After hundreds of hours of experiments, data, and poster work, students standing at their ISEF booth often still feel lost. The judging process involves rotating panels, rules that shift year to year, and conflicting information from every direction. When the awards are announced, they frequently have no clear idea which part of their project fell short.
science-fair-judge is a Claude Code skill built to address exactly this. It is grounded in ISEF's internal judging and scoring documents, layered with field experience, and structured as a simulated panel of three expert reviewers. It evaluates your poster across three dimensions — scientific content, scoring criteria, and interview preparation — and produces a single actionable report.
My background is in physics and engineering, so the skill has some tilt in that direction. Other project categories can use it too.
How It Works
Step 1: Reading Your Poster
The skill accepts three input formats:
- PPTX — analyzes text box positions and reconstructs reading order; does not simply dump raw text
- Images (PNG, JPG, etc.)
Point Claude Code, Codex, or OpenClaw at your poster file path and the skill loads it. You can also open it from the current working directory.
Tested on Claude Opus 4.6, GPT-5.4, and GLM-5.1. Other models may produce different results.
Step 2: Three Experts, Working in Parallel
Expert 1 — Science Content Reviewer
Every verifiable scientific claim on your poster — typically 8 to 15 — is extracted and checked against current literature:
- Is the cited data accurate?
- Is your experimental method standard in this field?
- Are your conclusions overstated?
- Are results "too good to be true" (e.g., R² = 0.999 in biological data is essentially impossible)?
Each claim receives a [Credibility: High / Medium / Low] rating, with supporting or contradicting papers from the last five years.
Expert 2 — Scoring Criteria Evaluator
This expert works from ISEF's official 100-point rubric:
| Category | Points |
|---|---|
| Research question | 10 |
| Design and methodology | 15 |
| Execution and data analysis | 20 |
| Innovation and impact | 20 |
| Presentation (poster + interview) | 35 |
As I have said before, the rubric does not determine awards — but the AI can use each rubric item to identify weaknesses. This expert also checks ISEF's hard physical requirements (board dimensions, prohibited items, chart standards) and flags missing elements with priority tags: Urgent / Important / Minor.
Most importantly, it renders a judgment: would a judge be willing to advocate for this project in the caucus? That question matters more than any numerical score.
Expert 3 — Interview Question Preparer
Beyond generic questions (motivation, obstacles), this expert prepares four categories of specific questions:
- Understanding and depth — do you actually understand what you did? (5–7 questions)
- Critical analysis — what are the limitations? what if the results were reversed? (4–5 questions)
- Ownership and independence — did you do this or did your mentor? (3–4 questions)
- Impact and future direction — what does this mean? (3–4 questions)
Each question includes: why judges ask it; what a strong answer looks like; what a weak answer looks like; and red-flag responses that signal you did not do the work.
Step 3: Cross-Validation
After each expert completes their section, the three reports are reconciled. If the content reviewer flags a data problem, the scoring evaluator's deductions adjust to match. The interview preparer targets questions at the weaknesses the scoring evaluator identified. The final output is an integrated whole, not three isolated opinions.
Step 4: The Report
A complete Markdown report is generated in your project directory, containing:
- Science claim verification table
- Detailed rubric scoring
- Missing elements with priority labels
- Full interview question preparation booklet
- Priority action list — split into "must fix before the fair" and "must be able to answer in the interview"
The Thing Most People Do Not Know About ISEF Scoring
Every judge submits a score from 0 to 100. That score goes through two rounds of normalization and determines initial ranking. Then, before final awards are decided, all projects appear ranked on a large screen — and the real decision happens in the caucus.
Judges sit together and advocate for the projects they believe deserve awards. Final placements are determined by majority vote in that room.
This means: if no judge advocates for your project in the caucus, even a high score may not produce an award. A judge advocates for a project because, in 10–12 minutes of conversation at your booth, they genuinely understood what you did and were moved by it. That is why this skill weights interview preparation at 35 points — the highest single category — and treats the caucus question as central.
Usage
/science-fair-judge
Or with an explicit path:
/science-fair-judge path/to/my-poster.pptx
| Context | Support level |
|---|---|
| ISEF | Full support — official rubric built in |
| Regional / state science fairs | Searches for current criteria and adapts |
| Regeneron STS | Partial — will search for STS criteria but output format may not suit; a dedicated STS version is planned |
Practical Notes
Use it 2–4 weeks before the fair — enough lead time to revise the poster and practice interview answers.
Do not trust the output blindly. Most of it will be accurate, but misconfiguration or partial reads do happen. My own setup runs Claude Code with the Oh-My-Claudecode plugin. GPT-5.4 tends to be more granular; Opus and GLM are faster.
Prioritize items tagged Urgent. Missing error bars, absent references, conclusions that do not address the hypothesis — these are signals of inexperience that judges notice immediately.
The skill does not do your research. It tells you whether your research holds up, whether your presentation meets the standard, and what you will be asked about in the room.
Source: github.com/tian2-ai/Tian2-skill — CC BY-NC 4.0