dhruba-datta onboard — first-run walkthrough
Fixes: new-user churn documented in MindStudio's Claude Code onboarding analysis and Medium "I installed 40 skills and Claude got worse" reports — users install too many skills at once, can't tell what's working, and bounce.
What this prevents
"I followed a tutorial, installed 25 skills, and now Claude is noticeably slower and ignores half of them. I uninstalled everything." — common Medium / Reddit complaint, Dec 2025
The claude-papercuts plugin ships nine skills. Most users don't need
all nine on day one. onboard is the curated install order: which
skill solves which problem, in priority order, with a one-sentence
"why" each.
What it does
Runs the bundled recommend.py, which:
- Detects which of the nine papercut skills are already installed
- Picks the next one to enable based on order
- Shows the install commands
How to invoke (the actual procedure)
Run the recommend script:
${CLAUDE_PLUGIN_ROOT}/skills/onboard/recommend.pyIf the user asked something specific (e.g. "what's the next one?"), forward the matching flag:
${CLAUDE_PLUGIN_ROOT}/skills/onboard/recommend.py --nextShow the script's output verbatim. Then ask the user:
- Which workflows do they actually do? (sessions/day, /compact use, run YOLO mode, ship code Claude wrote, author skills, etc.)
- Based on their answers, suggest the next 1–3 from the ordered list that match their workflow.
Never auto-install. The user runs the install command themselves — we're advisors, not installers.
The opinionated order (and why)
| # | Skill | When to enable |
|---|---|---|
| 1 | safe-shell |
Always first — irreversible-command refusal, no downside |
| 2 | token-x-ray |
Visibility into what's eating your context — informs all other decisions |
| 3 | amnesia-fix |
If you run >1 Claude session per day |
| 4 | compact-guard |
If you ever hit /compact (most users will) |
| 5 | done-prover |
If you ship code Claude writes (high-impact, mildly noisy) |
| 6 | unclear |
If you've ever run /clear and regretted it |
| 7 | skill-doctor |
If you author your own SKILL.md files |
| 8 | skill-budget |
If you have 10+ skills installed |
| 9 | subagent-broker |
Power-user — multi-stage agentic delegation |
When to auto-invoke
- User runs
/claude-papercuts:onboard - User asks what claude-papercuts does
- User asks which skill to enable first / how to start
- User says they're new to the plugin
- User asks for the install order or "what should I turn on?"
What this skill does NOT do
- It does not install skills. The user runs install commands themselves.
- It does not modify settings. No auto-config.
- It does not interview the user with a fixed quiz. You (Claude) tailor the recommendation to their actual workflow after running the script.
- It does not recommend skills outside the nine papercuts. For broader ecosystem recommendations, point users at https://github.com/dhruba-datta/claude-papercuts/discussions.
Configuration
# Full walkthrough
${CLAUDE_PLUGIN_ROOT}/skills/onboard/recommend.py
# Just the next skill to enable
${CLAUDE_PLUGIN_ROOT}/skills/onboard/recommend.py --next
# Machine-readable JSON
${CLAUDE_PLUGIN_ROOT}/skills/onboard/recommend.py --json
Deprecation plan
If the claude-papercuts plugin grows past nine skills, the
recommendation list needs to be re-curated. This file's SKILLS list
is the source of truth.