Skip to content

Fluency

A score per competency, per person. Not vibes — built from a static catalogue of 11 chat-observable indicators applied to every Enterprise-pathway session in a rolling 30-day window.

This is the 4D fluency framework — four habits that show up in real usage. In the app each habit carries a plain-language name.

CompetencyIn-app habitWhat it measures
DescriptionClear promptsHow clearly the prompt frames the task — context, constraints, acceptance criteria
DelegationRight task to AIHow well the engineer hands off versus over-specifying or babysitting the model
DiscernmentChecks the outputHow critically the engineer evaluates the response before accepting it
DiligenceVerifies before shippingHow reliably the engineer tests or cites the load-bearing claims before shipping

Where fluency surfaces

There is no standalone org fluency page. Scores surface in two places:

My AI → Overview — each person's own view. A habit score per competency, a plain-English verdict ("You work with AI fluently / well / reasonably / cautiously"), and a tip against the weakest habit. This is the personal, self-service surface at /me/ai-agents.

Agent insights drill-downs — a person's fluency rollup appears on their detail drawer when you drill into them from Insights → Agent insights. This is what an engineering manager opens before a 1:1 to spot whether someone's delegation has shifted.

How a session is scored

Every Enterprise-pathway session that gets ingested is run through the classifier. The classifier produces:

  • A boolean per indicator — present or absent in this conversation.
  • A summary block with rationale (used in the per-session drilldown).

Org-level prevalence is then presentCount / conversationCount per indicator. Competency scores are aggregated across their indicators with a per-indicator weight from the catalogue.

Limits of the model

Worth being honest:

  • Telemetry pathway orgs see no Fluency data — no prompt body, nothing to classify. Pick the Enterprise pathway if you want this.
  • The classifier evaluates patterns in observable chat. It doesn't know whether the resulting code shipped, passed review, or caused an incident.
  • Indicators are catalogue items, not custom prompts. We'd rather give you a stable benchmark than let every org drift its own.

Coding sessions only, today. The 4D framework and its indicator catalogue are calibrated against coding conversations — the indicators look for things like "specifies acceptance criteria" or "asks a clarifying question after a vague response," which read clean against an engineering session and noisier against, say, a marketing-copy or image-prompt session. Modality-aware scoring for image, video, audio and chat is on the roadmap. Until then, Fluency is silent on non-coding traffic — we'd rather show no score than a misleading one. See Catalog for what coverage your org has across modalities today.

What you do from here

  • Spot training opportunities. A team with low Description prevalence usually benefits from a prompt-quality workshop.
  • Catch model regressions. A sudden drop in Discernment across the org often correlates with a model release that confidently hallucinates more than its predecessor.
  • Calibrate hiring. Compare new joiners' first-30-day Fluency to the org baseline.

Empty state

If no Enterprise-pathway sessions have been captured yet, the fluency tiles show "not enough analysed sessions yet". Enable the Enterprise pathway in Settings → AI → Agent Policy.

Flowstate Documentation