On this page
The AI Context Debugger, under Settings, lets you paste a system prompt and a test question and see exactly how many tokens it costs, whether it fits comfortably inside the model's context window, and what the model actually replies. It is a token and prompt-testing tool, not a live replay of a real customer conversation.
AI debugger: first 5 of 6 steps
- 1Open Settings, then find the AI Context
- 2Paste your full system prompt
- 3Enter a test question
- 4Choose a provider to test with
- 5Click Run Context Test
Before you start
- The debugger needs the AI Assistant plan feature (Growth and above, or grandfathered) and at least one of an OpenAI or Anthropic key configured in Settings, then AI & Flows. It does not test with Groq.
- It does not automatically pull in your knowledge base content or a real conversation's history; you paste in exactly what you want tested, nothing more.
- Use it before shipping a long or complex system prompt, particularly one built for the AI Response flow node or the website chat widget, to catch a prompt that is unexpectedly large before customers ever see it.
- Only Owner and Admin roles typically reach Settings pages that touch AI configuration; if you cannot see the debugger, check your role's access.
- The tool tests the prompt in isolation, on a single question with no conversation history behind it, so it will not reproduce every nuance of a multi-turn conversation exactly. Treat its numbers as a strong signal, not a guarantee of identical behavior in production.
Steps
- Open Settings, then find the AI Context Debugger page. It shows a warning if no AI key is configured yet, with a direct link to add one.
- Paste your full system prompt into the System Prompt field. The field shows a live estimated token count as you type or paste, before you even run the test.
- Enter a test question, or use the default "Hello, what can you help me with?" if you just want a general check.
- Choose a provider to test with. OpenAI (GPT-4o Mini) and Anthropic (Claude Haiku) are the two options; whichever has no key configured is greyed out.
- Click Run Context Test. The tool sends your prompt and question to the real provider and reports back actual, not estimated, token usage.
- Read the verdict banner and stats. It reports the percentage of the model's context window used, the exact prompt and completion token counts, and the model's actual reply text.
A real-estate agency that has been adding rules to its AI Response node's system prompt for months, one edge case at a time, is a good example of when this tool earns its place: what started as 200 words describing the business can quietly grow into several thousand words of accumulated instructions. Pasting the current prompt into the debugger before adding one more rule shows whether it is still comfortably small, or whether it is time to trim before the next customer conversation pays for it in latency and cost.
What you will see
A short, focused system prompt returns a verdict like "Context is short, AI receives 100% of your prompt every time," with a low percentage bar. A very long prompt, built up over months of additions, might return "Context is very large, consider trimming your system prompt," with the percentage bar most of the way full. Either way, you get the model's actual reply to your test question underneath the stats, so you can judge the prompt's quality at the same time as its size.
The stats grid breaks the total down into four figures: the estimated system prompt tokens and question tokens (calculated before the API call, from a rough character count), and the actual prompt tokens and response tokens (reported back by the provider itself after the call completes). Comparing the estimated and actual prompt figures is a useful sanity check on its own, since a large gap between them usually means the character-based estimate undercounted something, most often a prompt with a lot of non-English text or unusual formatting.
Settings and options
| Setting or field | What it does | Default |
|---|---|---|
| System Prompt | The full prompt text to test; shows a live estimated token count as you type | Empty |
| Test Question | The question sent alongside the system prompt | "Hello, what can you help me with?" |
| Provider | OpenAI (GPT-4o Mini, 128,000-token window) or Anthropic (Claude Haiku, 200,000-token window) | Whichever provider has a key configured; OpenAI preferred if both do |
| Context used percentage | Actual prompt tokens divided by the selected model's context window | Calculated per test run |
| Verdict message | A plain-language read of whether the prompt fits comfortably, is large, or should be trimmed | Calculated per test run |
Troubleshooting
| Symptom | Likely cause | Fix |
|---|---|---|
| Both provider buttons are greyed out | No OpenAI or Anthropic key is configured for the organisation | Add at least one in Settings, then AI & Flows; the debugger does not test with Groq |
| Run Context Test does nothing | No system prompt was entered, since a blank prompt has nothing to measure | Paste your system prompt into the field before running the test |
| The debugger's reply does not match what a real flow's AI Response node said for the same prompt | The debugger does not include your knowledge base content or conversation history; a real flow turn includes retrieved chunks the debugger never sees | Paste the retrieved knowledge base content into the System Prompt field alongside your instructions to approximate a real turn more closely |
| The page says AI Assistant is not available | The organisation is below the Growth plan and not grandfathered | Upgrade the plan, or confirm grandfathered status if your organisation had a key configured before the gate shipped |
| The token count feels too high for a short-looking prompt | Estimated counts use a rough character-to-token ratio before the real test runs; the actual count reported after Run Context Test is provider-measured and more accurate | Trust the post-run "actual" figures over the pre-run estimate shown while typing |
Related reading
If a bot's system prompt seems too large or is producing inconsistent answers, use this tool before assuming a deeper problem; if the bot is not answering at all, the cause is more likely one of the gates covered in AI on but not replying than a prompt-size issue.