Prompting Playbook
You don’t need to speak AI. You just need to ask a good business question.
Agent Army is designed to let you ask questions the same way you would ask a strong analyst sitting beside you.
No code. No magic words. No elaborate prompt engineering.
But there is an important difference between:
“How are sales?”
and:
“How are Walmart POS sales trending for the last four weeks compared with the prior four weeks, and what items are driving the change?”
Both are understandable. Only one gives your agents enough direction to get immediately useful.
Agent Army MISSION Prompt Framework
When a question matters, give the agents a MISSION:
M = Metric
Tell us what you want measured. Agent Army’s semantic layer is built to map those terms to governed definitions rather than allowing the AI to invent what a metric means.
Sales dollars? Units? Inventory? Fill rate? Forecast accuracy? Deductions? On-hand? Store count? Margin? Use the business term your company normally uses.
Instead of: “How are we doing?”
Try: “How are POS sales dollars and units trending?”
I = Intent
Tell the agent what you’re trying to understand.
Are you trying to: compare performance, find an anomaly, diagnose a decline, identify an opportunity, validate a hypothesis, forecast what comes next, or prioritize where the team should act?
Instead of: “Show me inventory.”
Try: “Identify items where inventory levels are likely to create an out-of-stock risk.”
S = Scope
Define the boundaries of the question. Include the dimensions that matter, such as retailer, banner, category, brand, item, geography, store, distribution center, customer, or program.
Instead of: “What products are down?”
Try: “For Walmart US, which items are down in POS units over the last four weeks?”
S = Specific Timeframe & Comparison
Tell the agent when and compared with what. Don’t make the agent guess what you consider a meaningful baseline. Fuzzy dates are okay, but keep in mind how fiscal years and weeks are defined by retailer.
- Good: “Compare Walmart weeks 20–23 with weeks 16–19.”
- Better: “Compare POS sales dollars and units for Walmart weeks 20–23 against weeks 16–19 and the same four weeks last year.”
I = Investigate
Don’t stop at what happened. Ask why. This is where conversation becomes much more powerful than a static dashboard.
The agents are designed for multi-step analysis including filtering, joining, anomaly detection, forecasting, and root-cause investigation.
Useful follow-ups include:
- “Which items explain most of the change?”
- “Is this primarily distribution, velocity, inventory, or availability?”
- “Where did the trend begin?”
- “What changed from the prior period?”
O = Output
Tell Agent Army how you want the answer delivered when format matters.
Examples:
- “Give me the top 10 items.”
- “Show the result by item and week.”
- “Give me three bullets I can use in tomorrow’s sales meeting.”
You don’t need to specify an output every time, but doing so can save a round trip of follow up questions when you already know what you need.
N = Next Action
Finish analytical questions with the part dashboards often forget:
What should we look at next?
Try:
- “Prioritize the five issues with the greatest revenue impact.”
- “Which of these exceptions require human attention first?”
- “What should the sales team investigate?”
From Weak Prompt to Agent-Ready Prompt
Weak: “Why are sales bad?”
Better: “Why are Walmart sales down?”
Agent Army: “Compare Walmart US POS sales dollars and units for the last four completed Walmart weeks against the previous four weeks and the same period last year. Identify the items and stores contributing most to the decline. Investigate whether the primary drivers are velocity, distribution, inventory, or out-of-stock conditions. Rank the five largest opportunities by estimated sales impact and show the source metrics used.”
That’s not “prompt engineering.”
That’s a well-formed analytical question.
A Few Rules from Mission Control
Ask naturally.
Agent Army’s approach uses plain-language questions and governed analytics so the conversation can start with the business problem rather than SQL.
Be precise about business meaning, not database structure.
You shouldn’t need to know table names or joins. Tell Agent Army what the business question is. Let the system handle the plumbing.
Use your company’s real terminology.
If your organization says “POS Sales,” “Net Sales,” “Shipped Sales,” “Forecast,” or “Modular,” use those terms consistently. Semantic clarity is much more valuable than fancy prompting.
Don’t hide ambiguity.
If “sales” could mean POS sales, shipped sales, or invoiced sales, say which one you mean. If you genuinely don’t know, ask: “Which sales metrics are available, and how are they defined?”
Specify the grain when it matters.
Total Walmart, category, item, store, week, store-item-week, and distribution-center analysis answer different questions.
Ask comparisons explicitly.
“Sales are $4.2 million” is information. “Sales are down 7.4% versus the prior four weeks, driven primarily by three items” is insight.
- Ask for exceptions instead of oceans of data. Try: “Show me the exceptions that matter.” Not: “Give me every store.”
A useful analysis narrows a large set of records to the exceptions a person needs to investigate. Confirm the source scope, rules and validation with the team.
Use follow-up questions.
Use conversational exploration rather than treating every question as a disconnected query. A good analytical conversation often moves from what happened → where → why → what matters → what next.
Tell the agent when documents matter.
Agent Army can work across structured data and documents when the required sources are connected and the implementation supports the workflow. Confirm the scope before combining warehouse information with invoices, contracts, reports or document repositories.
For example: “Compare the contracted allowance rate in our agreement with the deductions actually taken this quarter and flag discrepancies.”
That’s a much richer question than asking the database or contract independently.
Ask for evidence when the answer matters.
Try:
- “Show me the source.”
- “Which metric definition did you use?”
- “What data supports that conclusion?”
- “Give me the lineage for this number.”
Ask for metric definitions, source references and the available audit trail so answers can be traced rather than simply trusted because an AI sounded confident. Confirm the traceability available in the scoped implementation.
Don’t lead the witness.
Avoid: “Prove that out-of-stocks caused our sales decline.”
Prefer: “Investigate the sales decline and identify the strongest drivers. Determine whether availability or out-of-stocks materially contributed.”
We want the answer the data supports, not the answer the prompt suggests.
Don’t ask the AI to guess.
If the available data can’t answer the question, the correct response is to say so. Governed business definitions are a guardrail: when a question cannot be resolved to agreed metrics, the missing definition needs review rather than invention.
At Agent Army, we’d rather give you a useful “I don’t know yet” than an impressive pile of nonsense.
Treat high-impact decisions differently from casual exploration.
AI can investigate, organize, compare, and surface evidence with remarkable speed. Humans still provide judgment, context, accountability, and approval where the consequences matter.