Novelty and superlative-claim guardrail for Claude Code + Grok — the moment a turn asserts a first/fastest/novel/record, it fires the invention discipline so the claim ships proven or scoped, never asserted.
World First is a Claude Code plugin for invention discipline and novelty-claim guardrails. A SessionStart hook primes the posture once per session. A UserPromptSubmit hook fires only when a turn actually makes a first / fastest / novel / record / "nobody has done this" claim — the moment overclaim risk appears — and stays silent on routine work. So the claim ships proven or scoped, never asserted.
Built for AI-agent, research, and benchmarking workflows where a novelty or superlative claim must survive review before it is committed, benchmarked-for-publication, or sent externally.
/plugin marketplace add 88plug/claude-code-plugins
/plugin install world-first@88plug
grok plugin marketplace add 88plug/claude-code-plugins
grok plugin install world-first@88plug --trust
No extra wiring: the hooks and the bundled world-first skill load automatically from the manifest. install.sh is a no-op that says so.
An invention is a measured delta past a confirmed limit, proven novel against a searched corpus, and reproducible by someone who is not you. The plugin makes each clause true before the word "first" is used. It targets one failure mode: the confident-but-wrong "first-ever!" that dies in review.
The UserPromptSubmit hook watches for claim-shape language — superlatives (fastest, best, unmatched), novelty/invention language, records/SOTA, and "beats vLLM/SGLang" comparatives. Detection lives in hooks/lib/detect.py as a broad regex match: false positives cost one line of reminder, a miss lets an unproven "first" ship. When it fires, the turn is asked to (1) name the confirmed measured limit, (2) provenance-search the closest prior art, (3) scope the claim to the measured delta vs the incumbent's tuned baseline, and (4) confirm a fresh-context refuter has tried to break it.
| Hook | Event | What it does |
|---|---|---|
prime.sh |
SessionStart |
Sets the invention-discipline posture once per session |
nudge.sh |
UserPromptSubmit |
Fires only when a novelty/superlative/record claim appears; silent otherwise |
- Confirmed limit first — entry is a measured wall, never a hunch. No confirmed limit, no campaign.
- Removed-constraint test — if the claim beats an incumbent, prove the incumbent structurally cannot express the mechanism. If a lock/GIL-bound engine could express it too, it is not your first.
- Refute, then measure vs a tuned baseline — a fresh-context refuter, the incumbent's best config (not defaults), then ablate which element carries the win.
- Provenance search — no unsearched mechanism is called new. Name the canonical prior art and scope the claim to the measured delta against it.
- Certify on a re-runnable artifact, not prose. Your own re-run never grants
reproduced.
Kills the failure modes: unscoped superlative, reinventing prior art, solving a non-problem, novel-but-perf-neutral, premature-first-before-refute, strawman baseline.
The full playbook is in the bundled world-first skill — invoke it when inventing, claiming a first, or auditing a "first" claim for overclaim:
- TRIZ contradiction → separation (
references/triz-cheat.md) — turn a technical or physical contradiction into an inventive move instead of a compromise; the four separation principles and the high-yield subset of the 40 principles for systems/kernels. - The removed-constraint test (
references/removed-constraint-test.md) — the highest-yield generator of a systems first, plus the dogma-mining move: mine incumbent GitHub issues/PRs/won't-fix/RFCs for admitted walls a lock-free substrate can cross. - The candidate-first template (
references/candidate-first-template.md) — one block per candidate; fill every row or the candidate downgrades.
It composes with scientific-method's invent.md (the falsification→invention engine, provenance protocol, certification ladder) and requires read-the-damn-docs before any API/kernel/FFI code — inventions die on stale-memory API guesses more than on bad ideas.
- break-dogma — a broken dogma is often the incumbent's removed constraint, the wall to invent past.
- scientific-method — the confirmed-limit entry and the measure-vs-tuned-baseline gate.
- read-the-damn-docs — required before any API/kernel/FFI code in the build.
bash tests/smoke.sh # prime always injects; nudge fires on novelty/superlative claims, silent otherwise
python3 .ci/validate_plugin.py . # manifest, keywords, hook paths, skill frontmatterFSL-1.1-ALv2.