v1.3.0: Catch AI mistakes before they cost weeks of compute
From paper-writing helper to ML-training discipline plugin in six PRs.
Six of the seven war stories in the README are now actively prevented.
The headliner
A second Claude reviews the first one's conclusions before they reach you.
The Stop hook is now an agent, not a static prompt (#5). Every time the assistant decides it is "done," a fresh-context Sonnet-4.6 spins up in its own session, reads the actual artifacts, and pushes back if the conclusions do not survive contact with reality. 300-second budget, isolated context, zero shared state with the run that produced the work.
Four new skills, zero config
Just describe what you want, they auto-trigger (#7).
| You say | What runs |
|---|---|
| "reproduce this arxiv paper" | Reproduce: 7-stage pipeline from arxiv URL to replication runs |
| "why is my loss diverging" | Debug: evidence-first probes, no guess loop |
| "compare run A to baseline" | Compare: same-epoch alignment across wandb / neptune / tb / mlflow |
| "launch the new training run" | Launch: pre-flight checklist before burning 50 hours of GPU |
Six new silent guardrails
You never invoke these. They catch the assistant being wrong (#7).
- destructive_path_guard: fabricated paths in
rm/mv/dd/ force-push - long_job_launch_inject: pre-flight before multi-hour ML launches
- outbound_artifact_reminder: unverified commands in teammate messages
- remote_inplace_edit_guard: in-place SSH edits to git-tracked files
- jargon_scrub: sprint-codes leaking into public commits
- timezone_scrub: timezone tokens drifting from the system clock
Fixes & polish
- Stop hook infinite loop on bare acknowledgments, fixed by @ccyrene (#8)
- All hook scripts migrated to the
hookSpecificOutputschema (#9) - Star history chart and support acknowledgment in the README (#6)
- README compacted 8% with no loss of UX or USPs (#10)
Install
claude plugin marketplace add fcakyon/phd-skills
claude plugin install phd-skills@phd-skillsZero MCP. Zero config. Works the moment it is installed.
Contributors
Full changelog: v1.1.0...v1.3.0
