Product development for teams whose AI writes the code

Proofboard takes a business objective and ships the work that tests it, faster, with the team you have, and with risks caught before the code is written and while it's being built.

Write the objective with a number on it. Proofboard finds the rules your org already has, proposes theses on how to reach it, each written so it can turn out wrong, and builds every commitment one reviewed checkpoint at a time with agents checking agents. What reaches your people is short: what was found, what was fixed, what passed.

See how it works, start to finish
Runs in your repo · Your agents and keys · Start with one epic
Objective example
Cut involuntary churn in half by end of Q4 Involuntary churn 4.1% → 2.0%
Theses · proposed by AI, challenged by the red team
Recovering failed payments cuts churn Wrong if: under 15% of failed charges recover within 60 days
Expiring cards get updated before they fail Wrong if: card-expiry failures don't fall by a third

The problem

AI writes specs and code faster than anyone can check them, and nothing ties the code back to the objective it was for.

01

The spec drifts from the objective

Nobody wrote down which business result the work was for, so nobody notices when the spec stops serving it.

02

The code drifts from the spec

An agent opens forty PRs over lunch. Your senior engineer reads three. The rest get a thumbs-up.

03

Nobody can say what happened

What was asked for, what was reviewed, what was found: a Slack thread and a half-deleted ticket.

How it works

One example, in the order it would really happen: a payments team with a churn objective. The rule conflict is found and settled before any code exists, and nothing is delivered while a rule it depends on is still open.

Steps 1 to 3 · Objective, Org Mind, a decision

The objective, and the rules it runs under

Write the objective with a number on it: a business goal with a measure, the way an epic should be.

Org Mind reads where your team writes things down, in Confluence, Jira, READMEs and the code itself, and works out the rules: what applies where, what's missing, and where two sources disagree, with a link to each source. A person settles each conflict once. When a rule is settled, it stays settled: every commitment that touches that part of the system is checked against it, and if the code drifts from it later, you hear about it. Org Mind also runs on its own inside the coding agent you already use.

Objective example
Cut involuntary churn in half by end of Q4 Involuntary churn 4.1% → 2.0%
Org Mind · billing example
Conflict Retry window for failed charges: 3 days or 7?
Confluence says 3 daysretry.py comment says 7 days
Settled · by a person 7 days. Confluence updated. Applies to billing/*.
Checked on every commitment touching billing
Gap No security owner named for payments.
Step 4 · Theses

Theses: how the objective might be reached

Proofboard proposes theses. Each thesis is a bet on how to reach the objective, written so it can turn out wrong, and the red team agent challenges them before anyone commits to one. Every thesis carries a "wrong if" line that tests it on its own measure, not the objective's. Recovery can reach 15% while churn doesn't move; Proofboard keeps the two questions separate.

Under each thesis sit the commitments that test it. A commitment is roughly a story, with its own spec and plan.

Thesis example
Recovering failed payments cuts churn Wrong if: under 15% of failed charges recover within 60 days
Commitments · under this thesis
Retry failed card chargesdelivered · 7 checkpoints
Dunning email sequencein progress
Card updater integrationspec
Steps 5 and 6 · Commitment, red team

Building a commitment

For each commitment, agents write the spec and the plan, and the red team agent challenges both before any code exists. Where it finds a hole, a person answers the question. Then the work is built one checkpoint at a time, each a reviewed step of the commitment, and the red team reviews each one before the next starts.

A finding is what the red team or green team reports, with a severity. Blockers are fixed and re-reviewed before the work moves on. Your team reads findings, not diffs.

Commitment · Retry failed card charges example
L
Spec and plan written by agents, challenged by the red teamThen built in 7 checkpoints, each reviewed before the next.
R
Finding · blocker“A retry can double-charge if the first attempt times out.” Fixed in checkpoint 4, re-reviewed.
Step 7 · Green team, delivered

Delivery, and what your team reviews

At the end, the green team agent checks the finished commitment against the spec's acceptance criteria, including the rule settled in step 3. Code can be well built and still solve the wrong problem, which is why this check is separate from the red team's.

What reaches your people is short: what was found, what was fixed, what passed. Your team approves, or sends back the checkpoint that missed. Delivered means the commitment met its spec. Nothing is delivered while a rule it depends on is still open.

Step 8 · Judged, twice · roadmap

Did 15% of failed charges recover? Did churn fall? The first question tests the thesis, the second the objective. A thesis can hold while the objective isn't met, and Proofboard never lets one stand in for the other. Measuring both inside Proofboard is on the roadmap.

Delivered · Retry failed card charges example
R
Red team: 1 blocker, fixedFixed in checkpoint 4 and re-reviewed.
G
Green team: 6 of 6 acceptance criteria metIncluding the 7-day retry window settled in step 3.
Approve Send back a checkpoint…
Judged, twice roadmap
T
ThesisDid 15% of failed charges recover within 60 days?
O
ObjectiveDid involuntary churn fall from 4.1% toward 2.0%?

Why it is faster

Where the time goes today is not typing code. It goes to grooming meetings, specs that get rewritten twice, a rule somebody knew but nobody wrote down, and senior engineers reading diffs. Proofboard does the grooming, the spec and the plan with agents, surfaces the rule before the first line is written, and builds whole commitments end to end. Your people answer the questions agents can't settle and read findings instead of diffs. The verification steps add minutes. The rework they prevent used to cost weeks.

Refactoring

Opportunistic refactoring, automated

Engineers fix the debt in the code they're touching when they have time. Proofboard does it every time. It keeps a record of what's wrong in your repo against the standards you already enforce. When a commitment's plan touches files with known problems, the agents fold the fixes that make sense into the work, as their own reviewed steps that run before the feature. You see what was folded in and can take any of it out. When the work lands, the record updates: fixed, still open, or came back.

Annual plan toggle · folded in example
billing/service.pyRaises HTTPException outside the API layer folded in
billing/crud.pyCommits a transaction inside a CRUD function folded in
billing/plans.pyCalls the payment vendor from business logic another team, in progress
2 folded in, run first Take any out
Code health · after this commitment example
39 open 8 in progress 32 fixed 1 came back
Jira, and where this sits

Bring in what you already have

Proofboard is where product work lives, from the objective to the delivered commitment. New work starts here. Existing work comes in from Jira: point it at a project or an epic, it groups the stories into theses, says why, maps every status to a stage, and shows you the whole tree before anything is written. Comments come along. Run it again any time, roll it back if you don't like it.

Import preview · epic PAY-100 example
PAY-112Retry failed card charges
PAY-118Dunning email sequence
PAY-131Card updater integration
PAY-140Annual plan toggle
PAY-144Proration on upgrade
→
Thesis Recovering failed payments cuts churn Why: all three act on charges that already failed.
PAY-112PAY-118PAY-131
Thesis Annual plans bring cash in up front Why: both change how plans are billed.
PAY-140PAY-144
Preview only · nothing written yet Confirm import

The agents, and your team

Four roles. Each answers one question, and none of the agents decides what to build.

L
Lead agent
“Build this checkpoint.”

Writes the spec, the plan and the code, one checkpoint at a time, to your team's standards.

Doesn't decide what to build or whether it's good.

R
Red team agent
“Is it built well?”

Challenges the spec, then the plan, then every checkpoint.

Findings by severity: blocker major minor. The lead fixes them before moving on.

G
Green team agent
“Is it what was asked for?”

Checks the finished commitment against the spec's acceptance criteria.

If the spec said “users can undo,” an undo button that doesn't undo fails.

Y
Your team
“Ship it?”

Answers the spec questions the agents can't settle. Approves the delivery, or sends back the checkpoint that missed.

Reads findings, not diffs.

A second model checks every run. It reviews the same work on its own and scores how much it agrees. Low agreement tells you where to look.

Bring your own agents and keys: Claude Code, Codex, Gemini CLI and Mistral Vibe. Assign any of them to lead, red team or green team and pick the model for each. Put a different model family on review if you want real disagreement. The board lives on your machine, and agents call their providers with your keys.

Start with one team.

Pick one team and one epic you haven't groomed yet. Leave your work email and we'll reply within two days with what a pilot looks like and what it costs.

Thanks. We'll reply within two days.