Best Prompts for Operations

Weekly reviews often become tours of status updates. Every team reports activity, the meeting runs long, and the real constraint remains hidden across queue sizes, handoff delays, rework, and unresolved dependencies. By the time someone names the bottleneck, the group has little time left to decide what to change.

A structured AI review can turn operating data into a sharper agenda. The aim is not to produce a polished summary. It is to identify where work is accumulating, show the evidence, and focus the team on a small number of testable interventions.

Why ordinary prompting fails

“Find our bottlenecks” invites the model to treat every delay as equally important. It may confuse a one-off incident with a recurring constraint, assume that correlation proves cause, or recommend broad process changes without owners or measures.

The prompt needs a defined period, a clear workflow, comparable measures, and rules for distinguishing facts from hypotheses. It should also resist false precision when inputs are incomplete.

Reusable prompt

Copy
ROLE
You are an operations analyst preparing a weekly bottleneck review. Your job is to identify constraints supported by the supplied data and propose small, reversible next steps. You do not invent metrics or claim causes that the evidence cannot establish.

REQUIRED INPUTS
1. Review period and comparison period: [dates]
2. Workflow stages and expected sequence: [list]
3. Weekly data by stage, such as arrivals, completions, backlog, age, cycle time, wait time, defects, rework, and capacity: [table]
4. Service targets or agreed thresholds: [list]
5. Known incidents, staffing changes, holidays, or policy changes: [notes]
6. Previous actions and their observed results: [list]

STEPS
1. Validate the input coverage and flag missing or non-comparable fields.
2. Map flow through the named stages.
3. Find material changes in backlog, aging, throughput, waiting, defects, and rework.
4. Rank up to three likely bottlenecks by impact on total flow, recurrence, and strength of evidence.
5. For each, distinguish observed facts, plausible causes, and unanswered questions.
6. Check whether a downstream blockage may be creating an upstream queue.
7. Propose one small intervention or diagnostic test per bottleneck, with an owner, measure, and review date.
8. Note what improved and should not be disrupted.

OUTPUT FORMAT
- Executive view: 5 bullets maximum
- Data quality warnings: table with issue and consequence
- Ranked bottlenecks: table with stage, evidence, affected work, recurrence, confidence
- Cause hypotheses: observed fact, hypothesis, disconfirming evidence needed
- Recommended actions: action, owner, measure, due date, decision at next review
- Questions for the meeting: 3 to 5
- Watch list: signals to monitor next week

EVIDENCE AND UNCERTAINTY RULES
- Cite each conclusion to a supplied metric, date, or note.
- Do not estimate missing values unless explicitly asked; label any calculation and show its inputs.
- Do not treat a simultaneous event as a proven cause.
- Use High, Medium, or Low confidence for each ranked bottleneck.
- If periods or definitions are not comparable, say so before drawing a trend.
- Prefer “the data suggests” to certainty when causal evidence is absent.

What to provide

Provide stage-level data in a consistent table, with definitions for every metric. Include both the current week and a useful comparison, such as the previous four weeks, if available. Raw counts alone can mislead, so include aging and waiting information where possible.

Add known context without pre-deciding the answer. A note that two team members were absent is useful. A note that “staffing caused the backlog” is already a conclusion. Include actions agreed in the previous review so the prompt can distinguish a new problem from a failed or unfinished experiment.

How to review the output

Start with the data quality warnings. If stage definitions changed or a week is partial, correct the comparison before discussing causes. Recalculate any derived figures and inspect the underlying work items behind the largest queue.

Then challenge the ranking. A large backlog may be intentional buffer stock, while a small queue of high-priority approvals may stop the entire workflow. Ask whether the proposed evidence reflects total flow and customer impact, not simply the most visible number.

Review each recommended action for scope. It should be specific enough to try within the next cycle, assigned to a named person, and paired with a measure that can show whether it helped.

Where it fails

A weekly dataset may show where work waits but not why. Interviews, work sampling, or direct observation may be needed. Aggregate metrics can also hide differences between work types, priority classes, regions, or customer segments.

The method is weak when teams change definitions from week to week, report only averages, or omit abandoned and reworked items. It should not become an automated performance score for individuals. Bottlenecks are usually properties of the workflow and its constraints, not proof that one person is underperforming.

Practical takeaway

A useful bottleneck review converts scattered measures into a short list of evidence-backed decisions. Keep the prompt stable, improve the inputs over time, and compare whether last week’s action changed the constraint. Try it in Agentic Workers to prepare the next review, then let the team challenge every hypothesis before acting.

<!-- X derivative: A weekly operations review should rank constraints with evidence, separate causes from hypotheses, and end with small actions that have owners and measures. -->