Skip to content

High-performance browser automation bridge and multi-instance orchestrator with advanced stealth injection and real-time dashboard.

1.8 KB
Raw Blame

PinchTab onboarding / install-UX persona lane

Agent-driven tests of the new-user onboarding experience. Each scenario pairs a user persona (skill level + intent) with a starting condition and a first-step goal, then a blind agent simulates that user reaching the goal using ONLY the user-facing docs, in a fresh container. We collect friction and turn it into fixes.

This is NOT a CI unit test — it needs Docker plus an LLM agent runtime (e.g. Claude subagents) to play the personas. The scenario definitions, container provisioning, and agent brief here make it repeatable.

How to run

# 1. Provision one clean container per persona (each with its starting condition)
tests/install-ux/setup.sh

# 2. For each persona, launch a blind agent with:
#      - tests/install-ux/subagent-context.md  (shared rules + the persona's brief)
#      - the persona's container name (e.g. ptux-sa)
#    The agent operates the container via `docker exec <name> bash -lc '...'`,
#    reads ONLY user-facing docs (README.md, skills/pinchtab/SKILL.md, docs/**),
#    and returns a friction log.

# 3. Tear down
docker rm -f ptux-sa ptux-sb ptux-sc ptux-sd ptux-se ptux-sf

Files

  • PERSONAS.md — the persona × condition × goal matrix.
  • setup.sh — provisions the per-persona containers with their starting conditions.
  • subagent-context.md — shared blind-agent rules + the per-persona briefs to hand each agent.

What good looks like

Every persona should reach its goal, and — critically — recover from each obstacle using the tool's OWN output/docs (no guessing, no source-reading). The friction log scores each step yes/partial/no on "could proceed from the messaging alone." See the repo's commit history for fixes that came out of earlier runs (no-browser guidance, IDPI remediation, session auto-start, config --help discoverability, etc.).