Stop Memorizing Prompts: The Complete Five-Step Method for Delegating Work to AI

For many people, the first thing they do when learning AI is go around collecting prompts.

“100 all-purpose prompts,” “prompts every boss must have,” “copy this line and AI instantly becomes an expert”… they end up with a huge collection, and when it’s actually time to work, they still don’t know which one to copy.

The problem isn’t too few prompts — it’s treating prompts as magic spells.

The same sentence, moved to a different company, a different data set, a different customer, can produce a completely different result. A prompt that works well for someone else may be useless for you. That’s because the AI has no idea what you sell, where your materials are, what your company’s rules are, or what counts as an acceptable result.

What distributors really need to learn isn’t how to memorize sentences — it’s how to fully brief a piece of work.

This piece gives you a five-step method you can use for the long haul:

Goal, materials, rules, deliverable, acceptance.

As long as these five things are spelled out, you can apply the same method to writing a notice, reviewing store visits, reading sales, or checking inventory. The tools change; the method doesn’t have to.

I. First, a Typical Failure: Why “Help Me Analyze This” Doesn’t Work

Suppose you send a sales spreadsheet to AI and just say:

Help me analyze this month’s business.

The AI will very likely tell you:

Pay attention to sales trends;
Optimize product mix;
Strengthen customer relationships;
Control inventory risk;
Improve team efficiency.

None of this is wrong, but not a single line gives an employee something to act on today.

Why?

Because “analyze the business” isn’t a task that can be checked off. You never said what problem you wanted solved, never explained which fields are in the spreadsheet, never told it how to handle returns and free gifts, and never specified who the result is for.

Even the smartest employee, handed a task like this, can only guess.

So before you delegate work to AI, first turn “I want it to help” into “here is exactly what I want it to deliver.”

II. Step One: Goal — Say Clearly What Problem You’re Actually Solving

A goal is not a broad direction — it’s the specific question this piece of work needs to answer.

“Analyze sales” is too broad.

It can be rewritten as:

Identify the flagship products whose sales dropped more than 20% this month;
Identify Class-A customers who haven’t placed an order in 30 consecutive days;
Identify the ten products with the highest inventory value and the slowest turnover;
Compare three promotion plans and determine which yields higher gross margin at the same cost;
Turn today’s sales reps’ visit notes into tomorrow’s follow-up list.

A good goal has three traits:

A clear subject: products, customers, stores, sales reps, or orders;
A clear time frame: today, this week, this month, or the past 90 days;
A clear problem: a decline, a backlog, an overdue payment, a stockout, or something that needs follow-up.

If a single sentence mentions “sales, inventory, receivables, the team, and next year’s growth” all at once, that’s a sign the goal hasn’t been broken down yet. Letting AI solve one core problem at a time is usually more valuable than throwing ten questions at it in one go.

III. Step Two: Materials — Without Your Own Data, AI Can Only Offer Common Sense

AI knows a great deal of general knowledge, but it doesn’t know your company’s business.

To get an answer that’s on target, you have to tell it what it can rely on.

Materials generally fall into three categories.

Category One: Factual Data

For example, sales detail, inventory tables, receivables tables, customer lists, visit records. Data should come with the time period, units, and field definitions spelled out.

Don’t just drop a spreadsheet. At minimum, add a line like:

This table is store-by-SKU-by-day sales detail. Sales quantity is a positive number; the return quantity is a separate field called “Return Qty.” The amount is in RMB. The reporting period is July 1 to July 26.

These few sentences keep AI from mistaking returns for sales, confusing cases with individual units, or mixing tax-inclusive amounts with tax-exclusive ones.

Category Two: Business Context

The same drop in sales can have entirely different causes.

The end of peak season, a manufacturer running out of stock, a sales rep leaving, rain in a region, a customer closing their store — any of these could be behind the change. AI can’t see this context, so it easily mistakes normal fluctuation for an anomaly.

So supply what’s already known:

The warehouse was out of stock from July 10 to 15; the sales rep for Region B left on July 12; there was no price adjustment this month.

Category Three: Company Rules

For example:

Class-A customers are only flagged after 15 days without an order;
“Near expiry” means less than 90 days remain before the shelf-life end date;
Any single transaction over RMB 20,000 must be confirmed by the boss;
Customer names must not appear in externally facing materials;
Free gifts don’t count as sales revenue but do count as promotional cost.

If these rules aren’t told to AI, it can only fall back on general understanding.

IV. Step Three: Rules — Tell It How to Calculate, How to Judge, and What Not to Do

Many people supply the data but not the rules, and the result is still unreliable.

Rules aren’t technical parameters — they’re your company’s way of doing things.

For example, to identify “dormant customers,” you need to spell out at least:

How long without an order counts as dormant;
Whether new customers are excluded;
Whether one-time bulk-purchase customers are excluded;
Whether a return counts as an order;
Whether the judgment is made at the overall customer level or by product category.

To calculate product profit, you also need to spell out:

Whether procurement cost is tax-inclusive or tax-exclusive;
How freight is allocated;
Whether rebates are counted once received or as projected;
Whether breakage, returns, and promotional costs are included;
Whether this is per-SKU contribution margin or the company’s overall accounting net profit.

The clearer the rules, the more the result can be audited.

You also need to add prohibitions. For example:

Don’t fabricate data that isn’t in the table; if something can’t be determined, label it “insufficient data”; don’t decide on my behalf whether to cut off supply; don’t output customer phone numbers.

This line matters a great deal. It doesn’t guarantee AI will never make a mistake, but it clearly marks the boundaries of the work and reminds the user that they must check the output.

V. Step Four: Deliverable — Don’t Just Ask for “Analysis,” Ask for Something You Can Act On Right Away

A long block of text from AI isn’t necessarily useful for work.

The deliverable should be designed around who will use it next.

For the boss to review, you might ask for:

A one-page summary;
A ten-item list of risks ranked from highest to lowest;
Each item containing only “problem, basis, recommended action, owner.”

For sales reps to use, you might ask for:

A one-store-per-row follow-up table;
Customer, issue, what to ask, deadline;
Grouped into today, within three days, and this week.

For finance to review, you might ask for:

Original order numbers preserved;
The calculation formulas and intermediate values shown;
Missing fields flagged in red separately;
No final credit conclusion drawn.

The more specific the deliverable, the harder it is for AI to paper over gaps with pretty language.

VI. Step Five: Acceptance — Without a Checking Standard, There’s No Real Completion

This is the step people most often skip.

Many people see AI’s output looking neat and orderly and simply assume it’s correct. But a tidy format and factual accuracy are two different things.

Acceptance checking can be done across five dimensions:

Completeness: Are any fields that should be there missing?
Accuracy: Can a spot check of a few numbers be traced back to the original table?
Consistent basis: Are time periods, units, returns, and rebates handled according to the rules?
Actionability: Does the result have an owner, a deadline, and a next step?
Safety: Has any customer information leaked, or has any decision that belongs to a human been overstepped?

It’s best to have AI run its own check first:

Before delivering output, verify item by item: was any data outside the table used, was any return overlooked, were cases and individual units mixed up. List anything that can’t be confirmed under “Needs Human Verification.”

But remember: AI’s self-check cannot replace a human spot check. When money, goods, and customer commitments are involved, the last line of defense is still a person.

VII. Full Walkthrough One: Turning “What to Do About a Customer Who Owes Money” Into a Usable Task

Vague question:

What should I do if a customer owes money and won’t pay?

AI can only offer generic knowledge about communication, negotiation, and legal channels.

Rewritten with the five-step method:

Goal: From the receivables detail, identify the customers that need priority follow-up this week — do not decide on my behalf whether to cut off supply. Materials: The attachment is a receivables table with fields including customer ID, customer tier, amount receivable, agreed payment term, due date, most recent payment date, and account owner; the amount is in RMB, and the data is current as of July 26. Rules: Days overdue = as-of date minus due date; sort first by overdue-days-over-30, then by amount; Class-A customers and new customers should be flagged separately, not directly classified as high risk; rows without a due date go into “Incomplete Data.” Deliverable: Output a table containing customer ID, days overdue, amount receivable, most recent payment date, account owner, recommended follow-up time, and items to be verified; sorted by priority; do not output contact phone numbers. Acceptance: The total receivables amount must match the original table; randomly spot-check the overdue days for three customers; no data outside the table may be used; list any items that cannot be determined at the end.

The value of this task isn’t having AI collect the debt for the boss — it’s letting finance quickly turn a few hundred rows of data into a verifiable follow-up list.

VIII. Full Walkthrough Two: Turning a Sales Rep’s Voice Memo Into a Store To-Do List

After finishing store visits, a sales rep leaves a voice memo:

Old Wang’s convenience store still has enough bottled water, but the beer in the cooler is running low. The owner said school lets out for break next week, so drinks might sell more slowly. That case of damaged goods from last time still hasn’t been dealt with. He’ll be at the store Friday afternoon.

Designing the task with the five-step method:

Goal: Turn the voice memo into a store visit record and a next-step to-do list.

Materials: The sales rep’s voice transcript, the store directory, and the company’s standard record format.

Rules: Only organize the facts that were actually stated; “might sell more slowly” stays as the store owner’s judgment and should not be rewritten as a certain sales decline; the damaged-goods issue is flagged as an after-sales to-do; if no quantity was given, write “quantity to be confirmed.”

Deliverable:

Store	Inventory Observation	Customer Feedback	Issue	Next Step	Deadline
Old Wang’s Convenience Store	Bottled water sufficient; beer running low, quantity to be confirmed	Drinks may slow down after school break	Damaged goods from last time not yet resolved	Check the damage claim; confirm beer inventory at the store Friday	Friday afternoon

Acceptance: The store name is correct; no quantities were made up; the damage issue and the restocking issue were not merged into one; every to-do item has an owner and a deadline.

As you can see, the five-step method isn’t complicated. It’s simply putting into words, in full, the experience a boss would normally pass on to a new hire.

IX. When the Answer Isn’t Good, Check the Task First — Don’t Switch Tools Right Away

If AI’s first output isn’t good, that doesn’t necessarily mean the model is bad. Work backward through the five steps first:

Is the goal too broad;
Are materials missing;
Were the rules never stated;
Is the deliverable vague;
Was there no acceptance standard.

If it’s a tone or formatting issue, you can keep asking for revisions. If it’s a data or methodology issue, go back to the materials and rules — don’t just say “be more professional.”

There’s also a situation where you must stop entirely: when the raw data itself is wrong. Duplicate customer codes, inconsistent date formats, returns not itemized separately — AI cannot fix business facts on your behalf. At most it can help you flag the suspicious points; someone in finance, the warehouse, or business ownership still has to confirm them in the end.

X. How to Use the “Five-Step Card for Delegating Work to AI”

The companion Five-Step Card for this piece isn’t a collection of prompts — it’s a task design sheet.

Every time you hand off a new piece of work, fill it out in order:

Goal: which specific problem is being solved;
Materials: which data and background are provided;
Rules: what basis to use, what’s off-limits;
Deliverable: who it’s for, what it should look like;
Acceptance: who checks it, and what they check.

The first time might take ten minutes to write. Once it works, save it, and next time you can reuse it with new data. That’s the point at which a single question starts to become the company’s working method.

Today’s Action

Pick one exchange from your most recent “unsatisfying” AI conversation.

Don’t switch tools, and don’t go searching for a new all-purpose prompt. Copy the original question onto the Five-Step Card, fill in the goal, materials, rules, deliverable, and acceptance, and try it again.

Put the two results side by side, and you’ll see it for yourself: what really separates the outcomes isn’t a magic prompt — it’s whether the work was fully briefed.

References

Cyberspace Administration of China and six other departments: “Interim Measures for the Management of Generative Artificial Intelligence Services,” 2023-07-13.
Internal project training method: the goal–materials–rules–deliverable–acceptance five-step method, 2026-07.
Internal project research draft: “Research Notes on AI Adoption Among FMCG Distributors,” 2026-07-24.

Next: “Without These Operating Tables, AI Can Only Talk in Pretty Words” — what actually stalls AI adoption for distributors is often not the model, but the data.