Speak-Spec-Review Loop
Dictate a rich specification, generate a draft, then narrate each review cycle.
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 6
- Confidence
- 99%
The Speak-Spec-Review Loop treats prompting as a conversation with a capable collaborator rather than as formal command writing. The user dictates a rich brain dump covering the objective, context, functionality, audience, style, and constraints. An AI system converts that material into a draft, prototype, or better-structured prompt. While inspecting the result, the user narrates reactions in real time and immediately submits those observations as the next revision request. This avoids the delay of silently reviewing, deliberating, and then typing a polished critique. The loop improves both speed and output quality because speaking lowers the cost of supplying details, while rapid feedback permits more iterations within the same amount of time.
Origin
The episode demonstrates the loop by dictating a reminders-app specification and then building and revising a personalized-dictation landing page in Claude. Extracted from Marketing Against The Grain.
Core principles
- 01Explain the task as though speaking to a smart collaborator.
- 02Provide context, constraints, desired behavior, and aesthetic direction together.
- 03Let the AI turn an informal brain dump into a usable specification.
- 04Narrate feedback while reviewing instead of composing a separate written critique.
- 05Use faster input to complete more iteration cycles.
How to run it
- 1
Assign the collaborator role
Tell the AI what kind of expert it should emulate and what broad job it must perform.
Pro tip Use a role tied to the output, such as an expert website-copy writer.
Watch out A vague role cannot compensate for missing task context.
- 2
Dictate the full specification
Speak the objective, background, required elements, intended audience, design direction, and constraints in one natural explanation.
Pro tip Include the details you would give a newly hired specialist.
Watch out Do not stop merely because the spoken prompt feels longer than something you would type.
- 3
Generate an initial version
Submit the dictated material and let the AI create a concrete first draft or prototype.
Pro tip Treat the first output as V1 rather than expecting final quality.
Watch out Over-editing the prompt before seeing any output delays useful feedback.
- 4
Review aloud
Inspect the result and speak observations as they occur, including what works, what is missing, and what should change.
Pro tip Provide the voiceover to your review in real time.
Watch out Generic feedback such as “make it better” gives the AI little direction.
- 5
Request targeted revisions
Turn the narrated review into explicit changes involving content, design, security, tone, or behavior.
Pro tip Preserve good elements while naming the exact deficiencies to address.
Watch out Changing every dimension simultaneously can make it difficult to judge improvement.
- 6
Repeat to sufficiency
Cycle through generation and spoken review until the artifact is useful enough for its intended next step.
Pro tip Optimize for faster cycles rather than a perfect first attempt.
Watch out Speed does not remove the need for factual, brand, or security review.
In the wild
Alan asks Claude to act as an expert website-copy writer, explains Willow and its personalization feature, and receives a first landing page. He then dictates that privacy and security should be emphasized and that the page should look more modern and sleek. Claude produces a visibly stronger second version with the requested security treatment.
→ A V1 and materially improved V2 landing page are produced in roughly five minutes.
Alan dictates the desired reminders workflow, celebration behavior, and green visual identity rather than typing a formal product specification. The resulting text supplies enough direction for Claude to begin building the app.
→ A detailed prototype request is created in seconds with less prompting friction.
Common mistakes
Stopping the prompt too early
Typing fatigue often causes users to omit instructions they already have in mind. Continue speaking until the important context, behavior, and constraints are captured.
Demanding perfection from V1
The mechanism gains its power from cheap iteration. Treating the first generation as final wastes the speed advantage of narrated review cycles.
Giving ungrounded aesthetic feedback
Terms such as modern or sleek help, but they become stronger when paired with concrete content, layout, or brand requirements.
Is it for you?
Best for
It is best for founders, marketers, engineers, and creators who can explain what they want more easily than they can type it.
Not ideal for
It is not ideal for tiny deterministic commands or tasks requiring exact syntax that is easier to paste or type.
From the transcript
“when you're just talking you could just like just sit there and talk for a minute or two and get much better details and prompt.”
“The hack here is do exactly what you just did and then have it turn it into the right prompt for you.”
“literally all you did was just provide the voiceover to your review in real time. And that allowed us to save a ton of time.”
From the episode
Everyone’s Using AI Wrong – This Is the Real Unlock