Backbrief

How to use Backbrief

This guide takes you from the download to your first grade, then through the full loop from brief to recorded decision. Every step and command name comes from the files in the download. It covers Backbrief 0.1.2.

Before you buy

What you need: Claude Code, installed and signed in. It runs in a terminal or in the desktop app. You also need a project folder.
No other accounts: Backbrief needs no other accounts and no keys. It runs on your own Claude Code access, whatever plan or billing it uses, and a grade uses that access like any session. The product has no server and collects nothing.
For video grading: Grading a video from a link or a file can use free local tools (yt-dlp and ffmpeg). Claude asks before it installs any. A pasted transcript needs none.
Price and terms: $29, paid once. One license, one person. Updates are free for as long as Backbrief is offered. All sales are final. Read the license terms and the refund policy before you buy.
What it does not do: Backbrief grades ideas and writes plans. It gives analysis, not advice. Publishing, sending and spending stay with you.

Install

Unzip, open Claude Code in the folder, and ask Claude to install it.

Step 1

Unzip the download.

Step 2

Open Claude Code in the unzipped folder.

Step 3

Type: read the README.md in this folder and install this for me

Step 4

Answer Claude’s questions. It asks for your project folder and whether Backbrief goes in every project or only that one.

It then asks about three optional layers. The first is permission rules that make Claude Code ask you before outward shell commands such as git push, curl or a deploy. The second loads your latest handoff note and recent decisions at the start of each session. The third records which Backbrief command you ran last, so /bb-status can report it. On a fresh install the permission rules are recommended and go on unless you say no. The other two stay off unless you say yes.

Step 5

Claude copies the pieces, runs /verify-install and offers /setup. The download’s README puts the install at about two minutes.

Global or per project: Global puts Backbrief in every project. Per project keeps it in one project, or keeps that project on one version while others move. Either way, each project needs its own scaffold.
You already have a setup: Say so, and Claude merges Backbrief without overwriting anything you have.
You are upgrading: Say “this is an upgrade from an earlier Backbrief product.” A plain copy skips files with the same name and would leave you on the old version.
You prefer to copy by hand: The README holds a longer prompt to paste and a by-hand copy for macOS, Linux and Windows.

Then run /setup

Run it once per project: Open Claude Code in your project, run /verify-install first, then type /setup. /verify-install confirms the install and whether its permission rules are live in this session. /setup then interviews you in plain language, one question at a time.
What it asks: What the project is. What you want done in the next few weeks. Anything Claude should ask you about first, beyond the standard stops (spending, sending, publishing, deleting). Anything already decided. Whether you keep a knowledgebase of documents and links the team should work from. Whether the project is a business to run through the Backbrief loop.
What it writes: Your answers go into context/strategy/current-state.md and context/strategy/current-priorities.md, in your own words where possible. Claude reads them back to you so you can correct them.
The default pace: Setup does not ask how much the team should drive. The pace starts at guided: Claude walks you through each step and asks before it runs anything. After you have seen the team do real work, /next asks once whether you want a different pace.

Your first grade

Run /grade-idea on one idea. You do not need a plan or a brief.

Run it

Type /grade-ideafollowed by your idea. If you type the command alone, Claude asks one question: “Describe your business idea in a sentence or two: what it is, who pays, and any numbers you already have.” Whatever you answer, thin or rich, is what gets graded.

The command needs nothing but that description. It needs no plan document, no earlier scorecard and no other file in the project.

The panel runs

Claude compresses your description into a brief of at most 300 words and adds no fact you did not state. Five advisors then score it in parallel. Each works through its own lens and sees only the brief and the rubric.

If the idea turns on what is legal, licensed, required, disclosed or taxed in a named place, Claude researches that point first. Without that check, the advisors label it an unverified local-law assumption and do not score it as risk coverage.

The scorecard

Claude prints the scorecard in the layout below and saves it to outputs/idea-scorecard-<slug>.md, where the slug is a short form of your idea.

Reading the scorecard

IDEA SCORECARD
Idea: <one-line summary of what you described>

Clarity of offer         X.X/10
Market realism           X.X/10
Financial viability      X.X/10
Execution feasibility    X.X/10
Risk coverage            X.X/10
Differentiation          X.X/10

Reachability (trial)     X.X/10   beside the rubric, not in the average

Average: X.X/10   Pass line: 8.5   Grade: <letter>

Lowest: <dimension name, or the tied dimensions listed> (X.X)
What the panel caught: <one sentence, cites the description or names the gap>

The panel's read: <pursue as-is / close the gaps first / walk away> - <one sentence>
First move: <the single next action, taken from the executor's own answer>

This is a stress-test, not a certification. A pass means no confident weakness was left un-named. A low score is the useful result.

Six dimensions

Each scores 1 to 10 and carries equal weight. Every score cites your own words.

Clarity of offer: who pays, for what, stated so a stranger understands it in one sentence.
Market realism: evidence that demand exists: named buyers, comparable products, real signals.
Financial viability: the unit economics close using the numbers you gave.
Execution feasibility: for you, with the time, money and skills you stated.
Risk coverage: the top risks are named, each with a mitigation or your explicit acceptance.
Differentiation: a concrete reason to pick this over the obvious alternative, including doing nothing.

Reachability (trial)

A seventh line asks whether a stranger could first find this, using only the channels, budget and audience you stated. It sits beside the rubric. It never enters the average, the letter or the pass line.

Average and grade

The average of the six scores sets the grade. The pass line is 8.5. A is 9.0 and above, A- is 8.5 to 8.9, B+ is 8.0 to 8.4, B is 7.0 to 7.9, and C is below 7.0. Where advisors disagree on a dimension, the scorecard reports the split instead of averaging it away.

The tags

Every figure that moves a score gets a tag. Verified means checked against a named source. Unverified means plausible and unchecked. Vendor claim means the source of the number sells something.

On a raw idea, nothing has been independently checked, so nothing is called verified. The numbers you supply count as unverified, or as vendor claims where their source sells something.

Lowest and first move

The scorecard names the lowest dimension and says what the panel caught there, citing your description or naming the gap. The panel’s read is one of pursue as is, close the gaps first, or walk away, with one sentence of reasoning. The decision stays yours. The first move is a single next action taken from the executor advisor’s answer.

What a grade means

A grade is a stress-test, not a certification. A pass means no confident weakness was left un-named. A low score is the useful result.

A passing score does not predict that the business will succeed, and a low score does not mean it will fail.

When your description is short

A short description caps what the panel can score, and low scores from missing information are the expected result. Add detail and run /grade-idea again. The score moves as the gaps close.

The full loop

When an idea holds up, four commands turn it into a plan, grade the plan and record your decision. Run them in order. Each one names the next. At the default guided pace, Claude offers each next step and waits for your yes.

/intake

Turns whatever you provide into a business brief: pasted text, a conversation, or files you put in inputs/. It writes outputs/business-brief.md.

It asks only for essentials it cannot find, at most five questions in one batch. It marks every estimate as an estimate. Sections you cannot answer stay under Open unknowns, and it invents no number.

/business-plan

Requires the brief. It writes workflows/active/business-plan-<name>.md with the next ninety days as units. Each unit has an owner and pass or fail acceptance criteria.

Before the Risks and Differentiation sections are final, it runs a pre-mortem (six months on, the business failed: why) and writes the strongest case for the obvious alternative, which the plan then has to answer. It uses only the brief’s numbers. Where the brief says unknown, the plan says unknown, and the grade prices that.

/grade

Five advisors score the plan against the six-dimension rubric. Each sees only the plan and the rubric. Claude saves the result to workflows/scorecard-<name>-loop<n>.md.

At 8.5 or above, it points you to /approve. Below that, it revises only the two lowest dimensions and regrades the whole rubric. Before each new pass it names the loop, reminds you the cap is three, and waits for your yes. Three grading passes is the limit.

If the plan is still below A- at the cap, the loop stops. You get the final scorecard, the top three unresolved weaknesses and your options: proceed at this grade, pivot, or kill. It never proceeds on its own.

/approve

Shows the plan’s offer line, the final scorecard and any open questions. You answer GO, NO or CHANGES. Below A-, a GO must take the explicit form “proceed at this grade.” Claude records your decision, with the date, plan name and grade, in .claude/memory/decisions.md.

On a GO, the plan’s ninety-day units route to Backbrief’s researcher, builder and runner, with the verifier checking each result. Publishing, sending and spending still stop for you. /approve never runs automatically and never on inferred consent.

A worked example: TrailNotes

The business

TrailNotes is an invented business, disclosed as one: a $6 a month weekly email of local hiking trail conditions for one metro. The figures are the example’s premise. The pipeline and every grading panel that ran on it were real.

Intake

The brief’s line on who pays: “Committed local hikers who go out most weekends and plan around conditions pay $6/month, monthly only, no annual plan yet, for the Thursday-night conditions email.”

Intake wanted five answers, including which metro. The owner’s notes held none of them, so each went under Open unknowns and nothing was invented.

Plan

The builder recomputed the cost stack from the brief’s own rates and did not agree with the owner’s total: “a $51.76 gap, 15.2% of the stated figure, that cannot be closed from the brief’s numbers.” The plan wrote “The plan does not pick between the two totals” and carried the gap as an open item.

Grade, loop 1

B, 7.1 average. The two lowest dimensions were Differentiation (5.9) and Market realism (6.4), both for claims nobody had checked: “The claim that no one else does this is the owner’s own, unchecked against any outside source, and the clone risk is unmitigated (risk 7).”

Revise and regrade

Each revision touched only the two lowest dimensions. Loop 2 scored B, 7.4. Loop 3 scored B, 7.6. The remaining defects needed the owner: naming the metro, running a freshness test, exporting payment records. The plan listed those checks as planned, and the panel scored them as not yet run.

The cap

At 7.6 the plan sat 0.9 short of the 8.5 pass line. The loop stopped. There was no fourth pass, nothing rounded up, and nothing unlocked. The final scorecard presented three options: proceed at this grade, pivot, or kill. Its last line reads: “A plan that plateaus is a result, not a failure of the system.”

The full run, with the scores from all three passes, is Part 2 of examples/worked-example-business.md in your project after install.

Other doors in

/grade-video

For a video that pitches a business or a money-making method. Type /grade-video and give a link or the path to a local file. If you give neither, Claude asks for it, along with one line on what you want to know: whether to do what it says, whether its numbers hold, or something else.

Before any fetch, Claude tells you the platform’s terms may restrict automated download and offers a local file or a transcript you have the right to use first. It never logs in to a platform.

It then writes a brief of at most 300 words that cites timestamps, and a steelman of the pitch: the strongest honest reading of what the video says. It builds a ledger of every claim that moves a score, each tagged verified, unverified or vendor claim. A number from someone selling something never lifts the score above what an unverified number would. The advisors then grade the opportunity on the same rubric.

Claude saves outputs/video-scorecard-<slug>.md and outputs/video-claims-<slug>.md. It never tells you to buy, join or sign up for what the video sells. It closes with: “This graded what the video alleges, not what the presenter earns. A claim tagged unverified is unchecked, not false.”

/pick-niche

For someone with no idea yet. It interviews you one question per turn on your hard constraints, weekly hours, money you can risk, skills and the work you refuse to do. Each answer goes into a capture file in context/discovery/ before the next question.

It then scans public complaints for gaps that fit inside those limits, using /find-gap, and returns three candidates. Each is graded on the six-dimension rubric and carries a first move for this week. It orders none of them and picks none for you. It recommends nothing to buy, join or spend on.

Claude saves the run to outputs/niche-picker-YYYY-MM-DD.md.

When something goes wrong

/verify-install: Type it any time something feels off after an install or an update. It prints a plain pass or fail for each check: the installed files, whether the optional permission rules are in force in the running session, whether the shipped reference documents match this release, and whether any commands or skills of your own shadow Backbrief’s. It starts by reporting the installed version from .claude/VERSION.
A missing .claude folder: If a project has context/ and outputs/ folders but no .claude folder, the install was left behind when the project was moved, copied, zipped or synced. The folder is hidden, which makes it easy to leave. /verify-install says so and offers to reinstall from the download.
A nested .claude folder: If /verify-install finds .claude/.claude/, the copy went one level too deep. It offers to move the contents up.
Not sure what to do next: Type /next. It works out where the project stands from the files on disk and offers the next step. Type /bb for the six core actions in plain language.
Help: Write to contact@ashecorp.com from the address you bought with.

Grade the idea. Get the plan.

$29, paid once.

Get Backbrief for $29

All sales are final: refund policy and license terms