Backbrief Business OS. The complete Backbrief system in one install.
Put a review panel between your idea and your money.
Backbrief Business OS runs a business decision the way a careful operator would, before any money or work moves. You describe the business in plain words. The system writes the brief, drafts the plan in 90-day units, stress-tests it against a public rubric until it earns a grade you can trust (or plateaus and says so), then waits for your explicit go before any execution starts. The full Backbrief build team ships inside: one download, one install, one setup.
You do not need to be good at prompting. /intake is a conversation: describe the business the way you would to a partner, and the system writes the structured brief. Your only job is to correct it until it is true.
$79, one-time. One price, no timer. All sales final; try the free kit first.
Checkout, then a delivery email with your download. You install it by telling Claude to install it. Set up in about ten minutes.
Page updated August 22, 2026. The changelog in the download names the current release.
Proof, not a highlight reel
The first business it graded was its maker’s. It said no.
The bundle ships with a real run: the author’s own pet memorial brand, NestPet, taken through the full loop, unedited. Two independent panels, a month apart, graded the same material C: 6.4 on the original July run, 5.9 on the August re-record. Both put the same two dimensions on the floor, and that agreement, not the decimal, is the signal. The plan itself had refused to pretend otherwise. Its market section states, in its own words:
“Revenue to date: $0. No listing, no sales, no conversion data exists... demand is untested.”
The owner invoked the early exit rule rather than spending two more passes polishing around a question only a live listing could answer. Presented with the three required options, the owner chose to pivot. And the gate held.
A second run ships beside it: TrailNotes, the product’s disclosed specimen business, taken through the full revise loop to the three-pass cap. It improved every loop, 6.9 to 7.3 to 7.7, and still never passed. The loop stopped at the cap, per the rule, instead of rounding up.
This is the trust feature. A system that only shows you its passing runs gives you no way to know what it does when the answer is bad. These runs show you exactly what it does: it names the weaknesses, stops when told, and keeps the ops team locked until a plan actually earns your go. Both scorecards are published on the inside page, beside the grading rule itself, verbatim.
C, 5.9 of 10. Below the 8.5 pass bar.
Financial viability, the disagreement kept in the record, not averaged away: “contrarian 2 (‘a margin floor with no reconciled cost stack is a wish’) vs. expansionist 5.”
“at /approve chose PIVOT. The ceo-gate held: nothing was unlocked, no ops agent was routed.”
Five steps. Each one leaves a file you can read.
Intake.
You talk about your business for ten minutes, or paste what you have, and it comes back as a one-page brief with your own numbers in it, which you correct until it is actually true.
Plan.
The brief becomes a business plan in 90-day units, sized to your real time, money, and skills, not a hypothetical team's.
Grade.
Five advisors with distinct lenses score the plan 1 to 10 against the public six-dimension rubric, and every score must cite the plan's own words.
Approve.
Your recorded go or no-go. Nothing downstream runs without it.
Route.
After your go, approved work routes to the ops team, and anything outward, spending, sending, or publishing, still stops for your sign-off. Every time.
The six dimensions the panel scores. Public, on purpose.
- Is the offer clear.
- Who pays, for what, stated so a stranger understands it in one sentence.
- Is the demand real.
- Actual evidence people want this: named buyers, comparable products, real signals. Not "everyone needs this."
- Do the numbers work.
- The unit economics close using your own numbers from the brief.
- Can you actually run it.
- With your hours, your budget, your skills, not a hypothetical team's.
- Are the risks named.
- The top ways it fails, each with a mitigation or your explicit, eyes-open acceptance.
- Why you.
- A concrete reason someone picks this over the obvious alternative, including doing nothing.
Scores run 1 to 10 with equal weights. An 8.5 average passes. That is an A-.
What the grade means, and what it never means.
A passing grade means the panel could not find a weakness left un-named. It does not mean your business is guaranteed to work.
The panel cannot wave things through. A legal claim about a named state or country goes to the researcher for retrieval before it can count as risk coverage, and every outside figure that moves a score is tagged verified, unverified, or vendor claim. A plan quoting itself is not verification.
Grading is capped at three passes, ever. If the plan is still below A- after the third pass, the loop stops and tells the truth: you get the final scorecard, the top three unresolved weaknesses, and your options. Proceed at the real grade, pivot, or kill. It never grinds past the cap and never quietly waves a weak plan through.
There is also an early exit. If one of the lowest-scoring dimensions cannot be fixed and you say so, the system stops right there and gives you the honest options instead of burning the remaining passes revising around it. A plan that plateaus is a result, not a failure of the system.
What’s in the bundle
Backbrief Business OS used to be an add-on pack that required the free Backbrief Kit installed first. It now ships complete: the whole build team is inside the same download. One install, one /setup interview, nothing to layer on afterward.
You’re the CEO. Your team lives in files.
Every agent, command, and rule is a readable Markdown file on your machine, and every step of the loop leaves an artifact you can open: the brief, the plan, the scorecard, the decision log. Nothing runs on a server you cannot see, and nothing about the product has to stay hidden until you pay: the complete file tree, the grading rule, and both real scorecards are published on the inside page.
The build team. An orchestrator that plans and routes, a researcher, planner, builder, reviewer, a cheap runner for mechanical work, a fresh-context verifier that checks finished work before you see it, and the five council advisors behind /council and /grade.
The operations tier. The team that executes an approved plan:
cmo.
Marketing operations: channels, offers, content calendars, launch sequences, with honest tool recommendations and disclosed affiliate links.
cfo.
A financial analyst, not an advisor: it recomputes the plan's unit economics from your own numbers, drafts budgets, flags risks, and outputs questions for your accountant or attorney rather than advice.
web.
Site and ads specs: what your pages and campaigns should say and do, written so they can be built without guesswork.
ops.
SOPs and day-to-day operations: the repeatable procedures that keep the business running.
The commands. The CEO loop: /intake, /business-plan, /grade, /approve. The build team’s own, including /grade-idea, /setup, /next, /verify-install, /critique, /council, /handoff, /pickup. Of those, /next is the one you will use most: it reads your files, tells you where you stand, and offers the next step, correct even in a brand-new chat.
The rules. Always-on guardrails, including the public grading rubric with its three-pass cap, the CEO gate that locks the ops agents until a plan is graded and approved, and the escalation rule that stops outward actions for you every time. On the shell-command path, that rule can also be enforced rather than followed: see the enforcement layer below, which the installer recommends on a fresh install and leaves off on a setup you already run unless you say yes.
The skills. Bundled, attributed, version-pinned MIT skills routed to the agents that use them: sales, marketing, ops, and finance procedures for the operations tier; SOPs, competitive analysis, and a standing AI-slop quality gate for the build team.
The knowledgebase. The team works from your sources. Drop in your product docs, guides, or a link to your docs site, and the researcher cites them, by filename or URL, before it reaches for general knowledge. Naming a document in chat is enough to have it filed.
Two docs tracks. A README for the Claude Code track, and a Cowork quickstart that installs and runs the same bundle entirely in plain English, no terminal.
The worked examples. The NestPet business run, brief to scorecard to pivot, exactly as it happened; the TrailNotes run, three revise loops to the cap and an honest plateau; and a real build-team run (a one-page FAQ planned, built, and independently verified).
Your whole team can run the identical setup. Now you can prove it.
Team Parity. Your whole team can run the identical setup today. Clone the repo and the same agents, rules, and skills land on the next machine. What you could not do before was prove it. Run /verify-install and read the parity line: the version, whether the guardrails are live in the running session rather than sitting unused in a settings file, and whether a personal file is overriding a shared one. Two teammates compare that one line, and matching lines mean you are running the same system. The rest of what a second person needs is here too: a roles template that puts approval authority on one named person, a decision-log convention that holds up with more than one writer, and an onboarding path that takes a new teammate from clone to first task without booking time with you.
Optional, and you are asked before it goes in
Every agent is told to stop. This is the part that does not rely on being told.
The escalation rule says outward actions stop for you: sending, publishing, deploying, spending. In every install that is an instruction the team follows, and most of the value is there, because the common failure is an agent that never considered the question at all, not one that considered it and pressed on anyway.
An instruction is not a lock. If a model overlooks the file, the action proceeds.
So the bundle ships an optional enforcement layer, and the installer asks you about it before any of it goes in. On a setup you already run it stays off unless you say yes. On a fresh install it is recommended, and goes on if you have no preference either way. On a yes, it merges Claude Code’s own permission rules into your project settings. Your existing settings file is never replaced. It is backed up first, and the merged result is re-read and parsed before the install calls itself done; if it does not parse, the backup goes back and enforcement stays out.
With the layer on, when an agent reaches for a matching outward shell command, Claude Code’s permission engine interrupts and asks you first, whether or not the agent remembered the rule. In an unattended run there is nobody to ask, so the command is blocked instead. It fails closed. The rules hold in the default permission mode and inside every subagent — and in headless runs they fail closed in every mode, bypass included, which we probed with a control rather than assumed. One honest exception, and it is your switch, not the agent’s: in an interactive session with permission prompts turned off, these ask rules are off with them — and /verify-install now probes for exactly that state and reports it, instead of letting a live canary imply more than it proves.
Where it stops, stated plainly
It covers a published, enumerated list of outward shell command shapes on the command path, plus edits to the files the layer itself lives in. Enumerated is the honest word: a command shape the list does not name is governed by the escalation rule, not the engine, and we benchmark real sessions to find out where that line actually falls rather than assuming. It does not read the inside of a script a model writes and then runs, and it does not cover MCP server tools, though the README beside it shows the one line that adds a server you have connected. For script interiors the stronger answer is OS-level sandboxing, which Claude Code offers on macOS and Linux. There is no Windows equivalent, so nothing here promises one.
And it guards the file it lives in
A permission rule that gets deleted stops applying on the very next command. Without protection, an agent holding file-editing tools could remove the stop and then take the action itself, in the same session, with nothing to prompt you. Editing those files with the file-editing tools asks you first. The same guard covers your project’s agent definitions and rule files. The guard makes the quiet path loud; it does not close every path, because a shell one-liner that rewrites a file is a different shape the command rules only partly name, and we say that instead of implying otherwise. It asks rather than blocks, because changing your own settings is legitimate and you may well want to. What should never happen quietly is the team changing the rules that govern the team.
Proof, not assurance
/verify-install runs a command that does not exist and expects the engine to block it, in the shell your sessions actually use. A blocked response is the permission engine itself confirming the layer is loaded and live. It reports presence and liveness separately, so a half-finished install cannot read as a clean one.
Your everyday commands, a commit, a local build, or a test run, match none of this and are untouched.
The free kit, next to the complete system.
| Feature | Free Backbrief Kit | $79, one-time Backbrief Business OS |
|---|---|---|
| The build team | Included, complete | The same team |
| The operations tier (cmo, cfo, web, ops) | Not included | Included, locked behind your recorded go |
| The CEO loop (/intake, /business-plan, /grade, /approve) | Not included | Included |
| Commands | The build team's, including /grade-idea | The same, adding the CEO loop's four |
| Rules | The always-on guardrails | The same, adding the grading rubric, the CEO gate, and business skill routing |
| Enforcing the escalation rule | The rule, as an instruction the team follows | Offered at install, off on a setup you already run unless you say yes: the permission engine then asks before a matching outward shell command runs |
| Bundled skills | The base set | Adds the sales, marketing, ops, and finance procedures |
| Worked examples | A real build-team run | That run, plus two graded business runs: a real C with a recorded pivot, and three revise loops to the cap |
| Updates | Public repo and plugin | Included: your purchase email always links the current version |
| Refund policy | Free, nothing to refund | All sales final; the free kit is the try-before-you-buy. Statutory rights unaffected |
The free kit is complete on its own and stays free. Get the free kit if you only want the build team. Backbrief Business OS is the complete system in one box. Read the full comparison.
Two tracks, one bundle.
On Claude Code.
You already run projects in a terminal. One paste installs everything in about two minutes, and every artifact the loop produces is a plain Markdown file in your repo.
Not technical.
The Cowork quickstart is the same bundle run entirely through Claude Cowork: you paste one install message, then you talk, you read, you decide. No terminal at any step.
Nothing else to install. The free Backbrief Kit stays free if you only want the build team; Backbrief Business OS is the complete system in one box.
What you need. All of it.
- A Claude subscription that runs Claude Code (Pro works; heavier grading runs simply use more of your plan's capacity) or Claude Cowork for the no-terminal track.
- Windows or Mac. About ten minutes to install and set up. No coding, no build step, nothing else to buy.
- Honest cost note: the five-advisor grade is real model usage on your own Claude plan. The bundle ships a token-discipline system that keeps it cheap by default.
One price. One promise.
$79 The launch window closed on schedule. The price moved a few days later, late in the buyers' favor, and stays here.
One-time purchase. No subscription beyond the Claude plan you already run it on.
Every update to the bundle is included: your purchase email always links the current version.
All sales are final: the download is immediate and the files are yours to keep. Try the free kit first; statutory rights are unaffected.
What happens after you buy
Checkout.
One-time payment through the secure payment link. No subscription starts.
The delivery email arrives.
Your download link plus the install instruction for your track, Claude Code or Cowork. The same email is where every future update reaches you.
About ten minutes later, you are running.
The install is one instruction: open Claude in the unzipped folder, type install this, and it does the copying and the checks. Then /setup interviews you about your business. Your first /intake can start the same sitting. From then on, /next always knows your next step.
Worth naming the alternative: verdict-only idea graders sell for around $39. They hand you a number and stop. The number here comes with the plan behind it, every score citing your own words, the file trail you can audit, and the team that executes after your go, with updates included. That is what the extra $40 buys.
Wondering who says a passing grade means anything, what it costs to run all in, or what happens when the analysis is wrong? The FAQ below covers the honest ones. Read the FAQ
Questions, answered plainly.
These are the questions a careful buyer should ask before trusting a system like this. Here are the honest answers.
Who says a passing grade actually means anything?
Nobody outside the process, and the system says so itself. A passing grade means the panel could not find a weakness left un-named. It does not mean your business is guaranteed to work.
What the grade buys you is discipline, not certainty. The rubric is public, every score must cite your plan's own words, and disagreement between advisors is reported rather than averaged away. The grade measures whether your plan survives a structured, cited stress-test. It does not measure your execution, your market's timing, or your luck. A plan can pass and still fail, and a plan can plateau below passing and still be worth doing with open eyes.
What if the money analysis is wrong?
Treat it as analysis, never as advice. The cfo agent is a financial analyst, not a financial advisor: it recomputes your plan's unit economics from your own numbers, drafts budgets, flags risks, and outputs questions for your accountant or attorney rather than advice. Its math is only as good as the figures you give it, and it marks estimates as estimates rather than treating them as facts.
Before you act on any financial, legal, or tax decision, verify it with licensed professionals. The system is built to hand them better questions, not to replace them.
What does it cost to run, all in?
The bundle price, plus the Claude subscription you run it on. Backbrief Business OS is a one-time purchase, and it requires a Claude plan that runs Claude Code or Claude Cowork. That subscription is the only ongoing cost the system needs.
The bundle itself is plain Markdown files: nothing to host, nothing to install beyond copying folders, no per-seat software, no hidden tool costs. When the bundle's marketing output recommends tools, free alternatives are named in the tool tables beside every paid pick, and it never recommends a paid tool for a need the free option covers at your current scale.
Do I need to be technical?
No. The bundle ships with two docs tracks. If you work in Claude Code, the README covers install and first run in a terminal. If you do not, the Cowork quickstart runs the same system entirely through Claude Cowork: you paste one install message in plain English, and from there you talk, you read, you decide. No terminal at any step.
Do I need to know how to prompt?
No. The system is built so that plain language is enough. /intake is a conversation: you talk about the business for ten minutes, or paste what you already have, and it comes back as a one-page brief with your own numbers in it. You correct the brief until it is true, and that correction is the skill, not prompting.
The same holds after intake. /setup interviews you, and if you are ever unsure what to do, /next reads your project files and offers the next step. There are no magic words to learn anywhere in the loop.
Is there a human behind this?
Yes. Reply to your purchase email and a human reads it and answers. That thread reaches the author directly. There is no bot between you and help.
What is the refund policy?
All sales are final: the download is immediate and the files are yours to keep, so there are no change-of-mind refunds. The free kit exists so you can try the system before spending anything. Faulty or missing downloads are made right, and statutory rights are unaffected.
That applies to the Backbrief Business OS purchase. The Backbrief Kit is free, so there is nothing to refund there.
How are tool recommendations chosen?
Openly. Some links on this site and in the bundle's output are affiliate links, which pay the author a commission at no extra cost to you, and they are disclosed as affiliate links wherever they appear. A free or cheaper option is named beside every paid pick, and it never recommends a paid tool for a need the free alternative covers at your current scale.
One more thing worth knowing: no affiliate exists for Claude or Vercel. We recommend them anyway, because the system runs on them.
What happens if my plan grades below passing?
The loop stops and tells you the truth. Grading is capped at three passes. If the plan is still below passing after the third pass, you get the final scorecard, the top three unresolved weaknesses, and three options: proceed at the real grade, pivot, or kill. It never auto-proceeds and never loops past the cap. There is also an early exit: if one of the lowest-scoring dimensions cannot be fixed and you say so, it stops immediately and gives you the honest options instead of polishing around it.
This is not hypothetical. The worked example that ships in the bundle is the author's own business grading C from two independent panels a month apart, the owner choosing to pivot, and the ops team staying locked. A second example ships beside it: a plan that improved for three real passes, plateaued below the bar, and was stopped at the cap. Remember what any grade means here: a stress-test survived, never a promise of success.
Do I need the free kit first?
No. Backbrief Business OS is complete: the build team ships inside the bundle. Unzip one download, paste one install message (or follow the Cowork quickstart), run /setup, and you are running. If you already installed the free Backbrief Kit, the bundle installs over it cleanly and the /verify-install check confirms the result.
Does anything actually stop an agent, or is it only told to stop?
Both, and the difference is worth knowing. The escalation rule is an instruction: every agent is told to stop before sending, publishing, deploying, or spending, and in practice that is most of the value, because the common failure is an agent that never considered the question at all. It is not a lock. If a model overlooks the rule, the action proceeds.
The bundle also ships an optional enforcement layer. The installer asks you about it: it is recommended on a fresh install, and on a setup you already run it stays off unless you say yes. With it on, Claude Code's permission engine stops a matching outward shell command and asks you before it runs, whether or not the agent remembered the rule, and blocks it outright in an unattended run where nobody can be asked.
What it covers is shell commands on the command path. It does not see inside a script a model writes and then runs, and it does not cover MCP server tools, though adding a server you have connected takes one line. On macOS and Linux, Claude Code offers OS-level sandboxing, which is the stronger answer for script interiors; there is no Windows equivalent, so nothing here promises one.