How RevampKey’s AI coding agents plan, build and test.
Describe what you want built. A team of specialist AI coding agents researches it, asks you what it cannot guess, designs the screens, writes the code in small steps and tests each one. Nothing is built until you approve the plan.
Follow “let clients book their own appointments” from idea to tested code.
Each specialist is an AI coding agent with one job. Pick a role on the left, or scroll. Everything below is interactive: sort the report, answer the question, approve the screens.
Researcher · investigates
Research lands as a report you can explore.
Before anything is planned, it reads your project and the web, then hands back a report. Not a wall of chat. Sort it, filter it, and see where every claim came from.
Reports / Booking readiness auditAUDIT · REV 1
RevampKey AUDIT
Booking readiness audit
What the appointments app handles today, and what self-serve booking needs first.
Four of the nine booking paths have no handler today.
9Paths in the flow
4Unhandled today
2Decisions for you
High
Two clients can confirm the same time at the same momentNo lock on a time slot today
High
A reschedule can leave the old booking behindThe old time is never released
Medium
A held time never frees if the browser closesHolds have no expiry
Medium
Reminders can go out at the wrong hour across time zonesTime is stored without a zone
Low
A cancel inside 24 hours is not flagged to the front deskNeeds your decision
Low
The confirmation has no calendar attachmentNice to have
Where your idea leaves a gap, it asks, one clear choice at a time, with a recommendation when it has one. Your answers are written into the plan, so nobody has to remember them later.
2 questions before I planOne more answer and it goes into the plan
Every path mapped. Every decision written.
The answers become a map of everything that can happen, including the unhappy paths, and a short list of rules the build must respect. You correct them here, while changing your mind is free.
Flow · booking lifecycle
Time pickedHeld for 5 min
Still free?Taken · offer next 3
Booking confirmedReminder sent
Spec · booking.md
Decisions
A time has at most one confirmed booking.
A hold lasts 5 minutes, then frees the time.
A reschedule moves the booking, never doubles it.
A cancel inside 24 hours is waiting for your answer.
Designer · designs
Screens you approve before any code.
You see what clients will see, including the awkward moments like “that time was just taken”, while changing it still costs nothing.
Pick a time
Pick a time
Shown in your time zone, London.
TUE 13 OCT
WED 14 OCT
09:3013:0016:30
We hold 11:30 for 5 minutes while you confirm.
Confirm booking
Just taken
That time was just booked
Someone confirmed 11:30 a moment ago.
11:30 on Tue 13 Oct is no longer free.
CLOSEST FREE TIMES
12:0015:00Wed 11:3011:30
Book 12:00
Booked
You’re booked
Haircut with Priya
Tue 13 Oct, 11:30
A reminder goes out on Mon 12 Oct.
RescheduleCancel
3 screens · version 2. Waiting for you.
Engineer · builds
Built in steps, each one checked.
The approved plan is split into small steps, each with its own checklist. Every finished step is saved as its own change you can read, so there is no giant pile of code at the end.
Plan · self-serve booking · 4 steps · 15 tasks
Slots and holds
4 tasks · 2 checks passed
a3f9c21feat(booking): hold a slot for 5 minutes
Done
Booking and rescheduling
5 tasks · 3 checks passed
7be04d8feat(booking): a reschedule moves, never doubles
Done
Reminders
3 tasks · 2 checks · working now
In progress
Front-desk view
3 tasks · 2 checks
Next
QA · checks
Checked by running it, not by reading it.
Each step is tested by actually using it, on every screen size, before the next one starts. A test that was never run is never reported as passing.
$ npm test -- booking ✓ a second client cannot confirm a held time ✓ an expired hold frees the time ✓ a reschedule leaves exactly one booking ✓ reminders follow the stored time 21 passed0 failed
Desktop, tablet and phone: every screen size rendered and looked at.
You answered 2 questions and approved 3 screens. The plan, the decisions and every change are saved in your project, so anyone can see why it was built this way.
Every path, written down
Flows and specs catch what a prompt would have missed.
Before screens are drawn, every path through the feature is mapped, including the ones people forget: the time that was just taken, the cancel that comes too late. Each decision you make is written into the spec.
Each specialist receives exactly what the one before produced and returns something you can read and check.
Specialist
Receives
Hands back
Researcher
Your question and your project
A report you can sort, filter and trace to its sources
Product manager
The report and your idea
Questions with a recommendation, then a written spec and a flow of every path
Designer
The flow and the spec
Screens for the key states, including errors, for you to approve
Engineer
The approved plan
Modules with checklists, each finished one saved as its own commit
QA
Each finished module
Checks that were actually run, on desktop, tablet and phone sizes
Built in steps
Small steps, each one committed and checked.
The approved plan becomes modules with their own checklists. Each finished module is a single commit, tied to the checks it had to pass, so there is never a giant pile of changes to review.
You stay in charge
Where you decide, and where it just works.
You approve the plan
Reports, plans, flows and screens come to you first. Correct the direction while changing it still costs nothing.
It asks before it assumes
Gaps in your request become clear questions with a recommendation, not silent guesses buried in code.
Everything is reviewable
The plan, your answers and every change live in your project, so anyone can see why it was built this way.
You talk to one AI coding assistant. For larger work it hands parts to specialist agents: a researcher, a product manager, a designer, an engineer and QA. Each works with its own context and reports back to you. How delegation works
Does every request go through all the steps?
No. A small fix is done and checked directly. A medium change gets a short plan, and a large one gets the full team. RevampKey decides how much structure a request needs, and a missing plan does not mean missing work.
When does it ask me something?
When your request leaves a gap that changes the result. It asks one clear choice at a time, recommends an option when it has one, and writes your answer into the plan so the decision is not lost.
Can I stop it before any code is written?
Yes. Plans, flows and screens are shown for review first, and you approve before building starts. You can correct the direction at any of those points. Plans and modules
How do I know the code works?
Each step is tested before the next one starts, and each finished step is saved as its own commit you can read. When the project has no tooling for a check, it says so rather than reporting a pass. Changes and verification
Keep exploring
Plans and reports
Reports you can sort and question, and plans split into reviewable steps.