SCORE IT YOURSELF
The skills behind the talk
Every number in the talk was checked by tooling before it went on a slide, and the reviewer that invented two papers was caught by one of these. They are Claude Code skills: a folder with a SKILL.md that tells the session how to do a job, sometimes with a small script beside it. Anyone who uses Claude Code can install them. They were written for my own setup, so the names, paths and habits in them are mine; change what does not fit.
Install
Download wwc2026-skills.zip and unzip it into ~/.claude/skills/, so that you end up with ~/.claude/skills/content-matter-expert/SKILL.md and so on. Start a new Claude Code session and the skills are available by name. Two of them need a key or credentials in the environment, noted below.
The five
- Content-matter expert content-matter-expert/
Becomes a source-backed expert on a topic before judging a draft: parallel research passes that never see the draft, a synthesis, then an adversarial verifier that fetches every load-bearing citation and marks it keep, weaken or cut. This is the skill that caught the two papers that do not exist. Needs a Claude Code session with web access and subagents.
- Debate with GPT debate-gpt-brad/
A structured argument between the Claude session and an OpenAI model, up to five rounds, stopping when two rounds in a row produce nothing new, then a written synthesis. Six modes, including a Red-Team pass that attacks whatever the two models agree on. The bridge script reads OPENAI_API_KEY.
- Debate with a model on Bedrock debate-bedrock-brad/
The same debate, with the adversary on AWS Bedrock so it comes from a different model family and runs on a company account with no personal key. Default adversary is gpt-oss-120b; pin another with BEDROCK_BRAD_MODEL. Needs AWS credentials in the environment.
- Peer-review panel peer-review-panel/
Four referees with distinct archetypes, a gate, a rebuttal round and a meta-review, for anything that will face real reviewers: an abstract, a manuscript, a study design. The references folder holds the archetypes, the rubric and the venue norms.
SKILL.md · references/panel-mechanics.md · references/reviewer-archetypes.md · references/review-rubric.md · references/venue-norms.md · references/anti-patterns.md · references/red-baseline-findings.md
- Voice pass captain-hook/
Rewrites a human-facing draft so it reads as competent human writing in the author's own register, strips the machine tells, and red-teams the result with a second model. The substance stays the author's; the skill only touches how it sounds.
How they were used on this talk
The abstract went through the peer-review panel and the GPT debate. The script's claims went through the content-matter expert, which is where the invented citations were caught. The business-card headline and the new slides were argued over with both debates, including the Red-Team pass. The handouts and the bio got the voice pass last. The two articles under the talk page, Build Your Own Agent Skills and Put Your AI's Sparring Partner on Bedrock, walk through how the debate skills were built and why the adversary sits on a different model family.
The rule that matters more than any of the files: nothing a model cites gets used until a person or a script has fetched the source and read it. The skills make that rule cheap to follow. They do not make it unnecessary.