HandbookAI assistant

Good prompts

How to ask so that the assistant does what you mean: give it a goal and a scope, say what you know and what it should find out, and check its work as you go. With example prompts for every stage of a study.

You do not need special wording. Write as you would brief a capable colleague who has never seen your product. The examples below use the garden hose reel from the tutorial; every example has a copy button.

Six habits that help#

Give the goal and the scope: What do you want to know, for whom, and where does the system end? Write it once into the model's Goal (or ask the assistant to record it with you); it reads the goal as its brief in every chat.
Say what you know — and what it should find out: Give the figures you have, with units. Name the ones you do not have and say “research these”. Otherwise the assistant cannot tell a fact from a gap.
Ask for sources: “Show me your sources” makes it keep the link with every figure, so you can check it later. See Research and review.
One task, or a plan: A small change is best as one short message. For a whole model, describe the end result and let the assistant plan; you can follow the plan card and step in. Avoid packing three unrelated tasks into one message.
Say how sure you need to be: “A screening is enough” and “this goes into a public report” lead to different amounts of research and different costs.
Correct plainly: Say what is wrong and what you want instead. You can type while it works — your remark waits as a chip and is read after the current step.

Build a model#

Model a garden hose reel set, cradle to gate: a 20 m PVC hose (about 2.6 kg) with brass fittings on a polypropylene reel, made in China and shipped to Hamburg. Research the masses you do not know and show me your sources. Group the processes by life-cycle phase.

Before you build anything, ask me what you need to know about the functional unit and the system boundary.

Add the use phase: the reel is used for 10 years, and the hose is replaced once. Put the lifetime and the number of replacements into parameters.

Replace the polypropylene of the reel with 50 % recycled polypropylene. Tell me which dataset you use for the recycled share and why.

Run the health check and fix what you can. List anything you are not sure about instead of changing it.

Research data#

What is the typical brass composition for garden hose fittings in the EU? Give a range and the sources, and say how confident you are.

Here is the supplier's datasheet. Take the masses of all parts from it, put them into parameters, and cite the page for each.

Find the sea distance from Ningbo to Hamburg, and the road distance from Hamburg to a warehouse in Kassel. Use the distances tool.

Use only my data and the background database for this — no web search. Tell me which figures are missing.

Scenarios#

Set up the end-of-life of the PVC hose in Germany from Eurostat, as parameters that sum to 1, and create a second scenario with 100 % incineration.

The hose mass is uncertain. Make scenarios for the low and high end of the range you researched and compare them with the baseline for climate change.

Which three assumptions change the result most? Make a scenario for each and tell me by how much.

Understand the results#

Calculate climate change with EF 3.1 and tell me which three processes contribute most, and why.

Is the polypropylene reel itself the emitter, or its supply chain? Split the score.

Explain the land-use result in plain words, as if to a product manager. What drives it, and is the magnitude plausible?

Set up the Results page to compare this model with the recycled-reel scenario for all EF 3.1 indicators.

Run a critical review of this model before I use the numbers, and go through the findings with me.

Prepare the report#

Report chapters are written on the Report page, not in the chat: the assistant in the chat has no tool for them and will point you there. Two things help the chapters:

Before writing, use the chat to get the model's record straight — goal and scope, sources, open review findings. The writer works from that record.
While writing, put chapter-specific wishes into Additional instructions for this chapter in the chapter editor. When the writer needs a decision from you, it asks above the chapter list; answer it and Update report works the answer in. See The full study report.

Check that goal and scope are complete for a report: functional unit, system boundary, cut-off rules, allocation, data quality. Ask me about anything missing and record it.

Which review findings are still open? Record each one as fixed, rejected with a reason, or open, so the report can list them.

An example of additional instructions for the interpretation chapter:

Address a technical audience. Compare the recycled-reel scenario with the baseline, and state clearly that the hose mass is the largest uncertainty.

Less good → better#

Less good	Better	Why
“Make an LCA of a hose reel.”	“Carbon footprint of a 20 m garden hose reel set, cradle to gate, for an internal design comparison. Research what you do not know.”	Goal, scope and the task for gaps are clear.
“The numbers are wrong.”	“The hose weighs 2.6 kg, not 4 kg. Change it and recalculate.”	Says what is wrong and what is right.
“Find data for brass.”	“Find the typical brass composition for hose fittings in the EU, with a range and sources.”	Names product, region and the form of the answer.
“Do everything for the report.”	“Check goal and scope for completeness and record the open review findings.”	One clear task the assistant can actually do.
“Use better data.”	“For the reel, look for a dataset closer to injection-moulded polypropylene in Asia, and explain the trade-off.”	Says what “better” means.
“Why is it so high?”	“Climate change is 18 kg CO2-eq. Which processes drive it, and is that plausible for a product of 4.5 kg?”	Gives the number and asks for a check, not only an explanation.
Three unrelated requests in one message	One message per task, or one goal with “make a plan”	Easier to follow, check and undo.

When it goes the wrong way#

Interrupt early. Type a remark during the run, or press Stop. A short correction now is cheaper than a long repair later.
Undo the turn and rephrase. Undo can take back all changes of one assistant turn at once (Work with the assistant). Then say what you want more precisely.
Ask why. “Why did you choose this dataset?” is answered from its actual work, not a guess.
Start a new chat for a new task. A long conversation costs more on every message and can drift.

Where to go next#

Work with the assistantThe chat dock, the plan, approval and undo. Research and reviewSources, confidence and a second opinion. The full study reportWrite the chapters on the Report page.