Understanding Your Attention Patterns

Understanding Your Attention Patterns

report attention analyzes how your time and focus move across projects when working with AI assistants. It measures active time, context switches, flow episodes, and concurrent-session orchestration — all from your local Claude Code logs.

Quick version

conscience report attention --days 7 --html attention.html

Unlike the other reports, attention covers every project by default, because switching between projects is what it measures. --project <path> narrows it to one.

What it measures

Metric	What it means
Active time	Estimated minutes you were engaged with AI sessions, per project and per day. Shown as a range reflecting attribution uncertainty.
Context switches	How often you moved between projects within an active window (a gap > 15 min is a resumption, not a switch).
Dwell	How long you stayed on one project before switching — median, mean, max.
Flow episodes	Sustained runs of engagement (default: 20+ min with gaps ≤ 10 min). Tagged as single- or multi-project. The waiting-on-AI rule prevents long agent runs from breaking genuine flow.
Orchestration	How much wall-clock time had 2+ AI sessions working concurrently, and how much of that overlap you were actively attending.

Reading the output

Per-project table shows active time as a range (e.g., "14–16 min") because attribution between consecutive prompts on different projects is genuinely ambiguous — the range's width is the honesty.

Flow episodes are tagged single-project or multi-project. If most of your flow spans multiple projects, that's a data point for reflection — does cross-project flow feel like richness or fragmentation?

Orchestration distinguishes "raw" overlap (any concurrent AI work) from "attended" overlap (concurrent AI work while you were actively prompting). The difference is time you kicked off agents and walked away.

The output ends with a reflection question generated from your data — conscience's anti-Goodhart mechanism. The numbers aren't a score to optimize; they're context for discernment.

HTML timeline

conscience report attention --days 14 --html timeline.html
open timeline.html

Produces a self-contained HTML file (no external dependencies) with:

• One section per day
• Swimlane per project showing touchpoint ticks and AI-working spans
• Theme-aware (works in light and dark mode)

Configuring thresholds

What counts as "idle" and "flow" is personal. Configure in conscience.yaml:

thresholds:
  attention:
    idle_minutes: 15            # gaps longer than this are idle
    engagement_floor_minutes: 2 # credit for an isolated prompt
    flow_gap_minutes: 10        # max gap inside a flow episode
    flow_min_minutes: 20        # minimum span to count as flow
    project_aliases:
      "/private/tmp/worktrees/*": my-project

project_aliases folds worktrees and scratchpad sessions into their logical project so they don't pollute your switch counts.

Honest-measurement caveats

• Active time counts only Claude Code interactions — reading, meetings, and non-AI coding are invisible. Totals are floors, not truths.
• Per-project attribution is a range because the heuristic's bias direction is unknowable per gap.
• Flow inference is a hypothesis generator for your own reflection, not a judgment.

Related

• Command Reference — full report attention flags
• Configuration — thresholds and project aliases
• Measuring Energy Cost — energy consumed during those sessions