HK//HACKERSHACKATHONS · KAGGLE · FOR BEGINNERS中文
HACKERS / Hackathons

UGREEN Hackathon 2026 on AI NAS: a beginner guide to building a Uliya Skill

UGREEN pays $46,000 for AI Skills that run on its AI NAS, and Round 1 needs no NAS at all: you submit a PDF of at most 500 words plus a ZIP holding a Markdown-only Skill package by Nov 15, 2026 at 11:45 pm PT. There is no code to write here. A Skill is a SKILL.md that hands work to six built-in expert agents, so the real test is whether that file passes the format rules. We downloaded the English starter kit (v0.7.3) and ran its own checker, scripts/check-skill.mjs, against test Skills. Three findings matter most. First, two headings in the kit's own English template (# Capability evidence and # Expert orchestration) fail the kit's own checker. Second, plain-English sentences can fail its security scan. Third, the catalog lists 78 NAS actions, and 27 of them sit behind an approval gate.

DEADLINE
PRIZE$46,000UGREEN
Official page ↗

Brainstorm word cloud

Terms from this hackathon's own brief, judging criteria, discussion board and submissions, plus our guide. Bigger = mentioned more often; wording that most Devpost hackathons share is left out. Click a word to see past winning projects that mention it.

Key facts

What you build
An AI Skill for UGREEN AI NAS that follows the Uliya Skill Development Standard. It is a SKILL.md file, plus optional Markdown files under references/, and it uses the supported UGOS CLI capabilities through registered expert agents. Custom executable scripts, general-purpose runtimes and direct access to NAS resources are not supported ↗
Prizes
Gold $20,000, Silver $10,000 and Bronze $6,000, plus two Most Creative Skills awards of $5,000 for the highest Innovation & Differentiation scores. Each project can win only one prize ↗
Two rounds
Round 1 (Oct 22 – Nov 15) is judged on the PDF and ZIP alone. NAS runtime evidence is not required, and missing it does not by itself lower the score. Up to 30 finalists then get one testing day on a remote iDX6011 Pro (64GB) between Nov 22 and 28 and submit a final package plus a public YouTube demo under 3 minutes by Nov 30. Round 1 scores do not carry over ↗
Round 1 PDF
Ten questions, answered in at most 500 English words in total. They cover the one-sentence pitch, user and problem, solution, core user flow, the difference from existing tools, which NAS capabilities are used at which step, the technical approach, development status, two core tests plus one edge case, and known limitations. Third-party components, their licenses and any external data flows must be named ↗
Package limits (from the kit checker)
ZIP at most 10 MB, each file at most 1 MB, at most 100 files. The frontmatter holds only name (lowercase and hyphenated, 2–64 characters, the same as the folder name) and description (at most 1,024 characters). references/ may hold only ordinary Markdown files, each linked from SKILL.md. Above 500 body lines you get a warning ↗
What the catalog contains
Our count of the kit's capability catalog (v2.0.0): 9 CLIs across 7 domains (file, album, monitor, document, knowledge, media, music), mapped to 6 delegate targets (file-agent, photo-agent, security-monitor-agent, document-agent, knowledge-agent, media-agent). There are 78 actions in total: 47 read-only, 27 state-changing and 4 mixed. 51 need no approval and 27 sit behind a gate: 15 gateway, 10 workflow and 2 conditional. The two playback controls are state-changing but ungated, while document translate and writing are mixed but gated. ug-photo (18 actions) and ug-filesystem (17) are the largest ↗
Who can enter
Individuals of legal age, or teams of up to 3. Organizations are excluded. Only listed countries may enter: the US, Canada (excluding Quebec), much of Europe, Japan, South Korea, Singapore, Hong Kong, Taiwan, Pakistan, the UAE, Nigeria and others. China, India, Brazil, Saudi Arabia, Vietnam, Ukraine, Bangladesh and Panama are among those excluded. A forum thread asks whether excluding Panama was a mistake. Check the list before you build ↗
Language and newness
All materials must be in English, and the demo needs English narration or full English subtitles. You may build on an older idea, but the Skill and its core workflow must be newly developed during the hackathon. The Round 2 project must be the same one you submitted in Round 1 ↗
Open forum questions
Participants have asked whether a gateway approval for ug-filesystem chtimes covers a whole batch, and which date the Photos timeline uses when a photo has no EXIF date. As of Oct 25 neither had an official answer on the page we checked, so a Skill that depends on them should state the assumption in its PDF limitations ↗
What Skill Rank shows
Of the 932 agent skills tracked on PocketPlay Skill Rank on Oct 25, 2026, none mention albums, EXIF, Synology or surveillance. One, invoice-organizer, already sorts invoices. A photo, camera-event or media-library Skill is a less crowded field than invoice or document sorting ↗

Idea angles

Our scoping ideas for beginners, built from the four official tracks and the actions that exist in the kit's catalog. They are not official briefs.

Trip album with a confirm step (Bring Memories to Life)

Use ug-photo search-picture to find photos by place and date, show the user the count, and run create-album only after the user confirms. This is the smallest end-to-end flow in the catalog: one agent and two actions. It works well as a first Skill, and you can then add OCR on tickets with vision-ocr.

Doorbell-to-report (Open Innovation)

Use ug-surveillance event-timeline, identity-resolve and event-evidence, which are all read-only and need no approval, to summarize who came to the door this week. Then use ug-knowledge save-text, after the user confirms, to store the note. Nothing in Skill Rank's 932 skills covers camera events.

Downloads cleanup that never deletes silently (Automate Your Workflow)

Use ug-filesystem ls, ug-filemgr scan-dedupe, and then file-workflow batch-move to a review folder after the user confirms. The kit's own example uses this flow. You stand out by reporting duplicates by size, and by stopping and reporting when the scan returns no results.

Contract and invoice tagger (Automate Your Workflow)

Use ug-filesystem find, ug-document extract and ug-filemgr tag-add-file. tag-add-file is one of only two conditional-approval actions, and invoice sorting already exists as a public skill. To differentiate, add ug-document reconcile to check totals rather than just sorting files.

Shoot-day selects for creators (Unleash Your Creativity)

Use ug-photo search-classify or vision-vlm to shortlist shots from one shoot, ug-filemgr mediainfo for video clips, and tag-set (gateway) to mark selects. Present a numbered shortlist and wait for the user to choose before tagging.

Subtitle fetcher for a movie night (Open Innovation)

Use ug-video search, then subtitle-search, and run subtitle-download (gateway) only with a real file id returned by the search. The kit's example warns against guessing ids.

Past winners in related hackathons

Winner-badged projects from ended Devpost hackathons with the same theme that mention terms from this brief, closest first. Read a few before you pick an idea.

  • Codi: Mobile AI Dev Butler Powered by Flutter & Serverpod — Codi turns your phone into a full-stack builder and web automation agent. Chat to create apps or automate tasks - no laptop needed. Built with Flutter + Serverpod for 6B mobile-first users worldwide. (Build your Flutter Butler with Serverpod)
  • TICIAN Cierre — Recovers VAT credits for Mexican businesses: processes CFDI, assembles the SAT file, and submits every close to a Vertex AI auditor that approves or rejects it with its reasons. (Build with Gemini XPRIZE)

Judging criteria

Rules: Round 1 starts with a pass/fail screen on whether the project fits the theme and proposes an AI NAS solution. Eligible projects are then scored on four equally weighted criteria, and Round 2 reuses the same criteria on the working Skill. Ties are broken by User Value first. The Most Creative awards use only the Innovation & Differentiation score. source ↗

User Value25%

A clear user problem, a clear workflow, useful feedback and a meaningful result. State in one sentence what the user gets at the end, such as an album, a moved folder or an evidence summary.

Innovation & Differentiation25%

A substantive new function, interaction or workflow compared with conventional AI tools. On its own this score decides the two $5,000 Most Creative awards, so combining domains (for example camera events plus documents) can pay off.

Technical Quality25%

Sound orchestration, appropriate use of supported CLIs and expert agents, and proper handling of permissions and errors. In practice that means handling all five delegation states, handling empty results, and confirming before any state change.

AI NAS Fit & Validation25%

Meaningful use of local data, storage, compute and privacy on the NAS rather than merely running there. Name the exact CLI actions at each step in PDF question 6.

Starter stack

  • UGREEN Hackathon Starter Kit Official page with the UGOS Skill Creation Tool (uliya-skill-kit). Put the folder in your assistant's skills directory; the kit names WorkBuddy, Claude and Codex ↗
  • scripts/check-skill.mjs The kit's own static checker: run node scripts/check-skill.mjs <your-skill-folder>. Aim for status static-pass with 0 errors. It needs Node.js; we used v20 ↗
  • data/capability-catalog.json The list of CLIs, actions, effects (read or state-change) and approval types. In the English v0.7.3 kit many description strings lost words in translation, so rely on the field names and action names ↗
  • Generation-time capability evidence The checker only accepts a heading named # Generation-time capability evidence, with lines in the form - `ug-photo` → `photo-agent` (backticks and the → arrow). The template's # Capability evidence heading with the - cli -> agent form fails with CAPABILITY_EVIDENCE_MISSING. Likewise, the template's # Expert orchestration heading fails with SKILL_REQUIRED_SECTION_MISSING; the checker wants exactly # Workflow or # Orchestration ↗
  • delegate command One per line: delegate --target photo-agent --instruction '...' with a single-quoted instruction and no pipes. Placeholders like <path> raise a warning. Calling cli.invoke from the parent Skill is an error ↗
  • Devpost project page Drafts can be saved until the deadline. After a submission period ends, that round's entry is locked ↗

Timeline

  1. Registration and Round 1 submission open (9:00 am PT)
  2. Round 1 deadline: PDF + ZIP (Nov 15, 11:45 pm PT)
  3. Round 1 judging, until Nov 19 5:00 pm PT
  4. Up to 30 finalists emailed credentials, assigned device and testing day (12:00 am PT)
  5. Round 2 opens; each finalist gets one assigned NAS testing day between Nov 22 and 28 (PT)
  6. Round 2 deadline: final package + public YouTube demo (Nov 30, 11:45 pm PT)
  7. Round 2 judging, Dec 1 10:00 am – Dec 5 5:00 pm PT
  8. Winners announced on or around this time (2:00 pm PT)

How a beginner can ship

  1. Check eligibility first

    Read the country list in section 3 of the rules. Residents of excluded places, and organizations, cannot win, however good the Skill is.

  2. Pick one user moment

    Choose one track and one end result (an album, a cleaned folder, a weekly camera report). Write PDF question 1 as a single sentence before anything else.

  3. Map it to real actions

    Open data/capability-catalog.json and list the exact CLI and action for each step. Mark which ones change data, since those need the user's confirmation and a waiting_approval branch.

  4. Write SKILL.md with the four required headings

    The required headings are # Goal, # Required input, # Workflow (or # Orchestration) and # Completion and failure, all as level-1 headings. Do not keep the template's # Expert orchestration, which the checker rejects. Add # Generation-time capability evidence with one - `cli` → `agent` line for each CLI you use.

  5. Cover every state and empty results

    Write a line for each of completed, failed, cancelled, waiting_user and waiting_approval; a missing one is an error. For any search or list step, include the literal phrase "no results" (or "empty result"). We found that "no photos were found" still triggers the EMPTY_RESULT_HANDLING_MISSING warning.

  6. Avoid the security-scan words

    The scan reads every line. Mentioning curl or wget, even in "do not use curl", is an error. So is any sentence where skip, bypass, ignore or disable comes within 40 characters of approval, permission, gateway or audit, even "never skip the approval step"; write "ask for approval" instead. Words like cat, rm, cp, mv followed by a space, or paths such as /volume1/, also fail outside delegate lines. In our test, "Include the cat photos too" was rejected. Code fences tagged bash, python or js are also errors.

  7. Run the checker until static-pass

    Use node scripts/check-skill.mjs your-skill. Our trip-album test Skill, written with the template's capability-evidence format, got 1 error and 2 warnings. After the fixes above it reached static-pass with 0 errors and 0 warnings. Renaming # Workflow to the template's # Expert orchestration alone brought back an error. Zip the folder: SKILL.md at the root, plus references/ only if each file is linked.

  8. Answer the PDF honestly about status

    Question 8 asks what is implemented and what waits for NAS validation, and Round 1 does not penalize missing NAS evidence. Name two core tests and one edge case (for example an empty album search) that you will record on your finalist testing day.

From the discussion board