Skip to main content

AI analysis modes

The platform uses AI to analyse coaching session transcripts against ICF competency rubrics. You upload or record a session, the platform sends the transcript to your chosen AI provider along with a system prompt containing the rubric, and the model returns competency scores with specific transcript evidence.

There are two modes: standard and deep analysis.

Standard mode​

Standard mode sends the transcript and rubric to the AI and gets the assessment back directly. Critical-evaluation guidance is included in the prompt so the model still produces differentiated ratings with specific transcript evidence rather than generic positive feedback. This is the default for local AI providers (Ollama and LM Studio).

Deep analysis​

Deep analysis adds a reasoning step. The model thinks through the transcript methodically before writing the assessment. This generally produces more thorough evaluations, particularly for PCC and MCC frameworks with many markers.

Deep analysis is available with:

  • Cloud providers that control their reasoning budget (Anthropic, OpenAI, xAI and others)
  • llama-server which enforces a reasoning budget at the sampling pipeline level via --reasoning-budget

Why local models run in standard mode​

After extensive testing with Ollama and LM Studio, I found that local models cannot control how long they spend thinking.

The model has a single output token pool shared between thinking and content. There is no API mechanism to say "stop thinking after N tokens." In testing, models consistently used 40,000 to 48,000 tokens on reasoning against a target of 5,000 to 7,000, leaving almost nothing for the actual assessment.

The result: 30-minute waits that always fell back to standard mode anyway.

I decided to make standard mode the default for these providers rather than have users wait half an hour for the same outcome. If you want to override this and enable thinking for Ollama or LM Studio, you still can in Advanced tuning, but be aware that assessments may take 30+ minutes and are likely to fall back to standard mode.

Provider capability tiers​

What this means in practice​

Standard mode is not lesser. The AI still evaluates each competency marker against specific transcript moments. The difference is that deep analysis lets the model reason through ambiguous evidence before committing to a score, which can matter for borderline markers.

For most coaching sessions, standard mode produces good, actionable feedback. If you want the extra reasoning step with a local model, llama-server is the path forward.

Future​

I am monitoring Ollama issue #10925 and LM Studio issue #988. When these providers add API-level budget control, thinking will be re-enabled by default.