Skip to content

CLAUDE.md

Latest commit

 

History

History
72 lines (53 loc) · 4.71 KB

CLAUDE.md

File metadata and controls

72 lines (53 loc) · 4.71 KB

ClickSense

Mousedown→mouseup latency as cognitive load / click confidence signal.

Origin

Andy Edmonds (2016) Bloomreach "Big Brains" talk — demonstrated that click hold duration correlates with emotional/identity engagement, not task difficulty. Political questions about Obama/Trump produced 110-125ms holds; obscure factual recall produced 75ms. The core range across n=200 crowdsourced participants was 50-120ms.

Reference: https://youtu.be/j38fm48gTgg?t=1348

Architecture

  • Core (src/clicksense.js): Vendor-agnostic mousedown/mouseup/touch timing. Filters drags, programmatic clicks, and press-and-holds. Walks DOM to find nearest meaningful target (link, button, role="button"). Supports data-clicksense attribute for explicit labeling.

  • Adapters: Thin wrappers for analytics backends.

    • adapters/posthog.js — flattens to PostHog properties
    • adapters/callback.js — buffers + flushes (for sendBeacon, Adobe Analytics, etc.)

  • Build: esbuild → IIFE (script tag), ESM (import), CJS (test/SSR)

Integration

Currently deployed on:

  • scrutinizer-www (PostHog, script tag)

Planned:

  • psychodeli-webgl-port
  • iblipper2025

Key design decisions

  • Capture phase listeners (see events before stopPropagation)
  • performance.now() for sub-ms timing (not Date.now())
  • Drag distance filter (>10px = drag, not click)
  • Touch support (touchstart/touchend) for mobile
  • data-clicksense="label" attribute for explicit target naming
  • No PostHog dependency in core — adapter pattern only

Analysis

  • PostHog SQL queries: queries/analysis.sql — 16 HogQL queries (data inventory, duration distributions, per-user baselines, target breakdown, session drift, approach dynamics, touch vs mouse, hyperlink/any-target deltas from baseline).
  • Dashboards:
    • Scrutinizer project (258589): Hold Duration × Approach Dynamics (117 events, Mar 2026 baseline).
    • Psychodeli project: primary production dataset (n=2,917 as of 2026-04-17 — 25× the Scrutinizer sample).
  • Key findings (Psychodeli, 2026-04-17, n=2,917):
    • Deceleration monotonic across hold buckets (-0.0012 → -0.0074). Holds and steepens vs Mar.
    • Velocity monotonic with hold bucket (0.62 → 1.02 mean, 0.35 → 0.52 median). New.
    • Distance monotonic with hold bucket (200 → 239 px). New.
    • Corrections peak in "normal" bucket (12.23). Holds.
    • Pre-click pause is flat at 123-128 ms for ballistic/normal/deliberative; only "extended" (>160 ms) jumps to 189 ms. The Mar "316 ms deliberative pause, 2.4× other buckets" finding does NOT replicate at n=2,917.
  • Property names (v0.1 core): duration_ms, click_x, click_y, drag_distance, input_type, target_tag, target_id, target_label, target_classes, target_href, target_text, approach_velocity_mean, approach_velocity_final, approach_deceleration, approach_corrections, approach_distance, approach_pause_ms, approach_linearity, approach_max_deviation, approach_trajectory_type.
  • Property names (v0.2 autoLabel, 2026-04-17): target_aria_label, target_title, target_name_attr, target_value, target_placeholder, target_name (computed accessible name), target_path (short CSS selector, depth configurable), plus target_data_<key> for every data-* attribute on the element (e.g. target_data_preset, target_data_mode). Default pathDepth: 3.

Research Data

  • Structured findings: docs/data/research-findings.json — 26 extracted numeric findings from cited papers and production data. Each entry has source citation, values, unit, and optional chart spec. Consumable by ascii-charts/chart.py --json.
  • CSV exports: queries/exports/2026-03-28/ — 5,201 events from 550 users (first data pull, Mar 2026).
  • Deep-dive docs: docs/deep-dive-*.md — 6 research summaries (Fitts's law, dwell time, cognitive load, implicit signals, related paradigms). The research-findings.json extracts the chartable numbers from these.

Cross-Project Research

This project is part of a connected research program:

  • Research log: ~/Documents/dev/research-log.jsonl — auto-captured WebFetch/WebSearch URLs across all Claude Code sessions. Query ClickSense-relevant entries: jq 'select(.project=="clicksense")' ~/Documents/dev/research-log.jsonl
  • interests2025 pipeline: Research URLs are ingested into Qdrant (node src/ingest_research_log.js --index) for semantic search across all projects.
  • Related projects: Scrutinizer (visual perception), Psychodeli+ (rendering), histospire (browsing behavior)
  • Session history: Each research log entry includes transcript_path for tracing back to the Claude Code conversation where the research occurred.