Dr. Use Case · Field Lab
Completed Field LabLab Nr. CF13

Why better prompts don't automatically use fewer credits

Reverse prompting, prompt optimisation, Cowork and a gelato use case that suddenly became rather expensive. What started as a session for the HELish Summit in Helsinki turned into a workload design experiment.

22 Sep 2026 12 min readReverse Prompting · Workload Design · Credits

Research question

How do I design an AI workload so the model does exactly the work the use case actually requires?

Use Case Ticket

Why better prompts don't automatically use fewer credits

Scenario

Mid September 2026: Alpine Gelato has to end the summer, in theory. Mango, Raspberry and Lemon are still selling surprisingly well, while Cinnamon Plum and Chestnut need to go into production, the first Christmas orders are coming in and Vanilla is becoming more important for the Q4 business. Baked Apple has demand and ingredients, but no complete product release yet. The management question: which summer flavours should be phased out, which core flavours protected, which autumn and winter flavours ramped up?

Test setup · Inputs

v1
  • Inventory
  • Sales Orders
  • Seasonal Demand History
  • Production Capacity
  • Procurement & Lead Times
First observation

“A summary says what happened. A briefing shows what matters now.”

01

Question Card

Starting question

Mid September 2026: Alpine Gelato has to end the summer, in theory. Mango, Raspberry and Lemon are still selling surprisingly well, while Cinnamon Plum and Chestnut need to go into production, the first Christmas orders are coming in and Vanilla is becoming more important for the Q4 business. Baked Apple has demand and ingredients, but no complete product release yet. The management question: which summer flavours should be phased out, which core flavours protected, which autumn and winter flavours ramped up?

Research question

How do I design an AI workload so the model does exactly the work the use case actually requires?

02

Setup Module

Test setup

Instead of writing an ever longer prompt, we started with reverse prompting. The SharePoint archive held planning documents from 2022 to 2025, from Season_Transition_Notes_2022.docx to FINAL_final_transition_notes.docx. A meta prompt was not meant to copy the output, but to reconstruct the method: objective, decision logic, sequence, business rules, quality checks. Out came rules such as: check Q4 demand before reducing core flavours; never assess inventory in isolation; review procurement risks before ramping up a flavour. In the second experiment the master prompt was optimised again, not shorter but more precise: clearer scope, explicit boundaries, stopping rules and a clear statement about which work the AI should explicitly not do.

  • 01Inventory
  • 02Sales Orders
  • 03Seasonal Demand History
  • 04Production Capacity
  • 05Procurement & Lead Times
03

Prompt Box

The original prompts

Meta prompt 01

Reverse prompting: recover the recipe

Reconstructs the method behind existing work results and derives a reusable master prompt from it.

You are an experienced prompt engineer and work design analyst.

Analyse the examples I give you as evidence of how this work functions when it is done well. Do not simply summarise them and do not copy them. Instead, reconstruct the underlying method.

Identify:

- the objective of the work
- the recurring decision logic
- information that materially influences the outcome
- the sequence of work steps or considerations
- recurring business rules and constraints
- useful output structures
- quality checks and decision points
- assumptions that must be re-examined on every run
- elements that are merely formatting, habit or historical noise

Then create a reusable master prompt that lets an AI system perform the same kind of work with new inputs.

The master prompt should:

- preserve the useful decision logic without blindly adopting historical conclusions
- clearly separate stable rules from assumptions that need to be re-examined
- define the minimum necessary information
- specify the expected output and decision categories
- flag missing or contradictory information instead of inventing it
- include explicit quality checks

Before the final master prompt, first show me which logic you extracted and which elements you deliberately excluded, including your reasoning.

Meta prompt 02

Prompt optimisation: narrow the workload

Reviews an existing prompt for unnecessary work, unclear scope and missing boundaries, and delivers a more precise version.

You are an experienced prompt engineer and AI workload optimizer.

Review the following prompt as if it were run repeatedly in an AI work environment where answer quality, unnecessary work and resource consumption matter.

Optimise it regarding:

- clarity of the actual objective
- minimum necessary context
- explicit scope and clear boundaries
- precise use of sources and tools
- redundant instructions
- unnecessary research, analysis or output
- clear decision criteria
- handling of missing information and uncertainty
- precise but complete output requirements

Do not shorten the prompt simply for the sake of shortening it.

Preserve everything that materially improves decision quality.

First show:

1. What is unclear or unnecessarily broad
2. What could trigger unnecessary work
3. Which information is missing
4. What should be removed, merged or more clearly limited

Then deliver:

A. The optimised prompt
B. What was changed and why
C. Which work the AI should now deliberately no longer do
D. Which remaining assumptions still require human judgement

Prompt to optimise:

[INSERT PROMPT]
04

Comparison Module

Results compared

Both runs use the same task and the same data. The difference is not the information, it is the instruction.

Broad vs. deliberately narrowed
Run A · Original prompt
generic

2,416 credits

Broad scope, no stopping rules: Cowork orchestrates more freely, researches more and produces longer output that nobody needs.

  • Scope too broad. The model keeps researching even though there is more than enough evidence already.
  • No stopping rule. No clear boundary for when the analysis ends.
  • Analysis-only test. 206 credits, exactly identical to Run B. Better prompts do not automatically save credits.

Same task, same data, same model, same effort level, only the instruction was different.

Output type:Description
Recommendation
Run B · Optimised promptcurated

2,024 credits

Precise scope, explicit boundaries and clear rules about which work the AI should not do. Result: 392 credits less, 16.2 percent.

  1. 01

    −392 credits (−16.2 %)

    Not a benchmark, but one concrete workload: the narrowed instruction measurably needed less.

  2. 02

    GPT-6 Astra: 1,717 credits

    The same optimised workload with Astra: roughly 15 % below Run B and almost 29 % below Run A, and noticeably faster.

  3. 03

    Vanilla: Protect / Increase

    2,110 units in stock look like overstock, but Q4 demand, open orders and a long procurement lead time flip the recommendation.

  4. 04

    Baked Apple: Delay / Investigate

    Demand and ingredients are there, but the product release is not complete. Demand is not automatically a production release.

Model choice is part of workload design: the strongest model was not the most expensive one.

Output type:Decision basis
vs.
05

Why this matters

Interpretation

First things first: in the small analysis workload both prompts cost exactly the same, 206 credits. So the nice story that better prompts automatically save credits is not always true. Only in the large end-to-end Cowork job did the narrowed instruction become visible: −16.2 percent. And GPT-6 Astra came in below that again. At the latest here it was clear: this was no longer a prompt experiment, it was a workload design experiment. Model strength, speed and resource consumption are not linearly connected, so “we always take the strongest model” is just as little of a strategy as “we always take the cheapest one”. And reverse prompting is not a pure prompt technique, it is a method for making existing organisational experience reusable. The best template for your next AI workload may not be a prompt catalogue at all, but the best management decision from last year, or indeed FINAL_final_transition_notes.docx.