Plans, Task Execution, and Verification

Introduction: ask for a plan before the first broad edit

When the task spans several files or editors, start by asking the AI Manager to plan the work. Use one concrete outcome. For example, ask it to add a feedback feature with a public form, a save action, and an admin review page. A good plan names the slices, the dependencies between them, and the checks you will use after each pass.

Outcome

State what the whole task
must accomplish

Slices

Break the work into
reviewable passes

Dependencies

Call out page, data,
backend, or docs links

Checks

Name how each slice will be
validated or reviewed

Ask the AI Manager to outline the work before it starts editing broadly.

Use the opening plan to define slices, dependencies, and checks.

Treat planning as part of execution quality, not as overhead.

Text

Planning prompt
"Plan this feedback feature before making broad edits. I need a Bootstrap 5 public feedback page, a backend save action with validation, and an admin review page. Break the work into the smallest useful slices, list the dependencies between them, and tell me what I should review after each pass in Wappler."

Screenshot: Introduction: ask for a plan before the first broad editClick to open the larger image

Turn checkpoints and validation into part of the plan

The strongest AI plans include how progress will be checked, not just what will be built. For the feedback feature, one checkpoint might be Design View and Structure after the page pass, another might be response-shape review after the Server Connect pass, and another might be Problems panel or test review after the final wiring pass.

Editor review

Design View, Structure,
Properties, or Code

Backend review

Server Connect contract,
validation, output shape

Safety checks

Problems panel, tests,
or tour checks when needed

Ask which validations should run after each slice instead of deferring everything to the end.

Keep restore points and touched-file review in mind when a pass may need rollback.

Use stop points where you can inspect UI, data flow, or generated structure before widening the scope.

Text

Checkpoint prompt
"For each slice in the feedback feature plan, tell me the validation checkpoint before we continue: which Wappler surface I should inspect, which output or state should exist, and whether any narrow check should run before the next pass."

Screenshot: Turn checkpoints and validation into part of the planClick to open the larger image

Execute in bounded passes

After the plan is sound, let the AI Manager execute one bounded pass at a time. For example, pass one might create the public feedback page only. Pass two adds the save action and response contract. Pass three wires the submit states and the admin review view. That makes corrections cheaper because each prompt only adjusts the current slice.

One pass

Ask for the next bounded
execution step

Review

Inspect the result before
expanding the scope

Correct

Use focused feedback to
adjust the current slice

Continue

Carry the same plan into
the next verified pass

Execute one bounded pass at a time so feedback stays specific.

Review each pass before the next one expands the scope.

Keep the plan alive as the thread that ties execution and validation together.

Text

Execution prompt
"Execute only slice one from the plan: build the public feedback page structure and form fields, keep Bootstrap 5, and do not create the backend action yet. After the pass, list the files or surfaces changed and the checkpoints I should review before approving slice two."

Screenshot: Execute in bounded passesClick to open the larger image

Recover from drift without throwing away the whole task

If the task drifts, use the plan and the last good checkpoint to recover. Do not reset the whole conversation unless the job has materially changed. First identify what is still correct, restore if needed, then ask the AI Manager for the next bounded correction based on the last verified state.

Reference point

Use the last verified slice
as the recovery anchor

Restore if needed

Roll back when the current
state is materially worse

Correct narrowly

Ask for the next focused fix,
not a whole restart

Use the last good verified point as the reference for the next correction.

Restore when the current state is materially worse, not just because one detail is unfinished.

Keep the same overall outcome in view so recovery does not turn into a new unrelated task.

Text

Recovery prompt
"Use the last approved checkpoint from the feedback feature plan as the reference. Restore if the current pass moved too far away, keep the existing page structure, and fix only the validation response handling. After that, tell me what to verify before we continue."

Screenshot: Recover from drift without throwing away the whole taskClick to open the larger image

Continue into broader AI workflows

Return to the prompting hub, or move back into the AI Manager and advanced routes for larger agentic work.

Choose the next execution-focused tour

Continue with the related tours that help you prompt better, enforce standards, or return to the wider AI Manager workflow.

Screenshot: Choose the next execution-focused tourClick to open the larger image