ML.ai plans on a lighter model and saves the frontier tier for writing the diff, so the same change costs about 30% less.
Add a dueDate field to tasks: schema, API, and the task card UI.
Read schema.prisma, route.ts and TaskCard.tsx
Routed into 4 stepsFrontier on every step $0.70
Standard$0.04 Standard$0.08 High$0.22 High$0.15Run completeSpent $0.49
Saved 3 files · typecheck passed
Describe a task or ask a question
for feature workDevelopers shipping ordinary feature work that is clear enough to describe in a sentence or two.
A field that has to reach the schema, the API route and the UI in one change.
New endpoints, settings panels and forms you can describe in a sentence or two.
The whole change across every layer, in front of you before anything saves.
Routing saves money on the steps that allow it. These four things are what keep the output worth shipping.
The schema and the route never made it into the chat.
What a pasted snippet can see
A generic chat model answers from the snippet you pasted in, not the actual codebase. It cannot see the three other files a real feature touches unless you paste those too, one at a time.
+dueDate DateTime?+{ title, dueDate }+<DueChip date={task.dueDate} />One change, three layers that must agree
A schema field, its API route, and the UI that renders it have to agree with each other. Reading and writing all three together is the part manual work gets wrong under time pressure, not any one layer alone.
“Added due dates to the task API”
A description versus the real diff
Being told what a change does is not the same as seeing it. Every edit here is a real diff against your actual files, so review means reading code, not trusting a summary.
Add an optional dueDate to Task, through the API and the card
Optional, as written: DateTime?
Route accepts and returns it
TaskCard shows it
What general will and won't guess at
general works from what you wrote, so a clear request gives you output predictable enough to hand off.
You write the schema change, then the API route, then find the UI three files later, testing each layer on its own.
Three files, edited one at a time
schema.prismaEdited
route.tsNot updated
TaskCard.tsxNot updated
Each layer written and tested on its own
One request describes the feature. ML.ai reads the schema, the API and the UI together and returns one diff for review.
Add a due date to tasks, from the schema to the card
general
general · Read · Plan · Write diff
Reads the relevant files, plans the edit across every file the change touches, then writes it as one reviewable diff.
architect
architect · Read · Plan · Write
Read-only. Returns the critical files and an ordered plan before general executes any of it, useful when the shape of the feature is not yet obvious.
Most daily work is plumbing: a new endpoint, a form, a column across three layers. ML.ai matches your patterns and holds the change behind a diff.
Explorer
general, the default agent, takes the whole task
Add a dueDate field to tasks: schema, API, and the task card UI.
Explorer
Add a dueDate field to tasks: schema, API, and the task card UI.
Read prisma/schema.prisma
Read src/api/tasks/route.ts
Read src/components/TaskCard.tsx
Planned one edit across all three
Describe a task or ask a question
Explorer
One diff across three files
Describe a task or ask a question
Explorer
One diff across three files
Allowed · schema, route and TaskCard saved in one diff
Describe a task or ask a question
Explorer
general, the default agent, takes the whole task
Add a dueDate field to tasks: schema, API, and the task card UI.
Where feature work usually breaks, and how ML.ai closes each gap.
dueDate across 3 filesWith ML.ai
schema.prismadate, optional
route.tsdate, optional
TaskCard.tsxdate, optional
A schema field renamed or retyped without updating the API contract and the frontend type that reads it, so the mismatch surfaces at runtime instead of at review time.
How handles it
general reads the schema, the route and the component together before writing, so the diff carries one consistent type across all three.
Where Task is usedWith ML.ai
route.ts+ dueDate in response
TaskCard.tsxrenders the card
tasks.test.tsfixture
A new field added to the model but never wired into the one API response that needs it, because that response lives in a file nobody thought to check.
How handles it
The agent traces how the touched files are used before finishing, so the schema change is never treated as the whole task.
One diff, every layerWith ML.ai
schema.prisma+1
route.ts+3
TaskCard.tsx+6
Each layer looks right on its own, and the full path first runs end to end in production.
How handles it
The diff spans every layer at once, so you review the complete path the request will take in one place.
The real edges of what general can do here, stated plainly instead of left for you to discover.
general won’t stop to ask whether dueDate is required. Write “optional” or “required, defaults to today” in the request, or it makes the safe choice itself.
Wants to write 5 lines
The diff waits behind Allow, Always or Deny. Pick Always if you don’t want to be asked about that kind of edit again this session.
Once a change spans more than a handful of files, plan it with architect, then hand the plan to general.
Plan · session store migration
Open question: hard cutover or dual write
architect returns the key files, trade-offs and a plan without editing anything. planSubAgent saves that plan to a file.
Root cause in session.ts
The store returns null before the session write finishes, so the middleware rejects a valid token.
Give ML.ai a failing test, an error message, or a stack trace and it reads the execution path, finds the actual cause, and proposes a fix as a diff you approve.
One pass across every file the change touches. It reads the relevant files first, plans the edit across all of them, then presents the full diff together rather than file-by-file.
Yes, that is the common case: general follows the change through schema, API, and UI in one task, the way the dueDate example above does.
Say what to change and it revises before anything is written. Nothing touches disk until you approve, so there is no partial-apply state to undo.
No. general cannot ask clarifying questions mid-task, so a vague request gets a guess, not a question. Say what should exist and roughly where, and it has what it needs.
When the change is big enough that the order of work matters, or the shape of the feature is not obvious yet. architect returns the critical files and an ordered plan before general executes any of it.
today, or talk to us about what is nextInstall the editor agent on your own machine, or book a call to talk through your team’s workloads
