Design Judge Skills: evidence-driven Agent Skills for design award submissions
Evidence-driven Agent Skills for design award research, evaluation, award matching, entry writing, and submission readiness.
At a glance
- What is it?
- A set of six installable Agent Skills that split design award work into search, evaluation, award matching, entry writing and submission checks. It ships a 22,125-record observation set and a benchmark table, but every deadline and fee still has to be re-verified against official pages at runtime.
- Who is it for?
- Adopt it if you already run Claude Code, Codex, OpenClaw, OpenCode or Hermes Agent and you want award research broken into modules you can trigger one at a time, with facts, inferences and open questions kept apart. Do not adopt it if you expect a winner predictor: the README states that the 22,125-record observation set is descriptive background and cannot be used to estimate award probability, and the skills explicitly avoid simulating unpublished jury preferences.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 27 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 19, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The gap Design Judge Skills fills between a design portfolio and an award entry
Award submission is not one task. It is a search problem (what won in this category), a judgement problem (how good is this work and how well is it shown), a routing problem (which award, which track, which category), a writing problem (what goes in each field, within a word limit), and a compliance problem (does the package match this year's rules). Most teams do all five in a shared folder and a chat thread, and the reasoning behind each decision disappears.
This repository turns those five into separately triggerable modules. The README describes the scope as award case retrieval, design evaluation, award matching, entry text preparation and submission readiness, and the skill index lists six skills: design-award-pipeline, design-award-search, design-evaluation, design-award-match, design-information-prep and design-submission-check. The audience is narrow and specific: designers and studios that already work inside a coding agent and want the award process to leave a trace. The stated configuration covers iF DESIGN AWARD, iF DESIGN STUDENT AWARD, Red Dot Product Design, Red Dot Design Concept, IDEA, DIA, K-Design, GOOD DESIGN AWARD Japan, Core77, James Dyson and EPDA.
How the six skills hand work to each other
The README prints the workflow as a diagram: design-award-pipeline sits at the top for requests that are uncertain or span stages, and it branches to design-award-search, design-evaluation and design-award-match. The match step feeds design-information-prep, which feeds design-submission-check. The modules are described as usable independently, and the pipeline skill is said to select only the necessary modules rather than forcing the full chain.
The design constraint that matters most is stated as a rule, not a feature: facts from the user's material, model inferences, and items awaiting user confirmation must be labelled separately. That is what makes the output auditable. It also means the skills will hand work back to you instead of filling a gap. If a field needs a fact that is not in your attachments, the information-prep skill is supposed to list the missing fact first rather than write around it.
The second rule is source priority. Award rules and cases are to be taken from official pages; search snippets and third-party pages are only for finding leads. The README says that where deadlines, fees, eligibility, categories and submission specifications are concerned, the skills require re-verification of the official page at runtime. Nothing about the current cycle is cached as truth, which is the right call for award rules that change annually, but it also means the skills are slower and more dependent on live web access than a static checklist.
Installing Design Judge Skills with npx skills and running a first evaluation
Every directory under skills/ is an installable unit. design-judge-shared is a shared support package, included automatically in a full install, and the README asks you to install it alongside design-award-search or design-award-match when you install those alone. The npx route needs Node.js 18 or higher and no global CLI.
List what the repository offers before installing anything:
npx skills add SeanJ1ang/design-judge-skills --listInstall everything globally into Codex, including the shared package:
npx skills add SeanJ1ang/design-judge-skills --global --agent codex --skill '*' --yes --copyFor Claude Code the same command takes --agent claude-code. For a single skill scoped to the current project, the README gives this form:
npx skills add SeanJ1ang/design-judge-skills --agent codex --skill design-evaluation --yes --copyOther agents supported by the CLI take repeated --agent flags, for example --agent openclaw --agent opencode --agent hermes-agent, and --all installs to every agent the CLI supports. To check what is installed and to update:
npx skills list --global --agent codex
npx skills update --global --yesStart a new agent session after installing or updating so the skill list refreshes. Then describe the task in plain language, or name the skill. The README's own example for a student concept is: use $design-evaluation, state that you are fixing the maturity level as a student concept, and ask for design quality, presentation quality, evidence confidence and Critical issues as separate outputs. The point of that prompt is the maturity label. You supply it, so the same portfolio can be scored under different assumptions instead of the skill silently guessing how developed the work is.
The benchmark table and what the 22,125 observation records actually do
The README reports a test in which all 1,612 entries of K-Design 2026 were collected and scored on 2026-07-03, before the winners were announced, so the scores could not be adjusted to match the result. The table compares base models against the same models with the evaluation skill applied at Top-115: GPT-5.5 goes from 32 hits to 41, GPT-5.4 from 33 to 47, and Claude Sonnet 4.6 from 19 to 44.
Those are the project's own numbers, reported by its author, and they come from one award cycle. The README does not describe how the 115 threshold was chosen, and it does not report a holdout on a second competition. Treat the table as a claim about one dataset, not as a general accuracy figure.
The evaluation module also carries 22,125 aggregated observation records drawn from iF, iF Student, Red Dot and IDEA winners or shortlisted works. The README is unusually direct about their role: they provide descriptive background, they do not change the core score, and they cannot be used to estimate award probability. That is the honest framing, and it is also the limit. A large corpus of past winners will not tell you whether your project wins; it can only tell you what the field around you looks like. Coverage, privacy and limitations are deferred to docs/benchmark-coverage.md.
Where the module boundaries cost you time
The single-responsibility rule is enforced hard. Retrieval, evaluation, matching, writing and submission checking are not supposed to cross into each other's territory. In practice that means a request like "is this good enough for Red Dot and write the entry" will not be answered in one pass. You evaluate, then match, then prepare text, and the pipeline skill exists precisely because that is annoying.
The second cost is verification. Because deadlines, fees, eligibility, categories and submission specs must be re-checked against official pages at run time, the submission check is only as good as the pages it can reach in that session. If the agent has no network access, or the official page is a PDF behind a form, the check degrades into a list of what it could not confirm. The README does not document an offline fallback or a cached rule set, so plan for a session with browsing available.
There is also a language surface to watch. The repository ships README.md, README_EN.md and README_JA.md, and the badge line lists Chinese, English and Japanese. The skill bodies are not described in the README as translated, so if your team works in English, check the SKILL.md files you install rather than assuming the English README reflects every instruction the agent will follow.
How this differs from a general-purpose agent with a long prompt
The obvious alternative is a single long prompt that asks the same model to research awards, critique the work, pick a target and draft the text. That approach is cheaper to set up and needs no installation, but it collapses the stages. Facts, inferences and unknowns blend into one confident paragraph, and there is no artefact that records which award rule was checked when.
A second alternative is a human award consultant, who brings jury-culture knowledge that no public document contains. The README draws that boundary explicitly: the skills align with published evaluation criteria but do not simulate unpublished jury preferences or internal processes. So the tool is not a substitute for someone who has sat on a jury. It is a substitute for the spreadsheet and the scattered notes.
A third comparison is the generic skills marketplace pattern itself. This repository is not a single skill but a package with a shared dependency, a pipeline router and a designated install order for the shared package. If you install design-award-search alone and skip design-judge-shared, the README warns that the skill depends on shared material. That dependency is easy to miss when you install one skill at a time.
Licence, maintenance and what upgrades look like
The repository is Apache-2.0 and the LICENSE file sits at the top level. For most teams that means you can use the skills internally and modify them; Apache-2.0 also carries an explicit patent grant and requires that notices be preserved. If you fork the skills into a commercial product, read the licence text and any NOTICE requirements with your own counsel rather than relying on a summary.
On maintenance, the last push to main was on 2026-08-24. The repository is not archived. There are no retrieved releases, so updates arrive as commits rather than versioned tags, which is why the README's update path is npx skills update --global --yes rather than a pinned version. Upgrading means pulling the current skills directory, and because the skills instruct the agent to re-verify award rules at runtime, a stale install is less dangerous than a stale rule file would be. The real upgrade cost is re-reading the SKILL.md files when the skill set changes, since the pipeline router's behaviour depends on which modules are present.
Editorial conclusion
Adopt it if you already run Claude Code, Codex, OpenClaw, OpenCode or Hermes Agent and you want award research broken into modules you can trigger one at a time, with facts, inferences and open questions kept apart. Do not adopt it if you expect a winner predictor: the README states that the 22,125-record observation set is descriptive background and cannot be used to estimate award probability, and the skills explicitly avoid simulating unpublished jury preferences. Before you rely on it, open skills/design-evaluation/SKILL.md and check how the maturity label you pass in changes the scoring, then confirm the target award's current categories and deadline on the official page, because the skills are written to re-verify those at runtime rather than cache them.
Frequently asked questions
What are the essential skills of a judge, according to Design Judge Skills?
The repository does not describe judging as a set of personal skills. It encodes published award criteria into a transparent scoring step, and the README states that the skills align with public evaluation standards while refusing to simulate unpublished jury preferences or internal jury processes.
What criteria does Design Judge Skills use to judge design competitions?
The design-evaluation skill scores two things separately: the design itself and its presentation, with evidence confidence and Critical issues reported alongside. The README says the evaluation module draws on aggregated observation records from iF, iF Student, Red Dot and IDEA, but that these records only provide descriptive background and do not change the core score.
Do I need engineering design skills to use Design Judge Skills?
No. The skills are aimed at design award work: retrieving verified winning cases, evaluating a design, matching it to awards and categories, drafting entry text and checking a submission package. The README frames the audience as people preparing award entries, not as engineers building the product.
Community notes