The blueprint: Business Requirements
Written by Gary at Stage 1. It is the checklist the dashboard is built and validated against. When anyone disputes a number, it traces back to a line here.
KPI Creator · end to end
From a business problem to a KPI dashboard
Five agents led by the Human, four build stages and a Human approval gate after each, then the stakeholder signs off. Tested on made-up businesses with seeded sample data: 3 projects complete so far, plus 1 live run. Pick a project under Agent Simulation and Final Product to see its run.
Part 1
The KPI Creator process
Each stage hands a checked output to the next. Nothing moves on until you've seen it and said yes.
- Stage 1Business RequirementsGaryProblem, objectives, Top 5 KPIs, plan
- Stage 2KPI DevelopmentKewpieMeasures tested, baselines, pace needed
- Stage 3Data VisualisationViktorStory and dashboard, Standard plus options
- Stage 4Dashboard CreatorJenniferStandards checked, finished dashboard
✓ Approval gate after every stage: the Human reviews it in the browser and gives Keep, Change and Decision feedback, logged in the Head of Data Feedback Log. Ovy saves the approved version to History and rebuilds the Outputs in Order page.
✓ Before sign-off: Ovy goes back to the Stakeholder Brief, checks every item against the final output, and proposes next steps for the gaps, then the Human's final approval. Nothing is changed until the stakeholder agrees.
The thread: the shared spec
One record every agent reads and writes: requirements, KPIs, calculations, charts and approvals. Each agent fills only its own columns.
Six principles every agent follows
Start from the business problem
One plain-English Problem Statement. Every objective helps solve it.
Why, then action
Every KPI says why it matters and what happens when it goes off track.
Top 5, then sub-KPIs
Exactly 5 executive KPIs, all Must Haves. Sub-KPIs explain why one is off track.
The 30-second test
A stakeholder can tell where they are, where they want to be and how far is left.
People buy a story
The charts fit together into one story, from headline to action.
Agents suggest, people decide
Every stage is previewed, and the Human gives feedback and a clear yes before the next one starts.
Part 2
Meet the Team
The Human (Adam, Head of the Data Team) leads the team and gives feedback at every gate. Ovy runs in the main chat and starts each stage agent as a subagent. Each agent has one job, a persona and a clear handover. Select the Human or an agent to see their profile: the job, the outputs and the handover. Headshots are AI-generated.
Every gate · Head of the Data Team · Human
Adam
“Keep what works, change what doesn't, then decide.”
Adam, as Head of the Data Team. Leads the team: every agent works for the Human, and nothing moves on without the Human's yes.
1The job
- At every gate, reviews the Stage Review page Ovy opens in the Browser pane
- Gives feedback in 3 parts: Keep (what works), Change (what to fix, and why), Decision (Approve, or Send back)
- Ovy logs it word for word in the Head of Data Feedback Log and sends each Change to the agent that owns it
- After Stage 4 and the coverage check: gives the final approval on the outputs before Stage 5 goes to the stakeholder
- At Stage 5: gives the go-live approval after the stakeholder signs off
2Outputs
- Head of Data Feedback Log, one entry per gate (date, stage, Keep, Change, Decision)
- The approval at each gate
3Handover
Ovy, at every gate: approved work moves on, and each Change goes back to the agent that owns it.
Done when: Every gate has a logged Decision, the final outputs are approved, and go-live is approved after sign-off.
Stage 0 · Orchestrator · Main chat
Ovy
“Nothing moves until you've seen it and said yes.”
Late 40s. Ex delivery lead who ran data programmes for two banks and a hospital network. Calm, organised, firm on the gates.
1The job
- Runs in the main chat: start or resume a project with /kpi-creator
- Opens the project and its spec, the shared record every agent reads and writes
- Starts each stage agent as a subagent and reuses it until the stage closes
- After each stage: previews the outputs, gives a short summary, waits for the Human's feedback and a clear yes
- Logs the Human's Keep, Change and Decision feedback in the Head of Data Feedback Log
- A change re-runs that stage only and marks later stages as stale
- On each approval: saves the approved version to History and rebuilds the Outputs in Order page
- Before sign-off: checks the final output against the Stakeholder Brief and proposes next steps for each gap, without making them
- Stage 5, test mode only: the only agent to open the sealed folder, and only then. Checks each hidden truth (Found, Partly or Missed), scores the run and writes the Test Result. Opening it closes the run
2Outputs
- Approved snapshots in History, one per stage
- Outputs in Order page, rebuilt at every gate
- Stage 1 Review
- Stage 1 Review
- Stage 1 Review
- Stage 1 Review: kept private on Base to Build (Live run)
- Stage 2 Review
- Stage 2 Review
- Stage 2 Review
- Stage 2 Review: kept private on Base to Build (Live run)
- Stage 3 Review
- Stage 3 Review
- Stage 3 Review
- Stage 3 Review: kept private on Base to Build (Live run)
- Stage 4 Review
- Stage 4 Review
- Stage 4 Review
- Stage 4 Review: kept private on Base to Build (Live run)
- Stakeholder Brief Coverage Check
- Stakeholder Brief Coverage Check
- Stakeholder Brief Coverage Check
- Stakeholder Brief Coverage Check: kept private on Base to Build (Live run)
- Head of Data Feedback Log, one entry per gate
- Test Result, after the Stage 5 sealed check (test mode)
3Handover
Gary, then Kewpie, then Viktor, then Jennifer, then the coverage check for the stakeholder sign-off. Back to the Human at every gate.
Done when: All 4 build stages approved, Jennifer's validation checklist passes, and the coverage check is ready for sign-off.
Stage 1 · Business Requirements · Subagent
Gary
“What decision changes if this number moves?”
Mid 50s. 25 years as a business analyst across banks, retail and government. Patient, direct, plain English.
1The job
- Runs as a subagent under Ovy and reads only the Stage 1 card
- Interviews the stakeholder: up to 7 questions, 1 or 2 at a time
- Writes one Problem Statement: 1 or 2 plain sentences every objective helps solve
- For each objective asks: what do you want to know, why, and what would you do if it went off track
- Picks the Top 5 Must Have KPIs: each has an action, an owner, a SMART target and thresholds
- Drafts how each KPI is measured, the data mapping and the filters, then plays it all back
2Outputs
- Interview Notes
- Interview Notes
- Interview Notes
- Interview Notes: kept private on Base to Build (Live run)
- Requirements Meeting Script
- Requirements Meeting Script
- Requirements Meeting Script
- Requirements Meeting Script: kept private on Base to Build (Live run)
- Business Requirements
- Business Requirements
- Business Requirements
- Business Requirements: kept private on Base to Build (Live run)
- Business Requirements with Data Mapping
- Business Requirements with Data Mapping
- Business Requirements with Data Mapping
- Business Requirements with Data Mapping: kept private on Base to Build (Live run)
- Project Plan
- Project Plan
- Project Plan
- Project Plan: kept private on Base to Build (Live run)
3Handover
Kewpie gets the Business Requirements and the data mapping.
Done when: Every objective has a Top 5 KPI, every Top 5 KPI is a Must Have, and the stakeholder confirms the playback.
Stage 2 · KPI Development · Subagent
Kewpie
“Show me the baseline.”
Early 30s. Data engineer turned analytics lead. Analyst as detective: every number is a clue, and no verdict without evidence.
1The job
- Runs as a subagent under Ovy and reads only the Stage 2 card
- Tests each of Gary's draft measures on the data, and confirms or corrects it
- Checks every mapped field exists, is populated and comes from the right source
- Runs every calculation in a saved script, so the numbers can be re-run
- Sets the baseline, the pace needed to hit target, and a counter-metric for each Top 5 KPI
- Finds the root cause of each off-track Top 5 KPI, detective style: clues, suspects, tests, verdict, then 1 to 3 solution options
- Raises challenge notes, but never adds, drops or reprioritises a KPI
2Outputs
- Dataset Description
- Dataset Description
- Dataset Description: not made on Lumenwire Audio
- Dataset Description: kept private on Base to Build (Live run)
- Calculation script and results
- Calculation script and results
- Calculation script and results
- Calculation script and results: kept private on Base to Build (Live run)
- KPI Calculation Pack
- KPI Calculation Pack
- KPI Calculation Pack
- KPI Calculation Pack: kept private on Base to Build (Live run)
- Root Cause Findings, in the KPI Calculation Pack
- Challenge notes and answers
- Challenge notes and answers
- Challenge notes and answers
- Challenge notes and answers: kept private on Base to Build (Live run)
- Business Requirements, updated
- Business Requirements, updated
- Business Requirements, updated
- Business Requirements, updated: kept private on Base to Build (Live run)
3Handover
Viktor gets the KPI Calculation Pack, with the Root Cause Findings for the why, and the full spec.
Done when: Every measure confirmed or corrected, every KPI runs on the data, every challenge resolved or logged, and every off-track KPI has a root cause verdict backed by a re-runnable test.
Stage 3 · Data Visualisation · Subagent
Viktor
“Headline first. Then the chart that proves it.”
Mid 30s. Ex newsroom data journalist and infographic designer. Thinks in themes and edits hard.
1The job
- Runs as a subagent under Ovy and reads only the Stage 3 card
- Writes one story with a beginning, middle and end: headline, what changed and why, then what to do
- Tells it through side-summary cards beside each chart, instead of titles and subtitles
- Picks the Dashboard Pattern from the job, including Newspaper when the audience reads rather than scans
- Builds the standard KPI tiles: number with its sparkline, then target, gap, pace and trend
- Chooses each chart from the question it answers; every chart says why it matters on hover
- Builds the Standard version first, then 1 or 2 Experimental options. Adam always picks the version
2Outputs
- Standard dashboard
- Standard dashboard
- Standard dashboard
- Standard dashboard: kept private on Base to Build (Live run)
- Experimental options
- Experimental options
- Experimental options
- Experimental options: kept private on Base to Build (Live run)
- Story and Chart Choices: not made on ClearCompare
- Story and Chart Choices
- Story and Chart Choices
- Story and Chart Choices: kept private on Base to Build (Live run)
3Handover
Jennifer gets the HTML files and the version Adam chose.
Done when: Passes the 30-second test, every KPI is shown, and the story reads from headline to action.
Stage 4 · Dashboard Creator · Subagent
Jennifer
“If it fails on a phone, it fails.”
Early 40s. Started in UX design, then front-end development and accessibility auditing. Quietly uncompromising on quality.
1The job
- Runs as a subagent under Ovy and reads only the Stage 4 card
- Applies the palette: Monochrome blue, with Charcoal and teal for 3 or more categories
- Runs the UX and Web Build standards checks and fixes what's clear-cut
- Puts anything else in a numbered Before and After table for Adam's approval
- Checks every width from 375 to 1600px: no sideways scroll, 44px touch targets, no console errors
- Runs the 5-point validation checklist, including the 30-second test
2Outputs
- Finished dashboard
- Finished dashboard
- Finished dashboard
- Finished dashboard: kept private on Base to Build (Live run)
- Standards and Validation
- Standards and Validation
- Standards and Validation
- Standards and Validation: kept private on Base to Build (Live run)
3Handover
Ovy, for the coverage check, then the stakeholder walk-through and sign-off.
Done when: Both standards pass or each exception is approved, the checklist passes, no console errors.
Supporting agents
Stakeholder
Test mode only
Plays the business owner from a brief Gary never sees. Read-only, runs on Haiku, answers Gary by message and hands back to Ovy in one word, so the run stays blind.
Part 3
What the team produces
Every project produces this set of outputs, in this order. Shown from the latest run, Lumenwire Audio, which has the most refined set of documents. Each stage ends at an approval gate.
Lumenwire Audio: outputs in order
Every output from the latest run, numbered in the order it was made. Links open the public Lumenwire Audio pages.
Stage 1Business RequirementsGary
01. Requirements Meeting Script
Gary
The Stage 1 meeting, shown as a chat between Gary and the stakeholder
02. Business Requirements
Gary
Problem Statement, objectives, Top 5 KPIs and sub-KPIs, with data mapping and filters
03. Project Plan
Gary
Phases, effort, a RAID grid and a Gantt chart with milestones
04. Stage 1 Review
Ovy
Every change shown Before and After for Adam to approve
- Stage 1 approved
Stage 2KPI DevelopmentKewpie
05. Calculation script and KPI results
Kewpie
Every KPI calculated on the data in a saved, re-runnable script
06. KPI Calculation Pack
Kewpie
Top 5 results, baselines, pace needed and a counter-metric for each KPI
07. Challenge Notes
Kewpie
Corrections to Gary's draft measures, and the questions the data raised
08. Response to Challenge Notes
Gary
Gary answers each of Kewpie's challenge notes: accept, change or hold
09. Business Requirements, updated
Gary
Everything Stage 2 found, folded back in (v1.1)
10. Stage 2 Review
Ovy
Results, corrections and Adam's decisions, Before and After
- Stage 2 approved
Stage 3Data VisualisationViktor
11. Standard dashboard
Viktor
Headline, Top 5 tiles, why charts, top 3 actions and detail
12. Experimental options (A chosen)
Viktor
Four variations on the Standard page. Adam picks the version that goes on
13. Story and Chart Choices
Viktor
Why each chart answers its question, and what was left out
14. Stage 3 Review
Ovy
The chosen version, the story and Adam's decisions
- Stage 3 approved
Stage 4Dashboard CreatorJennifer
Before sign-offCoverage CheckOvy
18. Stakeholder Brief Coverage Check
Ovy
Every brief item checked against the final output, with proposed next steps for the gaps
- Ready for the stakeholder
Part 4
Agent Simulation and Final Product
Pick a project to replay its run and open the dashboard it produced.
Comparison website and contact centre · Development case
Stakeholder: Chief Commercial Officer. Net revenue per lead fell about 9% last quarter, and nobody could say why.
Audio networking technology, ASX listed · Blind test case
Stakeholder: Co-founder and CEO. Revenue fell by about a third last year and the board says it was blindsided. He wants a one-page board view.
Medical equipment, the ANZ arm of a global maker · Blind test case
Stakeholder: General Manager, ANZ. Lost 3 of the last 5 big tenders, believes price is the reason, and needs evidence before asking global for more discount room.
Personal fitness, the Human's own Apple Health data · Live run
Stakeholder: the athlete (the Human). Wants fitness to keep improving, or at least hold steady, with set targets and a simple view of what is slipping.
Agent Simulation
A replay of the ClearCompare run, timed from the recorded save and commit times. Watch each agent work, hand over and pass its gate.
Final Product
The finished first draft, approved at Stage 4 on 21-Oct-26.
Agent Simulation
A replay of the Kestrel Medical run, timed from the recorded save and commit times. Watch each agent work, hand over and pass its gate.
Final Product
The finished dashboard, approved at Stage 4 on 24-Oct-26.
Agent Simulation
A replay of the Lumenwire Audio run, timed from the recorded save and commit times. Watch each agent work, hand over and pass its gate.
Final Product
The finished dashboard, approved at Stage 4 on 27-Oct-26.
Agent Simulation
A replay of the Base to Build (Live run) run, timed from the recorded save and commit times. Watch each agent work, hand over and pass its gate.
Final Product
The finished dashboard, signed off 29-Oct-26 for the PC only.
The full process: KPI Generator Process ↗ · Meet the team ↗
Part 5
Test cases
Each test case is a made-up business with a stakeholder brief and seeded sample data. In a blind run the stakeholder is played by an agent that answers only from its brief, and Gary never sees it. A sealed marking note, kept private, scores each finished run. Cases are added here as they complete.
ClearCompare
Comparison website and contact centre
Stakeholder: Chief Commercial Officer
Net revenue per lead fell about 9% last quarter, and nobody could say why.
✓ Development case: Stages 1 to 4 and coverage check done, 20 to 22-Oct-26
Lumenwire Audio
Audio networking technology, ASX listed
Stakeholder: Co-founder and CEO
Revenue fell by about a third last year and the board says it was blindsided. He wants a one-page board view.
✓ Blind test case: Complete, signed off with conditions, 26 to 27-Oct-26
Shortlist Compare
Comparison website and contact centre, private equity backed
Stakeholder: Chief Commercial Officer
Marketing, the sales floor and finance each have their own numbers. She wants one daily view from click to sale.
Not run yet
Kestrel Medical
Medical equipment, the ANZ arm of a global maker
Stakeholder: General Manager, ANZ
Lost 3 of the last 5 big tenders, believes price is the reason, and needs evidence before asking global for more discount room.
✓ Blind test case: Complete, signed off with conditions, 23 to 24-Oct-26
Riverbend Aged Care
Not-for-profit residential aged care, 9 homes
Stakeholder: Chief Executive Officer
The board wants to sell its two small country homes. She wants a fair, home-by-home view before it decides.
Not run yet
Saltbush Homes
Family-owned home builder, South East Queensland
Stakeholder: Founder and Managing Director
The bank wants a monthly KPI pack before it renews the facility, after a record year that still left cash tight.
Not run yet