Align Before You Execute
Work backward, align with the model, then execute the task
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 5
- Confidence
- 98%
This workflow divides AI-assisted work into two stages: alignment and execution. Begin with the intended outcome, then work backward with the model to define the approach, assumptions, constraints, and sequence of work. Instead of immediately requesting a finished post, product position, specification, or application, ask the model to propose how the task should be completed. Refine that proposal through dialogue until both sides share the same definition of success. Only then move into production. The mechanism reduces rework because structural mistakes are found while they are still inexpensive to correct. It also turns prompting from a one-way instruction into a collaborative planning process that can be reused across creative, strategic, and technical work.
Origin
Kieran Flanagan described this as his habitual back-to-front workflow while preparing a creator-hub build on Marketing Against The Grain.
Core principles
- 01Start from the desired outcome and work backward
- 02Separate alignment from execution
- 03Use dialogue to expose assumptions before generating deliverables
- 04Treat prompting as an iterative process rather than a single command
How to run it
- 1
Define the destination
State the result you want, the intended user, and the practical outcome the work should create. Focus on the destination before prescribing implementation details.
Pro tip Describe what a successful result lets the user do.
Watch out Do not begin with a vague request such as simply asking the model to make something good.
- 2
Ask for an approach
Have the model propose the structure, logic, phases, or decision process it would use. This makes its assumptions visible before it commits to a deliverable.
Pro tip Ask explicitly whether important requirements or edge cases are missing.
Watch out A confident proposal can still reflect an incorrect interpretation of the task.
- 3
Iterate toward alignment
Challenge unclear choices, add constraints, and correct misunderstandings through back-and-forth discussion. Continue until the proposed method matches the intended outcome.
Pro tip Resolve one major ambiguity at a time.
Watch out Do not mistake a long or polished response for genuine alignment.
- 4
Confirm the sequence
Ask the model to restate the agreed process as sequential steps or phases. Use that restatement as the execution plan.
Pro tip Check that each step has a clear input and output.
Watch out Watch for steps that hide unresolved technical or business decisions.
- 5
Execute the agreed task
Request the deliverable using the aligned plan as context. Review the result against the previously agreed definition of success.
Pro tip Preserve the alignment conversation in the same context window.
Watch out Do not silently change the objective during execution.
In the wild
Before building the creator tool, Kieran described the users, research stages, platform ranking, and creator-discovery requirements. He then asked Claude to turn the conversation into a functional specification and sequential build plan rather than immediately requesting code.
→ The model returned a structured specification that could be adapted for Perplexity Computer.
A marketer asks an AI to identify the audience, competitive frame, evidence requirements, and positioning structure first. After correcting the model's assumptions, the marketer requests the finished positioning document.
→ The final document follows an agreed strategic mechanism rather than an arbitrary template.
Common mistakes
Jumping directly to the deliverable
Immediate execution conceals disagreements about scope, assumptions, and success criteria until after substantial work has been produced.
Aligning without a concrete outcome
An extended conversation does not help if the user never specifies what the finished work must accomplish.
Treating the first plan as final
The first proposed approach should be inspected and corrected rather than accepted merely because it sounds authoritative.
Is it for you?
Best for
People using AI for complex writing, positioning, planning, or product-building tasks with multiple possible approaches.
Not ideal for
Simple factual or mechanical tasks where the desired output and method are already unambiguous.
From the transcript
“Yeah, I work back to front.”
“I think one of the best things you can do when working with these models is to first go back and forth until you align…”
“A lot of people just jump in and start to do the thing. But actually, it's really good if you go back and forth and…”
From the episode
Claude Broke. Perplexity Built the App Anyway