Create a guided first verified task and installation diagnostics #11

Workflow status: Done — merged in #18 and #29; walkthrough run end to end on 2026-10-31
Significance: Most significant
Effort: Medium
Release: 3.3.0 / 3.4.0

Problem

Installing standing policy does not itself demonstrate that a new user can complete a verified task or recover from setup problems.

Done when

• Diagnostics report active scope, policy revision, agent discovery/names, and conflicts without exposing secrets. (Add install diagnostics, a safe uninstall, and a first verified task #18, tools/diagnose.py)
• A documented example shows fail → correct → pass with actual command output. (Add install diagnostics, a safe uninstall, and a first verified task #18, docs/first-task.md)
• A new user can follow the walkthrough from install to verified result. (Warn about install permission prompts, from a real walkthrough run #29: run in Claude Code 2.1.316 against the published 3.4.0 plugin. Install from GitHub, the gated fix (the verifier subagent returned GATE: PASS), and the uninstall all worked. The one gap found, permission prompts during install, is now documented.)
• Cleanup preserves unrelated configuration and files. (Add install diagnostics, a safe uninstall, and a first verified task #18, tools/uninstall.py; confirmed in the run)
• Documentation separates demonstration from measured production benefits. (Add install diagnostics, a safe uninstall, and a first verified task #18)

Note from the run

In a session that blocks writes to CLAUDE.md, the skill installs nothing and reports every check as FAIL instead of working around the block. For that run, the scope was written with exactly what the installer writes (confirmed by check_install.py), so the rest of the walkthrough could be exercised.

Sources

Research H3 and H6.