Getting Consistent Results
from Claude & ChatGPT
Stop wrong-account mistakes and hallucinated data. Five proven strategies, from quick wins to advanced workflows.
TL;DR
LLMs don't automatically know how to use MCP tools correctly. They can confuse accounts, hallucinate data, or call tools in the wrong order. The fix is giving them the right context up front. Start with the Server Guide (takes 10 seconds), then layer on more strategies as needed.
Why This Happens
Understanding the root cause helps you pick the right fix.
Three Reasons LLMs Get Confused
- No built-in knowledge of your tools. Claude and ChatGPT know nothing about the MAP MCP server until you tell them. Without guidance, they guess which tools to call and in what order — and sometimes guess wrong.
- Multiple accounts look the same. If you manage several brands, the LLM has to pick the right profile for every API call. Without explicit context about which brand you're working on, it can pull data from the wrong account or mix results across accounts.
- Conversations drift. Over a long conversation, the LLM's context window fills up. Earlier instructions about which account to use or how to call a tool get pushed out, and the LLM starts making mistakes it wouldn't have made at the start.
Strategy 1: Load the Server Guide
The single easiest thing you can do. Takes 10 seconds and dramatically improves consistency.
Easiest — do this firstWhat is the Server Guide?
MAP includes a built-in tool called get_server_guide that returns a comprehensive instruction set teaching the LLM exactly how to use every tool correctly — which tools to call, in what order, what parameters to pass, and how to handle multi-account scenarios.
Think of it as an instruction manual that the AI reads before starting work.
How to use it
At the start of every conversation (or whenever you switch topics), just say:
That's it. The LLM will call get_server_guide, read the instructions, and follow them for the rest of the conversation.
Strategy 2: Install the Amazon Ads Skill
Give Claude persistent knowledge about MAP tools so you don't need to load the server guide every time.
Moderate — one-time setupWhat is a Skill?
A Claude Skill is a set of instructions that Claude loads automatically when relevant. Once installed, Claude already knows how to use MAP tools correctly without you having to ask for the server guide each time.
How to install it
We maintain a pre-built skill for Amazon Ads workflows. Install it by telling Claude:
Claude will read the repository and set up the skill. After that, it will automatically reference these instructions when you ask about Amazon Ads.
Strategy 3: Build Custom Skills from Wins
After a successful workflow, teach Claude to remember how it did it. This is the most powerful long-term strategy.
Moderate — do after successful workflowsHow it works
After Claude successfully completes a task — pulling a report, optimizing bids, running a wasted-spend sweep — you can ask it to save what it learned as a reusable skill. Next time you ask for the same type of task, it will reference that skill automatically.
What to say
After any successful interaction, try one of these:
Claude will save the exact tool sequence, parameters, and logic it used so it can replicate the workflow reliably.
Strategy 4: One Project per Brand
For agencies managing multiple accounts, this prevents the #1 mistake: pulling data from the wrong brand.
More setup — biggest payoff for agenciesThe problem with multi-brand conversations
If you manage 5 brands and talk to Claude about all of them in one conversation, it has to remember which profile ID belongs to which brand for every single tool call. As the conversation gets longer, it starts mixing them up or defaulting to whichever brand was mentioned most recently.
How to set up Projects
In Claude, create a separate Project for each brand you manage. In each project's instructions, include:
Now when you open that project, Claude already knows which brand you're working on — no room for confusion.
Strategy 5: Pick the Right Model
Not all models handle complex tool use equally well. The model you pick matters more than you'd think.
Easy — just check your settingsModel comparison for MCP workflows
MCP tool use requires the model to plan multi-step sequences, track state across calls, and interpret structured data. Here's how the current Claude models stack up:
| Model | MCP Tool Use | Multi-Account | Recommendation |
|---|---|---|---|
| Opus 4.6 | Excellent | Excellent | RECOMMENDED |
| Sonnet 4.6 | Good | Good | ACCEPTABLE |
| Haiku 4.5 | Limited | Poor | NOT RECOMMENDED |
Using ChatGPT?
If you're using ChatGPT with the MAP MCP server, use GPT-4o or better. The same principles apply — load the server guide, use projects to separate brands, and pick the most capable model available.
Putting It All Together
Quick-start checklist
Here's the recommended order. Each step builds on the last:
- Right now: Switch to Opus 4.6 (or at least Sonnet) in your AI client settings.
- Every conversation: Start with "Get the server guide from Marketplace Ad Pros."
- One-time: Install the Amazon Ads skill from our GitHub repo.
- After wins: Ask Claude to save successful workflows as skills.
- For agencies: Create a separate Claude Project for each brand you manage.
Ready to Get Started?
Connect your Amazon Ads accounts, load the server guide, and see the difference immediately.
Connect Your Accounts