Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
claude-papercuts
Ten Claude Code skills that fix bugs Anthropic closed as "not planned."
Anthropic's claude-code issue tracker has dozens of high-engagement bug reports closed with not planned. This plugin ships one skill per papercut. Each skill cites the GitHub issue it fixes — and gets deprecated the day Anthropic actually fixes it.
No telemetry. This plugin makes zero network requests.
Install
For now (pre-marketplace), clone and load locally:
git clone https://github.com/dhruba-datta/claude-papercuts ~/claude-papercuts
claude --plugin-dir ~/claude-papercutsOnce a marketplace is set up (planned):
/plugin marketplace add dhruba-datta/claude-papercuts
/plugin install claude-papercutsSkills in this plugin are namespaced. Invoke as /claude-papercuts:<skill>.
The papercuts
| # | Skill | Fixes |
|---|---|---|
| 1 | unclear | #39975 — /clear has no undo |
| 2 | done-prover | #5052, #10628, #20350 — Claude lies about completion |
| 3 | skill-budget | #30387, #34648, #16575 — skills silently vanish past the ~15K char budget |
| 4 | amnesia-fix | #14227, #27298, #43696 — no cross-session memory |
| 5 | token-x-ray | #39686 — auto-injected plugins waste 6k+ tokens silently |
| 6 | compact-guard | #24686, #26061 — plan-mode state lost on compact |
| 7 | safe-shell | (UpGuard/ClaudeLog postmortems) — YOLO rm -rf incidents |
| 8 | onboard | (Medium/MindStudio reports) — new users install too many skills and churn |
| 9 | skill-doctor | Root cause of #30387 — descriptions overlap with training |
| 10 | subagent-broker | #4182, #5528, #19077 — subagent delegation broken |
Full per-issue details in docs/issues.md. One new papercut drops on the 1st of every month.
What this is, in one paragraph
Most Claude Code "awesome skills" repos are append-only lists of 1,000+ SaaS wrappers. This one is the opposite — ten skills, each one a receipt for a specific issue Anthropic decided not to fix. Every README starts with the issue number. When Anthropic eventually ships a real fix, the corresponding skill becomes a no-op and gets deprecated with a date. We read the closed-as-not-planned issues so you don't have to.
Voice & conventions
- Trust, then verify. Every claim Claude makes that something is
"done" gets re-verified by the relevant skill.
- Refuse before warning. Risky commands block by default, not
prompt-and-continue.
- No emoji in commands. Read the docs, run the command, see the
output. No theatrics.
Testing
bash tests/run-all.sh271 tests covering plugin schema, all Stop hooks' behavior (snapshot pruning, claim detection, journal append/load), audit.py discovery + budget logic, regex extraction (bulleted and inline), unicode/control-char fidelity, concurrent invocations, and shellcheck on the shell scripts. CI runs the same suite on every push and PR.
Contributing
The monthly papercut drops on the 1st of each month. To propose the next one, open an issue with:
- A link to the Anthropic GitHub issue it would fix (must be a real
number, not a vibe)
- The reproducer steps
- A one-paragraph sketch of the skill that would address it
If the issue has been open for >60 days or closed-as-not-planned with >20 reactions, it's a strong candidate.
License
MIT. See LICENSE.
Acknowledgments
community whose issues are this project's bibliography
obra/superpowersfor the
methodology-as-skill framing