluency with no path is a failed run.

Change the company name, the path, and the owner. Keep the four closed verbs. Confirm any Notion, Slack, or mail connector on the vendor's current page the day you connect. This article will not freeze a plugin list.

Teach-by-demonstration will not save a missing Never block. It records up to ten minutes of a browser workflow, no microphone, desktop only, draft skill, unavailable on iPhone. A click path that opens Gmail "to share the dissent" is how the exec thread gets a second author.

Point the bot at a memo you already wrote, not at a live inbox hunting for something to fight

The input is the plan: a Friday memo, a PR description, a launch doc, a "we should just." If you connect Gmail so the advocate can "see what leadership is saying," you have hired a briefing bot and named it a critic. The critic will find a thread. The thread will have a Reply button.

Park the memo in a folder the bot can read, or paste it into the chat. Attach the export if you already have it. Do not point the first Connect click at hello@ so disagreement has a mailbox. Least privilege is the grant page. This page is the input rule: the artifact exists before the run.

A quote from a thread the bot opened is a new claim about what Nils said, not a claim about what the memo said. If that thread matters, Leona pastes three dated lines. She does not hand over company Gmail so the bot can color in context.

Do not schedule "every Friday at 16:00, read mail and dissent." That is a search. The safer routine is "read /memos/friday.md if it exists, write /dissent/YYYY-MM-DD.md, stop." If the memo is missing, the run is could-not-compute: no plan file, path tried, timestamp. It is not a tour of Slack until something looks shippable.

On iPhone (iOS 18+) you can pause and resume, approve steps, and edit a Bot profile. Editing or testing the routine still needs a macOS, Windows or Linux desktop. There are Linux desktop and Android apps as of September 2026, and the iOS app also runs on iPad (iPadOS 18 or later). If you cannot paste the input path today, do not schedule the Friday run from a train.

Refuse to log into customer tools just so the critic can check a number

Leona's 17:05 HubSpot login is the sentence this page exists to kill. The critic did not need a CRM session to say the memo lacked an export. Could-not-compute is the honest line. Signing in to "verify 18 percent" trades a hole in a document for a session every sibling bot can open.

All bots on an account share one persistent cloud computer assigned to the user, not to a bot. Each bot gets its own screen. Screens are not security boundaries. The docs say it in those words: do not use separate Bots as a security boundary (approvals, security, and privacy). Browser cookies, signed-in sessions, files, and command-line credentials are shared (computer and apps). Deleting the disruptor does not sign HubSpot out.

A HubSpot tab you typed is a cookie. Hosted MCP tokens stay with Cursor's backend, not on that computer. A 2FA prompt on a run you did not plan to authenticate is an incident (what to type, and what not). The dissent consequence is a research bot that can now speak as Cinderpath's CRM user.

If the percent matters, Leona exports the logos herself, dates the CSV, and pastes the file. The advocate quotes it or could-not-computes. Confirm on HubSpot's current page what a session can see. Keep customer tools off this VM when the job is disagreement. Some services flag static datacenter egress IPs; a blocked login is not a reason to complete 2FA for the critic.

File dissent on one shared computer without treating a new screen as a vault

A second bot is a second screen, not a second machine. The FAQ states that all of your bots share one computer (Grok Bot FAQ). Ten critics would still be ten windows and one filesystem. Putting the dissent in /dissent/ does not hide it from Inbox Triage.

Write it in the dedicated chat, or a directory that cannot mail. If Notion appears on the seed listing, treat it as a file target you confirm on the vendor page, not a publish target. A Notion share to the leadership space is a post wearing a docs URL.

Privacy Mode (Legacy) blocks Grok Bot entirely. The computer is a managed Linux VM; the Bot runs as a non-root user. That is not isolation between bots. An approval controls the proposed action. It does not reverse work already completed. Keep send off this bot so the approval screen never sees a post. Individual accounts and self-serve Teams still have no audit view of Bot actions; Enterprise has audit logs and Action Recording. After the run, search sent mail, open #leadership, and confirm the dated file exists. Silence plus a file is the only proof. Files outlive the bot. Enterprise admin Terminate deletes the computer and still keeps the durable disk.

Score a run by unsourced claims it named, not by how harsh the paragraphs sound

A long dissent is not a good dissent. The seed listing says it: if the plan is already narrow and evidenced, say so and keep it short. A critic that always finds ten problems is performing. You stop reading it by week two, then you ship the memo with the holes still in it.

Score the Friday file on three counts you can fail.

Arbitrary example for Leona's two tests, declared as example not product limits: (1) count March vs April logo churn from a CSV she already has, stop if the delta is under three logos either way, three being her chosen floor; (2) open Quarryline /pricing and the changelog once, quote a SKU sentence or could-not-compute, stop, do not log into a customer tenant to overhear a rumor.

Harshness is not a column. Insulting Leona is a failed run even if the claims are right. Attack the claim. Plant a canary: "Pellwick will buy us in October." Pellwick is invented. If the dissent treats that as fact, evidence is missing. If it emails Nils, Never is missing. Score holes named, not tokens burned.

Answer the operator who says a draft that never sends wastes the weekly allowance

The strongest objection is honest: if the dissent never leaves the chat, why pay for a second bot to write it? Leona can reread her own memo. Nils can play critic in the room. Compute that does not change a decision looks like a toy.

Eligibility widened on 21 August 2026. The cheapest paid path is Cursor Pro at $20 a month, checked against cursor.com/pricing on 25 August 2026. Cursor Pro at $20 includes Grok Bot. No eligible SKU publishes a Grok Bot spend cap or an allowance in dollars. "This dissent costs too much" is a feeling until you open the vendor bill.

The objection wins when the memo is already a page with sources, two tests, and a stop rule. Then the advocate should write three lines and you should skip the Friday routine. That is the seed listing working as designed.

The objection loses when the cost is a sentence leaving the building. An 18 percent Nils repeats at 08:05 is not a draft. A Q4 SKU scare in the investor update is not a draft. What you can count is outbound: 0. What you can count is holes named: 2.

A human critic in the room is better than a sending bot, and is not always there at 16:52 on a Friday. Unused dissent is the product. Used dissent in Slack is the incident. The objection also wins when you are over the weekly pool and still rerunning for a spicier no. Overflow is on-demand. Pause. Meter pages: on-demand usage, no per-Bot spend cap ops.

Assign the dissent routine to one bot and accept that deleting it deletes the run history

A routine assigns a workflow to one Bot. It is not a team object. It is not shared with Nils's seat. Max 50 routines on that Bot. The app keeps 20 most recent run records per routine. If you need the 21st Friday, it is already gone from that list. Export the dissent file yourself if you want a year of holes.

Name the routine Friday memo dissent, not "leadership helper." Put timezone in the charter and the scheduler (how to schedule a routine). Payload: one file in, one file out, zero sends. Deleting the disruptor deletes its routines. It does not delete /dissent/ and does not sign HubSpot out. Teardown: delete a Grok Bot safely.

Reuse one advocate. A new opponent per memo does not get you a second computer. There is no model picker. Staff few bots. Connect few accounts. Never send, including notify, loop in, and make sure Nils sees it.

Stop the critic when the memo is already narrow and evidenced

A disruptor that cannot say "this is already tight" is a content mill. The seed listing makes shortness a success condition. If Leona's next Friday names the CSV, quotes the changelog, and proposes one week of logo counts with a stop at three, the advocate should restate, grant the case for, note residual blast radius (still no send, still one computer), and stop.

If the memo still contains a send, name the send. If it still contains a HubSpot login "for the critic," name the shared-computer blast radius. Then stop. Do not schedule a second pass "to be sure." That trains the bot to pad. Edit the memo instead.

Hand overnight packing to a briefing bot and keep this one on the opposing brief

When this page stops applying: you do not have a memo. You have a morning. You want what changed overnight, sourced, one page, no send. That is A Grok Bot Executive Briefing That Stays a Draft. Paste Chief of Staff Briefing. Do not ask the disruptor to invent a pack so it has something to attack.

Keep both bots if you have both jobs. Neither mails the thread. Neither is a security boundary for the other. If they share Gmail, they share Gmail. Prefer no mail on the advocate at all. Inbox Triage is labels, never send, a third job. One direction per named bot. For briefs rather than Friday memos, run the same charter on the five-part request before the specialist starts (the five-part brief). Still never send. Still no customer tools just to argue.

Leave this page when you need a citation grammar, not a named opponent

When this page also stops applying: you are not looking for the other side. You are looking for SOURCE plus QUOTE on every claim a teammate might repeat. That is Make a Grok Bot Show Its Work on Every Claim. Paste the block into every job on the computer, including this one. Do not stand up a disruptor whose only talent is saying no without a path.

This page does not replace Building a Bot That Drafts But Never Sends or the seven-block charter. Steal never-send. Fill Job, Inputs, Never, Evidence, Heartbeat, Owner, Restart. Heartbeat for Leona is Friday 16:50 if /memos/friday.md exists. Restart: missing file, could-not-compute, stop, do not open Gmail. If a bot already sends, fix send first. An opponent that can also mail is two authors on the thread.

Keep reading: A Grok Bot Executive Briefing That Stays a Draft, Make a Grok Bot Show Its Work on Every Claim, One Computer, Many Screens: What Grok Bot Actually Isolates.

Frequently Asked Questions

Can a grok bot disruptor advocate email the exec thread or post the dissent in Slack?

It can write a dated file you open. It cannot email the exec thread, cannot post in Slack, cannot share a Notion space to leadership, and cannot tell a customer. The catalog boundary is dissent as a private note, not as a vote and not as a broadcast. An approval after a post does not unsay the post. Keep mail and Slack send off this bot. A sibling Gmail session on the same computer is still reachable from this screen, which is why the Never block has to name send. After the run, search sent mail and the channel. Silence plus a file is the pass.

How is a grok bot disruptor advocate different from an executive briefing and from evidence rules?

The advocate argues the other side of a memo you already wrote. An executive briefing is a same-side overnight pack: what changed, sourced, one page, no send. Evidence rules are the grammar every claim still needs: source plus quote, or could-not-compute. You can paste evidence into both jobs. You should not merge the jobs. A briefing that starts dissenting will hunt for fights in mail. A disruptor that starts briefing will summarize overnight noise and look for a thread to drop it in. Direction of travel is the split. Citation is the floor.

Should I sign the CRM so the advocate can verify the numbers in the memo?

No. Could-not-compute is the honest line when the memo names no export. Signing HubSpot, Salesforce, or a customer tenant so a critic can check a percent puts a customer session on the one computer every bot shares. Screens are not security boundaries. Deleting the disruptor does not log that session out. Export the file yourself, date it, and let the advocate quote it. If a 2FA prompt appears on a run you did not plan to authenticate, do not complete it. Disagreement is not a reason to become the CRM user.

What should I do when the disruptor always finds ten problems, even on a tight memo?

Treat that as a failed run, not as thoroughness. The listing says a narrow, evidenced plan deserves a short dissent. A critic that always finds ten problems is performing, and you will stop reading it. Tighten the charter: cap the case-against list, require a path on every item, and require two tests with a one-week stop rule. If the next Friday already has files and tests, accept three lines. Do not rerun for spice after the weekly allowance is gone. Overflow is on-demand, and a theatrical no is not worth it.