Carving, Construction, Curation and Scaffolding: Four Workflows for Advanced Prompting
Last week, we discussed a basic, broadly-applicable prompt template to get better results with LLMs for simple tasks. But what if you have more complicated tasks?
If you’ve ever asked a chatbot to draft a grant proposal, summarize a complex report, or help prepare for a high-stakes meeting, you know the results can be uneven. Sometimes the AI feels like a brilliant collaborator; other times it leaves you with generic writing, shallow ideas, or just complete nonsense.
With tasks that are difficult for AI (which are hard to identify in advance!) you can improve your results by using multiple prompts. I think of this as three distinct approaches: Curation, Carving, and Construction.
1. Curation: Selecting and Refining Outputs
Think of curation like being a museum curator. You generate multiple outputs and then carefully select, edit, and polish the best ones.
When to use it: Brainstorming sessions, title generation, divergent idea generation.
How to use it:
Ask the model for 10 different ideas, not one. Ask for more than you need; when you start getting crappy answers, you’ll know you got all the good ideas it can offer with that prompt and are scraping the bottom of the barrel.
Push for variety (you can use the divergence-encouraging prompts we talked about in the Brainstorming post! And don’t forget to create new chats often)
Keep your critical thinking hat on: your job is to sort and shape the outputs, not accept them wholesale.
Curation works best when you want to consider a broad range of options and when the task itself has few steps.
2. Carving: Reshaping Existing Material
Carving is like sculpting from marble—you already have a block of text or ideas, and you’re refining it step by step.
(I know this image is ridiculous— I gave up after three prompts attempts. You get the idea!! Check the LLM disclosure at the bottom to see what happened here.)
When to use it: Long-form content with lots of context and arguments or themes throughout. For example, drafting a memo, report from information you already have, or rewriting existing for a different audience.
How to use it:
Enter a lot of context and ask the LLM to generate a large block of text. Ideally, bring your own sources, thesis statement, and even an outline.
Ask it to add, take away, and incorporate new information until it matches your goals.
Quick Tip: If you are using an LLM that has this feature, ask it to create a document. It will put its output into a pane next to your chat, rather than in the chat, so you can scroll through the output while you chat. If you ask it to change a particular paragraph, it will integrate those changes itself, rather then spitting it out in the chat for you to incorporate yourself.
Curation works best when you have a very good idea of what the content should be and need to shape it into a particular form.
3. Construction: Building Step by Step
Finally, construction is the architectural approach—you’re building a complex output piece by piece, ensuring each layer is solid before moving on.
When to use it: Complex, high-stakes, or interdependent tasks (e.g., grant proposals, strategic plans, performance evaluations).
How to use it:
Break the project into parts, just like you would if you were project planning or delegating.
Identify which parts a human needs to do and which parts an LLM can help support.
Supply the AI with lots of context (RFP text, partner agreements, prior reports).
Quick tip! Use a project here! It will be easier to manage context and pieces of the puzzle.
Start with a strong “foundation block” (e.g. a project description or introduction). Gather lots of context, write your prompt carefully, get feedback on it, and spend a lot of time making that block excellent.
Put that foundation block back into your project files (or the context for prompts for future sections).
Build sequentially. Making sure each block is reviewed and added to the context. This will support dependencies and reduce the need for cleaning up the project at the end.
Construct your final product and fact-check it! Look for hallucinated facts or promises.
Construction is the most time-intensive but also the a safe approach for complex and risky tasks where accuracy and coherence matter most.
4. Scaffolding: Support Without Output
Sometimes, the safest way to use AI is not to let it touch the final product at all. For especially risky, high-stakes projects, you may prefer to keep human hands on the actual writing. You may still be able to use AI to make the process more efficient and improve the quality of the product.
When to Use It
Scaffolding is ideal when:
Accuracy, tone, or compliance requirements leave no margin for error.
You’re working with sensitive or confidential material.
You want to protect human authorship while still benefiting from AI’s pattern-spotting abilities.
How It Works
Instead of asking the AI to create anything that will go into the ultimate output, you lean on it to do support tasks.
Examples include:
“Here is my list of sources. What am I missing?”
“Here is my report. Can you identify places where I’ve split infinitives?”
“Please flag any claims that should be cited.”
“Which sections rely too heavily on one source?”
Scaffolding preserves human ownership and accountability for the final product. It’s a way of using AI that maximizes safety and control, especially when reputational risk, compliance, or funding is on the line.
Why This Matters
Single prompts can write a mean e-mail, but you often need a prompt workflow to get the most out of AI when you have complex, sensitive, and values-laden tasks. By choosing the right prompting workflow, you can turn AI into a more reliable partner while keeping human judgment firmly in the driver’s seat.
Have you tried any of these approaches in your own AI work? Do you lean more toward curation, carving, construction, or scaffolding? (Do you have any ideas for alternatives to “scaffolding” that start with a “C”??)
__
LLM disclosure:
I asked ChatGPT 5 “Can you write a blog post referring to the [redacted] slides about Carving, Construction, and Curation approaches to multi-shot prompting?”
This output was quite good, much to my surprise. I did edit it with more details, especially in the intro, but the structure was pretty spot on.
“Can you create coordinating 1:1 images, one representing each of curation, construction, and carving?”
(No, it cannot. It created one very generic image of a computer screen. NEXT)
“Can you create a 1:1 image of a hand selecting one of several blocks from a pile? the block that is being selected should be #00b4c4 and the rest should be light grey.”
Ugh. This looks JUST like the image i made for my slides. FOR CONSTRUCTION. (The pile looked like a pyramid.) I am going to switch chats and try something else.
“Can you create a 1:1 image of a hand selecting one of several blocks from a pile? the block that is being selected should be #00b4c4 and the rest should be light grey. The pile should look like the image from the [redacted] deck for curation.”
This worked!
Thank you! Can you make one that complements this image aesthetically and depects a large block (the same color blue) about to be carved with a chisel?
Love this, but can you create a gash in the block where the curl could have come from? Right now it's just resting on top of the block, not coming from it.
ok can you remove the curl and make it look as if the chisel is being pusehd down rather than pulled?
This seemed like it was not going any where, so I actually just went with the first one because it failed in a funny way, instead of an annoying one.
Can you create one that is similar to the construction image from the slides? Some should be in the blue and one or two in grey.
Then, at a later date:
I'd like to add a section to the multi-shot post about scaffolding. Where projects are very risky, you might not want AI to touch any of the output at all. In this case, the AI offers support for the human who is building the content themselves. Prompts like, "Here is my list of sources. What am I missing?" Or "here is my report. Can you identify places where split infinitives/make claims that should be cited/refer to a particular source?" Can you draft this section of the post for me?