Lesson 3.3.2
Build and enrich a list without opening the app
What this costs
The same as doing it in the app. Set a ceiling in your first message.
Before you start
A working MCP connection from lesson 3.3.1.
The same three moves, in conversation. Source, curate, enrich. The difference is that you describe each move instead of configuring it, and the agent picks the tools. Everything you learned about conditions and cost still applies, and now you have to say it out loud rather than click it.
Open with the constraints, not the request
The single biggest difference between a good and a bad MCP session is the first message. An agent given a goal and no constraints will pick the most thorough path available, which is also the most expensive one.
I want to build a prospect list using the Kuration tools. CONSTRAINTS, follow these for the whole session: - Tell me the estimated credit cost before running anything that spends credits, and wait for me to confirm. - Never run an enrichment on more than 50 rows without asking first. - Filter before enriching, always. Do not enrich rows we are about to cut. - If a tool returns nothing on a row, tell me whether it was skipped or genuinely not found. - Keep a running total of credits spent and show it after each step. THE LIST I WANT: Independent bakeries in Lisbon that have their own website. About 40 of them, with the owner or manager and a contact email. Start by telling me which builder you would use and what the sourcing step will cost. Do not run it yet.
Six constraints and then the request. Those six lines are the whole cost discipline from the Practitioner track, restated in a form an agent will follow.
- 1
Ask for the plan before the run
Which builder, which columns, in what order, and what each step costs. Read it. This is the moment to catch an agent that was about to enrich before filtering.
- 2
Approve one step at a time at first
Until you trust the pattern, confirm each spending step individually. After a few sessions you will happily approve a whole plan at once.
- 3
Make it filter out loud
Ask what the row count is after each filter. An agent that never mentions row counts is probably not filtering.
- 4
Score before you enrich, because scoring is free
Same reasoning as lesson 2.2. Tell the agent to score first and enrich only above a threshold.
- 5
Ask for the result as a table in the conversation
Read twenty rows before exporting. You are checking the same things you would check in the app.
- 6
Export, and reconcile the credits
Ask for the final balance and compare it to the running total. If they disagree, something ran that you did not approve.
No constraints in the opening message
- What you see
- A large spend on a thorough plan you did not ask for
- The fix
- Paste the constraint block first, every session.
Agent enriched before filtering
- What you see
- Full row count on an expensive column
- The fix
- Make filter before enrich an explicit constraint, and ask for row counts after each step.
Vague market description
- What you see
- A list that is technically correct and useless
- The fix
- The agent inherits your precision. Describe the market as tightly as you would configure a builder.
Never checked the balance afterwards
- What you see
- A surprise at the end of the week
- The fix
- Balance at the start, running total throughout, balance at the end.
Check yourself
0 of 2 answered1.What is the single biggest difference between a good and a bad session?
2.Which constraint matters most?
Try it yourself
Using the constraint block above, build a 40 row list end to end in conversation. Do not open the Kuration app at any point.
- You have done it when
- A finished list, and the credits you actually spent match the running total the agent reported.
- Credit budget
- Whatever ceiling you set, and you approved every step of it.