Turn messy ideas into useful AI work.

I start messy. The work does not have to stay messy.

My brief for DLVX marketing was a lot. I wanted stronger proof on Instagram, Bangkok campaigns, useful AI education, a mailing list, better search visibility and a website that made the company feel established. I also wanted bigger, longer-term partnerships without abandoning the smaller projects that can become those relationships.

Those ideas belonged together in my head. They were not yet a brief someone could execute. I was mixing an ambition, possible channels, things that annoyed me about the current work and instructions to start publishing. If an AI took every sentence as an equally important task, I would get a busy marketing department before I had a clear marketing decision.

This is where I want AI to help. I do not want to spend an hour making my thoughts presentable before I can ask for help. I want to give it the rough material and get back a better working brief, with the uncomfortable choices still visible. The method below works in a ChatGPT or Claude conversation. It does not depend on a magic phrase, a particular model version or pretending the assistant knows my business already.

By the end, you should have one decision, a compact context packet, an executable task and a way to reject a polished but weak result. Keep the worksheet at the end open while you try it on your own problem.

Ask for the brief before you ask for the strategy.

Here is a cleaned-up paraphrase of part of my actual request: “Make our Instagram show proof and AI expertise. Research Bangkok and create four or five creative directions. Think about ads for ten to twenty services. I want big companies and real partnerships, but keep the easier website projects too.”

There is enough direction there to start. There is also enough ambiguity to waste a week. Does “big companies” mean an enterprise procurement team, a local company with meaningful revenue, or an agency looking for a delivery partner? Is the immediate outcome followers, a qualified enquiry, or evidence a salesperson can send after a meeting?

I would put this prompt below the rough brief:

Turn the material above into a working brief before proposing campaigns.

Identify:
1. The business outcome I actually want.
2. The first audience and buying situation worth addressing.
3. Constraints I explicitly stated.
4. Ideas I mentioned that are possible methods, not requirements.
5. Assumptions you would need to make.
6. Decisions that would materially change the next piece of work.

Recommend one first deliverable. Explain its strongest competing option.
Do not turn my ambition into an achieved company claim.
Ask only questions that block useful progress. Continue with work that
can be done without those answers.

Return a brief I can approve or correct in two minutes.

The important move is separating my desired outcome from my suggested methods. “We need a newsletter” may really mean “People should have a reason to keep hearing from us.” A newsletter is one answer. A useful recurring lesson, a workshop or an existing customer update could be another. I want that distinction before the assistant designs an email template.

Anthropic’s guidance recommends clear instructions, relevant context and examples that show the intended result. I use those principles here; the workflow and judgment calls are my recommendations, not an official recipe. Claude prompting guidance.

The better brief makes a choice.

For that marketing request, an executable first brief could look like this. It is an illustrative reconstruction, not a report that a campaign achieved these results.

Decision: Which first campaign should earn a qualified conversation
with an established Bangkok business?

Audience: A business owner or operations lead with an existing enquiry
flow and a visible handoff problem.

Outcome: A relevant conversation about connecting the website,
enquiries and follow-up. Followers are a supporting signal.

First deliverable: One campaign concept, two creative treatments,
and a destination brief for the same service and audience.

Proof: Use documented DLVX project scope. Identify partner delivery.
Do not invent revenue gains, deployed AI features or client quotes.

Constraint: Keep the wider brand credible for long-term partnerships.
A website project can be the first engagement without defining
the whole company as a cheap website supplier.

Unknowns: Channel demand, commercial readiness and paid-media budget.
Research the first two. A proposed budget is not spending authority.

Acceptance: A buyer can identify the problem, see relevant evidence,
and understand the next step. Each claim has a source or is removed.

Notice what is absent: twenty simultaneous service campaigns. That idea is still available. It is no longer allowed to consume the first deliverable just because I mentioned it enthusiastically.

I would reject a brief that simply turns my message into headings labelled SEO, social, ads and academy. That is organisation, but it has not made the decision easier. The better output tells me what to do first and what evidence could change that recommendation.

Build a small context packet you can carry between tools.

A long chat is a poor place to hide the only current version of your business. Keep a short packet outside the conversation, in a document your team can maintain. It should be understandable to someone who was not in yesterday’s meeting.

My packet would have five parts: the company and audience, the current decision, approved evidence, operating constraints and the last accepted state. Link to the detailed source rather than pasting an entire history into every task. If a source is old, say when it was last checked.

This classification prevents a subtle failure: a confident instruction becoming a fact in the next draft. “We can beat other agencies” expresses my belief and ambition. It is not independent evidence for publishing “Bangkok’s number one AI agency”.

ChatGPT Projects can keep related chats, files and project instructions together. That is useful for an ongoing subject, but I still want a deliberately maintained source of truth. Having an old file available is not the same as knowing it remains correct. Check the current product instructions before relying on a particular memory or sharing behaviour. Projects in ChatGPT.

Claude Projects can also hold project knowledge and instructions. I would put stable reference material there and keep a short accepted brief in each task conversation. Do not assume a separate chat has inherited every earlier discussion. Claude: creating and managing projects.

In either tool, give different outcomes their own conversations: one for the campaign decision, another for reviewing the actual creative. Carry the accepted brief between them. A ChatGPT project is not a direct connection to a folder on my computer, and a Claude project is not automatically the source repository for my website. The tool may change; approved facts and decisions should remain portable.

Give research a question it could answer against you.

“Research the best opportunities” is an invitation to bring me whatever looks interesting. For this marketing decision, I want a narrower question: what evidence supports targeting enquiry handoff problems first, and what evidence would make website redesign a better first offer?

Ask for sources that can actually establish the claim. Official platform documentation can explain how a feature works. It cannot prove that Bangkok business owners want our service. A competitor’s pricing page can show its offer. It cannot prove customers buy it. A simulated buyer response is useful criticism, not customer research.

Research the decision in the brief, not the whole industry.

For each useful finding provide:
- The claim, source link and date checked.
- What the source directly supports.
- Your inference, clearly separated.
- How it changes the recommended first test.

Look for evidence against the preferred option as well.
Do not invent market size or treat search results as purchase intent.
Finish when more research would not change the next deliverable.
List any consequential question still unanswered.

I would rather have three sources that change the work than thirty links decorating a recommendation. If the evidence is thin, the next step can be a small test with an explicit hypothesis. The assistant does not need to manufacture certainty to earn its place.

Make the critic responsible for finding a real defect.

The first draft should now be specific enough to disagree with. Give a separate review pass the brief, the draft and the same evidence. Ask it to assess the result, not rewrite everything in its preferred tone.

Review this deliverable as a skeptical buyer and delivery lead.

Find material defects only:
- Claims the evidence does not support.
- A mismatch between audience, promise and next step.
- Work the team could not deliver under this scope.
- A missing assumption that could reverse the recommendation.

For each defect, quote the relevant line, explain the consequence,
and propose the smallest correction. If a section is good enough,
say so. Do not invent a defect to fill a quota.

End with: accept, revise, or stop, and the evidence for that judgment.

A second model can be helpful, but different branding on the chat window does not make it independent evidence. Both assistants can agree on the same unsupported claim. The source still matters.

For example, if the campaign draft describes an existing commerce project as proof of an AI customer-service agent, the critic should ask for evidence that the agent was deployed. Removing that phrase is a better correction than adding a vague disclaimer underneath an impressive claim.

I want the reviewer to distinguish a defect from taste. “The promise names a feature we have not verified” is material. “I prefer a different adjective” usually is not. Otherwise the review becomes another machine for generating work.

Keep the accepted state, then stop the loop.

Save the brief, the evidence and the accepted deliverable together. Record what changed and why. A practical version note might say: “Version two: audience narrowed to operations-led buyers; removed unsupported automation claim; paid launch remains dependent on budget and account readiness.”

When you resume in another conversation, give it that accepted state first. Ask it to identify any missing information before changing the direction. This avoids a fresh assistant enthusiastically rebuilding the strategy from the oldest attachment.

My stopping rule would be simple: the deliverable answers the chosen decision, material claims are supported, the next action is clear and the remaining uncertainty does not prevent that action. A draft can pass while the wider commercial hypothesis remains unproven. More rewrites do not turn a hypothesis into a customer.

If two review rounds keep finding the same problem, change the input or scope. Do not ask for “an even better version” indefinitely. If the source is missing, get the source. If the audience is too broad, choose an audience. If the promise cannot be delivered, change the promise.

Try it on a real decision.

Exercise one: separate direction from evidence.

Take a messy paragraph from your own notes. Mark each sentence as a fact, direction, hypothesis, unknown or aspiration. Then ask ChatGPT or Claude to classify it independently.

Pass condition: every commercial claim has a source or is explicitly unverified. Where your classifications disagree, resolve the difference from evidence. A sentence sounding confident is not a pass.

Exercise two: cut the first deliverable in half.

Ask for a working brief, then remove half the proposed output while preserving the business decision. For the marketing example, one audience, one offer and two creative treatments are enough to prepare a coherent test.

Pass condition: the smaller deliverable still lets someone decide or act. If you only shortened the document while leaving ten priorities, try again.

Exercise three: plant one bad claim.

In a private practice copy, add a clearly invented result such as “this campaign doubled client revenue”. Give the critic the real source packet and the altered draft. Never publish the practice claim.

Pass condition: it identifies the unsupported statement, explains why the available evidence does not establish it and removes it without inventing a replacement result. This tests your review process more usefully than asking whether it is rigorous.

The worksheet I would keep beside the chat.

Copy this into a document and save it with the accepted output. Complete the short fields yourself; let the assistant draft the longer parts and show you where it needs evidence.

SESSION WORKSHEET
Decision to make:
Person who will use the result:
Business outcome:
First deliverable:

CURRENT CONTEXT
Verified facts and source links:
Founder instructions:
Hypotheses to test:
Unknowns that matter:
Actions authorised now:
Actions that need a later decision:

ACCEPTANCE
What must be true for this to be useful?
Which claims need source receipts?
What would make us reject the recommendation?

HANDOVER
Accepted version and date:
Material changes since the previous version:
Evidence checked and remaining gaps:
Next action, owner and trigger:
Stop condition:

The reward is not a beautiful prompt. It is being able to hand the work to another person or another tool without losing what was decided. That is when AI starts helping me direct the business rather than giving me another pile of ideas to manage.

Take the templates into your next session.

Download the prompts and worksheets as a plain text file. Adapt them to one real job, then use the exercises above to judge the result.

The templates are free. Copying and downloading happen in your browser; these controls do not upload your business information.