A Practical Prompting Workflow for Research, Summaries, and Decisions

1. Find your starting level and choose one real task

Before we touch prompts, decide what you’re actually trying to do.

You’ll get more out of this guide if you work on a real decision or research task, not a toy example.

Use this quick self-check to pick your starting level:

Beginner

You’ve asked an AI general questions ("Explain photosynthesis"), but never used it to support a real decision, and you don’t save prompts.

Intermediate-ish

You already paste in text or links and ask for summaries. Sometimes it’s helpful, sometimes it’s vague, and you’re not sure how to fix it.

If you’re not sure, assume beginner. The workflow still works; you’ll just move faster.

Now choose one concrete task you care about this week, for example:

You want something with real stakes but low risk. Not a medical diagnosis, not a legal strategy. Think purchase decisions, learning a topic, comparing options, or drafting a summary for work.

Write your question in one sentence. We’ll use that exact sentence in the first prompt.

2. Set up a simple research sandbox

Large models like Claude Opus 4.7 or GPT-5.1 can search the web for you, but beginners often get better results by controlling the inputs themselves.

For this first run, we’ll keep it manual on purpose:

1. 1

   Open a document (Notes, Notion, Google Doc, whatever you like)

2. 2

   Find 3–6 sources about your question

   For example: a review article, a product page, one or two blog posts, maybe a forum thread.
3. 3

   Copy the most relevant sections (not entire 20-page PDFs) into your document

   Aim for 500–2,000 words total.

Label each chunk clearly:

[Source 1 - Wirecutter review]
...pasted text...

[Source 2 - Manufacturer specs]
...pasted text...

You’re building a small, controlled context that we’ll feed into the model. This is a poor person’s retrieval-augmented generation setup: you choose the context instead of a search API doing it for you.

Once that’s done, open your model of choice (ChatGPT, Claude, Gemini). We’re ready for the first attempt.

3. First attempt: a copy-paste workflow you can run today

We’ll start with a single, moderately-structured prompt. Treat this as source code, not scripture: you’ll edit it later.

Paste your gathered sources into the chat, then follow them with this prompt, replacing the parts in ALL CAPS with your own text.

You are helping me research and make a decision.

My question:
[ONE-SENTENCE QUESTION HERE]

My constraints and preferences:
- [1–3 SHORT BULLETS, e.g. “budget around $500”, “I work from home 3 days/week”, “prefer established brands”]

I will paste several sources. Rules:
- Only use information from these sources.
- If an answer would require outside info, say what’s missing instead of guessing.
- If sources disagree, call that out.

Tasks:
1) Extract a concise fact table with columns: Source, Key facts, Numbers/estimates, Caveats.
2) Write a summary (max 200 words) of what these sources say that’s relevant to my question.
3) List the 3–5 main tradeoffs or uncertainties that matter for my decision.
4) Give a tentative recommendation *for me*, based only on these sources and my constraints. Include a short "because" paragraph.

Format the answer with clear headings.

Run this once, as-is. Don’t tweak it mid-stream. The goal of the first attempt is not perfection. It’s to generate something concrete enough that you can inspect what went right and wrong.

4. What good vs poor results look like

Now you need to read the model’s answer like a skeptical collaborator, not like an oracle.

Use this quick inspection table as you read:

As you skim, mark issues directly in the answer (or in your notes):

Treat the first output as a test, not a verdict. You’re not asking “Is this correct?” in a yes/no sense; you’re asking “How is this system behaving with the instructions and evidence I gave it?” Once you see the behavior, you can change the instructions instead of blaming or trusting the model at random.

If your result already looks “good” by this checklist, skip ahead to the section on decisions. If it’s mixed or bad, keep reading; this is where the workflow pays off.

5. Tightening the prompt after a weak first attempt

When the first output is weak, beginners tend to either give up or paste a longer, louder prompt. Neither helps much.

Instead, you’ll make targeted adjustments based on what went wrong.

If the answer was too generic

Tell the model exactly what went wrong and what you want instead. For example:

Your previous answer was too generic and could have been written without reading my sources.

Retry the task, but:
- For every key claim, add in parentheses which source it came from.
- Remove any advice or statements that are not clearly supported by the sources.
- If a detail is missing from all sources, explicitly write "Not specified in the provided sources".

Then redo the fact table and 200-word summary with these rules.

If the recommendation ignored your constraints

Explicitly restate them and force the model to use them as filters:

Your previous recommendation did not respect my constraints.

Restate my constraints in a short bullet list from my perspective.
Then, using only the facts already extracted, answer in this order:
1) Which options clearly violate any constraint? List and explain.
2) Which options fit all constraints but involve tradeoffs? Describe the tradeoffs.
3) Given that, make a revised recommendation and show a 3-line "because" chain that mentions specific constraints.

If sources conflicted or the model smoothed over disagreements

Make disagreement first-class:

Identify any places where the sources disagree or emphasize different things.

Create a short section called "Source disagreements" with bullets like:
- [Source X] says ...
- [Source Y] says ...
- Practical implication for my decision: ...

Then, update the recommendation so it clearly states which side you’re leaning toward and why.

You’re not trying to invent the perfect mega-prompt. You’re debugging the prompt based on the model’s observed behavior.

6. Using the same loop for longer research

Once you’re comfortable with this 3–6 source sandbox, you can scale the pattern instead of starting over.

For bigger questions, the workflow looks like this:

1. 1

   Break the question into sub-questions

   Example: for “EU AI regulations,” you might split into definitions, timelines, and obligations for your type of organization.
2. 2
   Run the same question → gather → summarize → tradeoffs loop for each sub-question separately.
3. 3
   Then ask the model to combine the sub-summaries into a higher-level view.

When context window is a concern (check your model’s token limits in the provider docs), don’t paste 20 pages at once. Have the model help you compress:

Here are 3 long documents about [topic].

Task: For each document, create a 200–300 word "summary for future AI context" that captures key facts, definitions, and numbers.

Format:
[Doc 1 - short title]
Summary: ...

[Doc 2 - short title]
Summary: ...

You then use those compressed summaries as your “sources” in the main workflow. It’s not perfect, but it keeps you within the context window while preserving most of the useful structure.

7. Decisions: turning summaries into clear calls

Research and summaries are only useful if they lead somewhere. The last step is turning the structured information into an explicit decision.

Start by asking the model to rewrite its own reasoning in a way you can quickly sanity-check:

Based on everything above, create a decision brief with:

1) My question (1 sentence)
2) My constraints (bullet list)
3) Options being compared (2–4 items)
4) A brief argument *for* each option
5) A brief argument *against* each option
6) Your recommendation and a 3-step reasoning chain

Keep the whole brief under 400 words.

Read this brief with the same skeptical lens as before. Ask yourself:

• Does the recommendation actually follow from the arguments, or does it jump to a conclusion?
• Is there any hidden assumption that matters to you but isn’t stated?
• If you showed this brief to a colleague, would they understand why you’re leaning one way?

If the brief feels close but not quite right, that’s the time for one or two clarifying follow-ups, not a full redo. For example: “Rewrite the recommendation assuming I care more about long-term reliability than upfront cost.”

Think of the model as a fast, opinionated intern. You’re still the one signing off.

8. Common mistakes and how to avoid them

You now have a basic workflow. To keep it useful over time, avoid a few patterns that quietly ruin results.

Mistake 1: Asking the model to “do research” with no constraints.
You get generic blog-post output because you never gave it specific evidence. Fix: either use built-in browsing plus explicit instructions about what to look for, or do the small manual-gathering loop we used here.

Mistake 2: Treating a single prompt as a magic key.
Models change. A prompt that worked beautifully in 2024 might behave differently on a 2026 model. Fix: keep prompts short and legible, and rely on the inspect → adjust pattern instead of copying “ultimate prompts” from social media.

Mistake 3: Skipping the inspection step.
If you don’t read the first output critically, you can’t tell if improvements are real or just more words. Fix: keep using the fact table, source mapping, and constraint checks.

Mistake 4: Overloading the model with context.
Pasting 30 pages at once and asking for “key insights” usually leads to mush. Fix: chunk documents, summarize each, then work from those summaries.

Mistake 5: Using AI where precision stakes are too high.
For medical, legal, or financial decisions with real downside, treat this workflow as a way to prepare questions for a human expert, not to replace them.