Four-Part Direct Prompt Checklist
Give reasoning models clear requests, boundaries, and only relevant context.
- Difficulty
- Starter
- Time to result
- ~days to results
- Steps
- 5
- Confidence
- 96%
This checklist treats a prompt as a compact work brief rather than an elaborate incantation. First, state the ask in simple, direct language. Second, avoid forcing a modern reasoning model to display or follow a user-specified chain of reasoning when it can reason internally. Third, use delimiters to distinguish instructions, source material, examples, and output requirements. Fourth, supply only the context that materially helps with the task. These constraints improve signal-to-noise ratio and make the prompt easier for both humans and models to inspect. The checklist is a starting configuration, not a substitute for iteration: users should review the response and clarify the brief when the output reveals ambiguity.
Origin
Extracted from Marketing Against The Grain during a discussion of OpenAI's guidance for prompting its o1-preview reasoning model.
Core principles
- 01Treat prompting like clear communication with a colleague.
- 02State the request simply and directly.
- 03Let reasoning models handle their internal reasoning.
- 04Separate prompt sections with explicit boundaries.
- 05Exclude context that does not help complete the task.
How to run it
- 1
Write the Direct Ask
Describe the desired task and output plainly. Remove unnecessary scene-setting and indirect language.
Pro tip Write as if briefing a capable colleague who is unfamiliar with this specific assignment.
Watch out A simple prompt still needs a concrete objective.
- 2
Remove Forced Reasoning
Delete instructions that ask the model to expose or mechanically follow a long chain of reasoning. Focus the prompt on the result, constraints, and evidence required.
Watch out Do not confuse removing chain-of-reasoning requests with removing necessary process constraints.
- 3
Add Delimiters
Mark the boundaries between instructions, context, examples, and source data. Use consistent labels or structured sections.
Pro tip Choose a delimiter format that makes the prompt easy to scan and reuse.
Watch out Inconsistent boundaries can make source text look like an instruction.
- 4
Trim the Context
Retain only information that affects the model's decisions or output. Remove background material that is merely interesting.
Pro tip Ask whether the answer could change if each context item were removed.
Watch out Trimming must not remove facts needed for accuracy or identity verification.
- 5
Iterate from the Output
Inspect where the response diverges from the goal and revise the brief accordingly. Repeat until the prompt performs reliably.
Pro tip Change one important variable at a time when diagnosing weak output.
Watch out Do not mistake a polished first response for a robust reusable prompt.
In the wild
A host uploads two LinkedIn profiles and a sample of an Actors on Actors exchange, then directly asks the model to create questions for a comparable conversation. The documents are separated from the task instructions and provide only relevant context.
→ The model produces a strong initial interview-question set that can be refined conversationally.
Common mistakes
Overengineering the First Prompt
Spending hours crafting a supposedly perfect prompt delays the faster learning available from an initial attempt and iteration.
Adding Irrelevant Context
Excess background can distract the model from the actual decision and weaken instruction priority.
Forcing Visible Reasoning
Prescribing step-by-step chain-of-reasoning can conflict with how newer reasoning models are designed to process requests.
Is it for you?
Best for
People writing prompts for modern reasoning models that can perform complex reasoning internally.
Not ideal for
Tasks whose primary difficulty is missing domain knowledge or the absence of examples showing what good output looks like.
From the transcript
“The first is be simple and direct in your asks. Be very simple and direct in your asks.”
“The third one is use delimitators. So delimitors are how do I segment my section?”
“And the fourth is basically only give it the context it needs.”
From the episode
From Beginner to AI Expert in 30 Minutes: The 5-Step Framework You Need