Quarrybank

AskApproveBuildGateMergeHQ

01 · The ask

A plain request with real data becomes a refined ticket.

The marketing manager asks for a conversion improvement and attaches last month's data. The agency's rules attach themselves. The product owner agent writes the ticket: approach, files, what "done" means, effort. It joins the sprint plan.

Proves. Shaped by the agency's own rules, not a prompt. Nothing starts untracked.

Claude · marketing managerQuarrybank layer on

You

pricing-oct.csv 2,140 rows · last month's conversions

Heatmap hotjar.link/pr-oct · mobile and desktop

Claude

Read the file. October converted at 2.1% overall and 1.4% on mobile; 61% of visitors left before reaching the form. The heatmap agrees: on mobile the form sits two screens down.

Applying your rules: brand voice, design system, CRO checklist. No new components, copy stays in voice.

done

Ticket SITE-112 filedMove the form into the hero on mobile; two headline variants on the A/B flag. Done when: form above the fold, variant live behind the flag, tests green, review pass. Cap £4.00.

done

Added to sprint plan 4212 of 15 tickets. The client manager approves the plan once; the lanes do the rest.

done

Copy brief COPY-14 queuedPricing hero copy and a supporting article, in the client's voice, through the gate.

queued

Baseline recorded for the client report2.1% → target 3% by end of November. The ledger will show the change.

Anything to change before it builds?

AskApproveBuildGateMergeHQ

02 · One approval, not twenty

The client manager approves the plan once. The lanes fill themselves.

Fourteen tickets, each with its definition of done and cost cap, one approve. Auto-start fires them into the build lanes on the agency's own machine, with a cloud box as fallback.

Proves. Sprint-level approval. Owned hardware doing the work.

HQ · Mission control · on decksprint 42

Sprint plan 42 · 14 tickets · cap £484 s · 8 m · 2 l · every ticket has a DoD and a cost cap

Approve plan

builder-1idle

builder-2idle

builder-3idle

builder-4idle

tester-1QA

tester-2idle

tester-3offline

reservestandby

SITE-112 · effort mLift pricing-page conversion: form above the fold, two headlines on teston deck · site repo

ADS-457 · effort sQ4 campaign landing variant so paid traffic stops bouncing to the homepageon deck · ads repo

CMS-203 · effort mCase-study block so account managers publish proof without a developeron deck · cms repo

FORMS-88 · effort sCapture UTM on every lead so attribution is right in the client reporton deck · forms repo

HUB-31 · effort lMove HubSpot form auth to the new key before the old one expires Fridayon deck · protected path

COPY-14 · effort sPricing hero copy and a supporting article in the client's voiceon deck · content line

AskApproveBuildGateMergeHQ

03 · Watch the build, meter the spend

Every session is watchable and metered. Nothing runs away.

The paid ads executive's landing variant is building in lane 2. Every read, edit, command and test result streams into HQ. Turns, spend and time are capped; a breaker stops the run if a cap trips. Click a step on the rail to see what happened there.

Proves. Visibility and cost control while the work happens, not after.

HQ · session · ADS-457 · builder-2streaming

Turns0 / 80

80 turns, effort m

Spend£0.00

£4.00 per session

Cache hits—

of reads served from cache

Elapsed0m

50 min per session

templates/campaign-q4.php+41 −6blocks/hero/+58ab/flags.json+4breaker: armed · trips at any cap

AskApproveBuildGateMergeHQ

04 · Copy goes through the gate

Content passes a judge that never marks its own work.

The pricing hero copy and its supporting article are drafted in the client's voice skill, then checked twice: hard rules a script can test, and editorial rules a separate model judges. One claim fails. It is rewritten and passes.

Proves. Quality is enforced, not hoped for. (Code takes a different route: QA runner and a DoD reviewer, next scene.)

Content line · COPY-14 · pricing hero + articlegate

RUN 1Gate · 1 of 9 failedclaims-supported · "3x" has no source

AUTORedraft sessionwriter · 2 turns · £0.11 · one line rewritten

RUN 2Gate · 9 of 9 passapproved 14:52 · published to WordPress

RULESET brand-voice v7 · aeo-quality · run 1

DETERMINISTIC
word count 2,480 / 2,500PASS
no em dashesPASS
internal links ≥ 3PASS
focus keyword in openingPASS
AI-JUDGED
answer-firstPASS
human voicePASS
claims supportedFAILED
no paddingPASS
UK EnglishPASS

VERDICTSEND BACK

Draftpost · /pricingvoice: client-b2b-plain2,480 words

Pricing that scales with usage, not with seats

Hero copy · H1 + standfirst · supporting article below

Start on the free tier, pay for what the team actually runs, and see the bill before it arrives. No minimum seats, no annual lock-in.

Teams switch in a day and cut tooling cost by 3x. Two lines of code to connect, no extra infrastructure, and the same SLAs as the rest of the platform.

Gate · claims-supported "3x" and "in a day" have no source in the client brief. Sent back 14:43:12.

Redraft session · auto rewrote one sentence from the migration guide. 14:44:02. No other text touched.

Supporting article · "How usage-based pricing changes a team's first month" · 2,140 words · 4 internal links · answer-first

Content passes the gate. Code passes QA and a DoD reviewer.

AskApproveBuildGateMergeHQ

05 · Reviewed, repaired, merged

Reviewed, repaired, merged. In that order.

A pull request is tested by the QA runner, then judged by a DoD reviewer against the ticket's definition of done. Pass, and the merge train takes it one at a time with tests rerun. Fail, and a repair session fixes it on the same branch. A change to auth or client data waits for a named person.

Proves. Docs and tests merge themselves, normal code merges after review, sensitive paths wait for a human. Honest about repair.

HQ · merge train · site repo3 in queue

ONE AT A TIME · TESTS RERUN ON EVERY CARRIAGEposition

ADS-457 · PR 221 · Q4 landing variantMERGED 15:02

docs + template

QA 42/42DoD 7/7

SITE-112 · PR 222 · Pricing form above fold2nd · rerunning tests

normal code

QA 38/38DoD 6/6

HUB-31 · PR 223 · HubSpot form auth3rd

Review requiredprotected path · config/auth

REPAIR · CMS-203 · builder-3repairing

14:21:30DoD review: 1 of 7 failed

14:21:31relevance threshold untested at 0.5

14:21:55repair session 1 of 2 · same branch · £1.88 of £7.00

£1.88 / £7.00then: back to the train

The method inside

The rules ship with a way of working, not just a style guide.

Campaign management, design systems and editorial practice, codified from people who ran agencies and products for a decade. Sprint goals and retros, definitions of done, concerns logged at every planning, a client view every week. Your team inherits it on day one.

CAMPAIGN MANAGEMENT

Small sprints, one approval

Goal-sized sprints of 10–20 tickets, approved once. The reviewers and the train handle item-level decisions; a person only sees what touches client data, config or spend.

DESIGN

Approved systems, no new components

Each client starts with a brand kit: design system, voice, claims policy. Pages compose from it; the review rejects anything outside it.

EDITORIAL

Answer first, claims sourced, no padding

Writers draft in the client's voice from a research dossier. Every claim traces to a source or it does not pass.

SessionsEvery live session with its journey, stream, files and tests. Click a step, see what happened there.

LedgerEvery output ever, with verdict, operator and cost. Filter by client, sprint or person.

Client viewThe same ledger with one toggle: outputs, verdicts and turnaround for the client, never costs or model names.

The operating layer for agencies that run on AI.

Paid ads, marketing and client teams who want AI to do the work, with rules, limits and a review before anything ships.

A plain request with real data becomes a refined ticket.

The client manager approves the plan once. The lanes fill themselves.

Every session is watchable and metered. Nothing runs away.

Content passes a judge that never marks its own work.

Reviewed, repaired, merged. In that order.

What the company sees. Every session, every lane, every pound.

Your agency HQ

The rules ship with a way of working, not just a style guide.

Small sprints, one approval

Approved systems, no new components

Answer first, claims sourced, no padding

A few agencies a year.

Bring one brief. We run it through live.