Prompting That Does Not Waste Credits
Outcome: a set of phrasing habits that make an MCP session cost what you expected, plus a check on your own recent sessions.
- Surface
- MCP server
- Level
- Beginner
- Uses
check_credits- Credits
- 0
- Prerequisite
- A few sessions from lessons 02–06
Five habits
Name the tool
“Use find_work_email” fires reliably. “Get their email” leaves the assistant choosing among several plausible tools, possibly differently on different days.
Cap every run
“Cap at 25.” Without a cap, a search that matches 4,000 rows may attempt 4,000 rows.
Filter before you enrich
“Show me the list first, then I will tell you who to enrich.” Filtering is free; enrichment is not.
Close the scope
“Do not call any other tools.” A helpful assistant will otherwise decide the company details would be useful too.
Stop mid-chain deliberately
“Stop there.” Two-stage prompts — source, review, then enrich — are the single biggest saving available in an MCP session.
Phrases that cost money
| Phrase | What it does |
|---|---|
| ”Tell me everything about…” | Calls every plausible tool |
| ”Find all the…” | No cap — attempts the full result set |
| ”Enrich these companies” | Ambiguous: which enrichment, and how many? |
| ”Get me some leads” | Unbounded sourcing and unbounded enrichment |
| ”And anything else useful” | An open invitation |
| ”Do it for the whole list” | Whatever the list length happens to be |
The pattern: anything unbounded in scope or count. Every one of those is fixed by naming a tool and a cap.
The most expensive prompt in normal use is asking for enrichment before filtering. “Find people who commented on this post and get their emails” enriches the competitor’s employees, three consultants and a job seeker — at full price, before you have seen any of them. Always split it into source, review, enrich.
A prompt template that works
Using the SyncGTM MCP server, <goal in one line>.
Steps:
1. <tool_name> with <parameters>. Cap at <n>.
2. <tool_name> using the output of step 1.
3. <tool_name> on the survivors only.
Return: <exact fields>, as a table.
Rules:
- Do not call any tools other than those named.
- If a step returns nothing for a row, leave it blank and continue.
- Stop after step <n> and show me the result before continuing.Six lines of structure that turn an unpredictable session into a predictable one.
Watch the balance
Check before
check_credits at the start of a working session.
Check after
The difference is what the session cost, and it is the only honest measure.
Compare against your estimate
A large gap means a step ran on more rows than you thought.
Find the step
Usually the one without a cap.
Do this for a few sessions and you will develop an accurate intuition for what things cost — which is what stops the surprises.
Recovering a runaway session
If a run is clearly doing more than you intended:
- Stop it. Interrupt the assistant rather than letting the chain finish.
- Check the balance to see what it actually spent.
- Re-prompt with caps rather than trying to redirect mid-chain.
- Do not ask it to “undo” — enrichment calls are billed and cannot be reversed.
Do this now
Look back at three recent sessions
Find the prompts that lacked a cap or a tool name.
Rewrite one using the template
Run both versions on a small input
Compare the credits used.
Adopt the two-stage pattern
Source, review, enrich — for anything that produces a list.
Check credits before and after your next real session
Write your own worst phrase down
Everyone has one they reach for.
Check your work
- Every prompt names its tools
- Every list-producing step has a cap
- Enrichment always follows a review step
- You measured a session’s cost against your estimate
- You have rewritten at least one habitual prompt
Where this breaks
An assistant that is being helpful is the main source of unexpected cost. Asked for a contact card, it may decide company details, tech stack and recent posts would round out the picture — each a billed call you did not request, and none of them reported as a surprise. “Do not call any other tools” is one line, and it removes the entire category.
Further automation
Once a chain is right, save it as a skill so the caps and tool names are baked in and nobody has to remember them. The GTM agents library has ready-made ones; AI-Powered GTM lesson 05 covers writing your own.
Next lesson
08 — Draft outreach in the conversation, using what you pulled instead of a template.
Reference for this lesson: MCP prompting guide, check_credits, MCP tools, Credits.