Have AI List the Items Before Building a Large Section── Make listing the items the first stage and agree the scope before design, and gaps and extras get fixed before anything is built

01Before the new section was built, the operator had AI list its items

In September 2026 the operator set out to add a new section to their site, one that would collect practical know-how on using AI. The planned contents included tips for each office application, such as word processing and spreadsheets, choosing between AI models, building and running AI agents, and linking AI to documents kept in cloud storage.

The operator did not hand this over as a single instruction to build the section. The request split the work into four stages and fixed their order. First, list the items the section would cover. Second, draw up a design. Third, write the program that would do the work. Only last, create the pages.

The request named several items the operator already had in mind. It did not claim that they were the full set. Producing the complete list, including what was missing, was the first job given to the AI.

The point worth checking is whether this order, list first, is a pattern anyone can use when asking AI to build something large.

02An item list is a draft that settles the scope with the AI before building

Ask an AI to build a large section in one go and it decides the scope for itself as it works. The person who asked learns that scope only when the pages are finished.

Asking for the list first reverses the order. The AI proposes the items the section will cover. The person reads the proposal, adds what is missing and strikes what is not needed. Design and building start only once there is an agreed list.

03Scope errors grow more expensive the later they are found

A mistake in scope costs more the further along the work it is discovered. A 2004 report by NASA engineers put numbers on this. Taking the cost of fixing a requirements error during the requirements phase as 1, they estimated 3 to 8 during design and 21 to 78 during integration and test.

Those figures come from large space projects, which are nothing like asking an AI to build web pages. The ordering still holds. Whatever you leave undecided has to be fixed later along with everything built on top of it.

The AI adds its own source of drift. Anthropic's published best practices for Claude Code note that Claude tends to read a request conservatively and build the smallest thing that satisfies it, because it cannot see how much the person was hoping for.

In a large section, that unseen scope is wide. Asking for the list first means the scope is settled on paper instead of being left to the AI's guess.

04The list here means a table of things to build and things to leave out

An item list is neither a table of contents nor a design. A table of contents orders material for the reader. A design records how something will be built. The list comes before both and settles only one matter: the set of things to be built.

The fourth column is the one people forget. Items the AI added may drift from the person's intent. When origins are marked, the person knows which lines need the closest reading.

Some work does not need this step. A fix to a single page, or a job whose scope is already settled, gains nothing from it. If the items number a handful and fit in your head, a list only adds effort. Use it when the output spans several pages or sections and you cannot state the whole of it yourself.

05In material review, it pays off before building a guide or a set of FAQs

In pharmaceutical companies, people in promotional material review and medical affairs increasingly ask AI to produce large bodies of work. Examples include an internal guide to review criteria, a set of answers to frequent questions, and training material for new staff.

The value of such work depends on its scope. In a review guide, the choice of criteria makes up most of the content: claims about efficacy, the presentation of comparisons, the handling of safety information, the rules for citing studies. A criterion missing from the guide will also go missing from the reviews of anyone who relies on it.

Asking for the list first lets the team discuss missing criteria before anything is written. A reviewer compares the AI's proposed criteria with company rules and past findings, then adds and removes. Kept as a list, the result also gives the next person who revises the guide a place to start.

Website teams have long worked this way. Nielsen Norman Group recommends a content inventory, a page-by-page list of existing content, together with an audit of its quality before a redesign. The purpose is to decide what to keep and what to drop before rebuilding.

06A list splits the work into stages, so each AI step is easier and can be checked

There are two reasons the list works. The first is that only the person knows the intended scope. An AI can propose many candidate items quickly. Choosing among them takes someone who knows the goal and the readers. The list is the point where the AI's proposing is separated from the person's deciding.

The second is that splitting the work makes each AI step smaller. Anthropic's engineering article on building agents describes prompt chaining, in which a task is broken into steps and each step works on the output of the one before. A check between steps, which the article calls a gate, shows whether the process is still on track. Its example is to write an outline, check it against criteria, then write the document from the outline.

The operator's four-stage request has that shape. List, design, program and pages run in order, each stage feeding the next. The list comes first because the other three all depend on it.

07From tomorrow, give seed items, get the list back as a table, and agree before design

There are four steps.

1. Give seed items. Write down the items you already have in mind and say the list is not complete. State that nothing should be built yet.
2. Get the list back as a table. Ask for four columns: item name, one-line content, left out, and origin.
3. Add, strike and reorder. Read the list and settle the scope. Mark uncertain items as deferred instead of deleting them.
4. Pass the agreed list to the next stage. Have the design drawn from this list. When the work is done, have the AI check that every item on the list made it in.

The same best practices suggest starting the build in a fresh session once a specification exists, and reviewing finished work against that specification. An agreed item list is the smallest form such a specification can take.