Skip to main content

Instructions

Step-by-step process for how the Agent should approach tasks. Instructions are the methodology—the "recipe" for analysis.

Why This Matters

Instructions determine:

  • Quality of analysis - Good instructions lead to systematic, complete analysis
  • Consistency - Same instructions produce similar outputs for similar questions
  • Efficiency - Well-designed steps prevent wasted effort on dead ends
  • Completeness - Instructions ensure nothing important is missed

Without instructions, an Agent might:

  • Jump to conclusions without exploring data first
  • Present findings without validating them
  • Miss important context or confounding factors
  • Present incomplete analysis

The Anatomy of Good Instructions

Good instructions follow a logical flow:

  1. Understand/Clarify - Get context before diving into data
  2. Explore - Understand the data available
  3. Analyze - Execute the core analysis
  4. Validate - Check the findings for accuracy
  5. Interpret - Explain what it means
  6. Recommend - Suggest actions

This mirrors how expert analysts actually think.

Real Example: Poor vs. Good Instructions

Poor Instructions

Instructions:

1. Query the data
2. Calculate metrics
3. Present results

Problems:

  • Too vague (which data? which metrics?)
  • Missing validation step
  • No context gathering
  • Might miss important context

Better Instructions

Instructions:

1. Understand the business question
- What decision does this insight inform?
- What stakeholders will use this?
- What time period matters?

2. Explore available data
- What tables contain relevant data?
- Check data quality (nulls, outliers, timing)
- Identify any data gaps

3. Calculate core metrics
- Primary metric relevant to question
- Supporting metrics for context
- Segment breakdowns if relevant

4. Compare to baselines
- vs. historical performance
- vs. targets/goals
- vs. peer benchmarks if available

5. Validate findings
- Check using alternative methods
- Look for data quality issues
- Verify numbers are reasonable

6. Interpret findings
- What does this actually mean?
- Why might this be happening?
- What's NOT explained by the data?

7. Recommend next steps
- What action should be taken?
- What questions remain unanswered?
- What validation would increase confidence?

This is much completer and more systematic.

Instructions Template by Domain

For all Agents, use this structure:

<INSTRUCTIONS>

1. CLARIFICATION/EXPLORATION PHASE
- [Step 1: Understand what's being asked]
- [Step 2: Identify relevant data sources]
- [Step 3: Check data quality/availability]

2. ANALYSIS PHASE
- [Step 1: Calculate primary metrics]
- [Step 2: Segment or drill down]
- [Step 3: Compare to baselines/benchmarks]

3. VALIDATION PHASE
- [Step 1: Cross-check using alternative method]
- [Step 2: Look for data quality issues]
- [Step 3: Verify numbers are reasonable]

4. INTERPRETATION PHASE
- [Step 1: What's the key insight?]
- [Step 2: Why is this happening?]
- [Step 3: What assumptions underlie this insight?]

5. RECOMMENDATION PHASE
- [Step 1: What action should follow?]
- [Step 2: What validation would help?]
- [Step 3: What questions remain?]

</INSTRUCTIONS>

Real Domain Examples: Instructions in Detail

Example 1: Customer Churn Analyst

<INSTRUCTIONS>

1. CLARIFICATION PHASE
- Confirm what "churn" means (subscription ended? No activity X days? Other?)
- Identify time period to analyze
- Confirm which customer segments matter (all? By company size? By geography?)

2. DATA EXPLORATION
- Check subscription table for completeness
- Verify churn event dates are accurate
- Look for data quality issues (missing dates, duplicates)
- Identify which data fields explain churn (reason codes, support tickets, usage)

3. CHURN CALCULATION
- Calculate gross churn rate (% of customers lost)
- Calculate revenue churn (% MRR lost)
- Segment by: company size, industry, acquisition source, tenure
- Calculate cohort retention curves (how long do cohorts typically stay?)

4. ROOT CAUSE ANALYSIS
- Correlate churn with: usage patterns, support tickets, feature adoption
- Identify which segments have highest churn
- Look at churn reason codes (why did customers say they left?)
- Compare to historical churn (is this increase real or seasonal?)

5. VALIDATION
- Cross-check churn calculation using two methods
- Verify any unusual segments have sufficient sample size
- Check if churn reasons align with observed behaviors
- Look for seasonal patterns (Jan always higher, Q4 lower, etc.)

6. INTERPRETATION
- What's the key churn trend?
- Which segments are most at-risk?
- What's the likely cause for each segment?
- What's surprising or unexpected?

7. RECOMMENDATIONS
- For each high-risk segment: specific retention actions
- Early warning signals to monitor
- What further analysis would help?

</INSTRUCTIONS>

Example 2: Inventory Optimization Analyst

<INSTRUCTIONS>

1. CLARIFICATION PHASE
- Confirm scope: all SKUs or specific categories?
- Which locations/channels are in scope?
- What's the planning horizon (1 week? 1 month? Quarter?)
- Are there business events (promotions, seasonality) to account for?

2. INVENTORY DATA EXPLORATION
- Assess current inventory levels by location
- Understand sales velocity (units/day) for each SKU
- Check for data quality (phantom inventory? Timing lags?)
- Identify lead times and replenishment cycles
- Calculate holding costs (storage, capital, spoilage)

3. DEMAND FORECASTING
- Project demand for each SKU (normal trend + seasonality)
- Account for known events (promotions, seasonality)
- Estimate demand variability/risk (fast movers vs. slow movers)
- Calculate safety stock needed to cover variability

4. OPTIMAL INVENTORY CALCULATION
- Calculate Economic Order Quantity (EOQ) for each SKU
- Determine safety stock levels based on demand variability + lead time
- Set target inventory levels (EOQ + safety stock)
- Compare current levels to targets

5. OPPORTUNITY IDENTIFICATION
- SKUs below safety stock (stockout risk)
- SKUs above 2x target inventory (excess/dead stock)
- Slow-moving items approaching obsolescence
- Opportunities to rebalance across locations

6. VALIDATION
- Verify calculations using two methods (if possible)
- Spot-check a few SKUs with domain experts (are targets reasonable?)
- Check if historical stockout/overstock patterns align with model
- Verify any unusual recommendations make sense

7. RECOMMENDATIONS
- Immediate replenishment needs (next 1-2 weeks)
- Rebalancing opportunities (move inventory between locations)
- Items to hold/reduce orders on
- Next steps for improvement

</INSTRUCTIONS>

Example 3: Manufacturing Quality Analyst

<INSTRUCTIONS>

1. CLARIFICATION PHASE
- Confirm quality metrics to focus on (defect rate? Cpk? First-pass yield?)
- Identify which production lines or time periods
- Confirm inspection method (100% automated? Sampling?)
- Are there known recent changes (new supplier? Equipment?) to investigate?

2. DATA QUALITY ASSESSMENT
- Check inspection data completeness
- Verify time stamps are accurate
- Assess inspection method consistency
- Look for any data gaps or anomalies

3. QUALITY METRICS CALCULATION
- Calculate defect rate (% or PPM)
- Calculate process capability (Cpk)
- Calculate first-pass yield
- Segment by: production line, shift, time period

4. TREND ANALYSIS
- Create control charts (looking for out-of-control signals)
- Identify when quality changed
- Distinguish normal variation from special cause variation
- Look for patterns (by shift? by supplier batch? by equipment?)

5. ROOT CAUSE ANALYSIS
- For each quality issue, identify potential causes
- Correlate quality with: process variables, equipment maintenance, material supplier
- Interview production teams about recent changes
- Look for relationships (temperature drift + dimensional issues?)

6. VALIDATION
- Verify control chart signals (are they real or noise?)
- Cross-check data from multiple sources
- Confirm suspected root causes with observation (if possible)

7. RECOMMENDATIONS
- For each identified issue: specific corrective action
- Expected impact on quality metrics
- Timeline and resource requirements
- What follow-up validation is needed?

</INSTRUCTIONS>

Instructions Design Principles

  1. Logical Flow Instructions should follow a natural progression:

    • First understand what you're trying to solve
    • Then explore what data you have
    • Then do the analysis
    • Then validate it
    • Then interpret it Bad flow: "Calculate metrics, then explore data, then validate" (backwards!)
  2. Anticipate Failure Points Instructions should address common mistakes:

    • "Don't compare physicians without risk adjustment" (prevents unfair doctor-shaming)
    • "Always verify sample size before making claims" (prevents statistical errors)
    • "Account for seasonality before calling something a trend" (prevents false alarms)
  3. Build in Quality Gates Instructions should include validation steps:

    • Cross-check using two methods
    • Verify unusual findings
    • Check for data quality issues
    • Compare to historical patterns
  4. Balance Flexibility and Structure Instructions should:

    • Provide enough structure for consistency
    • Allow flexibility for different questions
    • Be specific enough to be useful
    • Not be so rigid they prevent good judgment

Good: "Calculate churn by segment (they may vary significantly)" Bad: "Calculate exactly 5 segments using these exact fields"

Common Instruction Mistakes

MistakeProblemExampleFix
Too vagueAgent doesn't know what to do"Analyze the data""Calculate churn rate by company size and compare to 3-month average"
No validation stepAgent doesn't check workMissing cross-validationAdd step: "Verify using alternative calculation method"
Missing context gatheringAgent misses important business contextDoesn't ask about known promotionsAdd step: "Confirm any planned events affecting expected patterns"
No prioritizationAgent might analyze irrelevant thingsSpend time on 100th-order effectsAdd step: "Focus first on [high-impact area], then drill deeper if time allows"
Not domain-awareAgent doesn't account for domain-specific issuesComparing small hospitals to large ones without risk adjustmentAdd step: "Risk-adjust all comparisons for [domain-specific factor]"
Share this page