The AI Fluency Framework.

The four-pass pipeline.

Every score on your fluency profile comes out of the same four passes, in this order. Each pass writes to its own table, so a failure in pass three doesn't corrupt the work passes one and two already did.

1. Pass 1 · Upload & redact

   You drop a ZIP export from Claude or ChatGPT. We parse each conversation, strip secrets, emails, tokens, and obvious PII, and stage one row per conversation. The redactor runs before anything else touches the text. The raw export is never persisted.

2. Pass 2 · Per-conversation extraction

   Each redacted conversation goes to an LLM that returns one structured signal: dimension scores (0–10), outcome, topic, techniques detected, failure modes, and an optional exemplar excerpt. Conversations are processed independently so partial failures don't block the rest.

3. Pass 3 · Aggregation

   A second LLM rolls every extracted signal into one user-level aggregation: overall scores, deltas, confidence intervals, monthly progression, biggest gap. This is the row the credential reads from.

4. Pass 4 · Synthesis

   A final pass picks the most informative exemplars per dimension and writes the human-language descriptors that show up on your profile (“what you do well,” “where you stretch”). Synthesis fails open. You still get scores even when descriptors can’t be written.

90-day retention.

Raw conversation text is deleted automatically 90 days after the conversation was started, on a rolling basis. The signals extracted from each conversation, your aggregated scores, and the exemplar excerpts you've flagged all persist past that window: only the raw bodies are dropped. You can also delete the raw text at any time from Settings → AI Fluency Data.

The practical effect: your scores never disappear, but the original conversation text behind a six-month-old score is gone. If you want to re-score against a newer rubric, you upload a fresh export.

How this is selected.

Every prompt that appears on your AI fluency profile, as a worked example, as a trajectory pivot, as evidence under a dimension, came through the same four-stage pipeline. It's deliberately conservative: most of your prompts never appear on the page, and that's the point. We want to surface the few that actually demonstrate something, not dilute the signal by showing everything.

1. Stage 1 · Eligibility

   We start with every human turn in the conversations you uploaded. A turn is eligible if it's a real prompt: substantive, not a one-word reply, not a system or assistant message. Anything that doesn't clear the eligibility bar is dropped before the analyzer ever sees it.

2. Stage 2 · Prefilter

   Eligible turns are scored cheaply with four boolean heuristics: does the prompt break work into subtasks, does it reference an earlier turn, does it explicitly ask the model to verify, and does it follow a failed model response. Each signal moves the prefilter score; only turns above a fixed threshold proceed to the next stage. This is the gate that keeps the cost of analysis bounded. Most prompts never need the LLM.

3. Stage 3 · LLM evaluation

   Prompts that clear the prefilter are evaluated by the analyzer LLM against a versioned rubric. The evaluator returns a structured diagnostic: what the prompt did well, where it broke down, and the root cause of any pivot it can identify in the surrounding turns. Confident "nothing notable" verdicts are stored as nulls and never shown.

4. Stage 4 · Selection for display

   Worked examples and trajectory cards are picked from the evaluated set with two goals: cover the five dimensions evenly, and show the most informative behavior for each. We prefer recent prompts, prefer ones with clear root causes, and explicitly de-duplicate near-identical prompts so the same pattern doesn't dominate the surface. Everything you see on your analysis page came through these four stages.

If a prompt you expected to see is missing, it's almost always because it was filtered out at stage 2 or stage 4. Either the prefilter judged it ordinary, or the selector preferred a similar prompt that covered the same behavior more clearly.