Inspiration
Useful creative material often sits unused across folders: a photograph without its context, a half-finished journal entry, a voice note that contains the seed of a story. Found Stories explores how those pieces can become a new photo essay while preserving the distinction between visible evidence and remembered experience.
What it does
The Skill inventories a user-authorized collection, gathers visual and textual evidence, proposes story directions, and drafts a photo essay with an ordered image sequence, quotations, source references and missing-information notes. A user can choose a direction or request automatic selection. Saving requires an actual destination and the platform's approval workflow. Originals stay unchanged.
How we built it
The submitted package follows the Uliya Skill Development Standard and was authored with UGREEN's official v0.7.7 Skill Creation Tool. It delegates to registered file, photo, document and knowledge experts using catalog-supported capabilities. It contains only SKILL.md and directly linked Markdown references.
A separate local development prototype uses Python and Ollama to inspect fictional sample images and generate a fresh essay. It validates known source IDs and exact quotations before exporting HTML and Markdown. It is a development experiment, not evidence that the Skill has run on UGREEN hardware, and is not included in the ZIP.
Current progress and limitations
The UGREEN package passes static checks. Eighteen local prototype tests pass, and live local model calls have produced new essays from a second sample collection. The prototype reads voice-note transcript sidecars; it does not perform speech recognition. Model interpretation and writing quality still require human review. UGREEN installation, actual expert response shapes, model routing, audio transcription and approved saving remain pending target-device validation.
What is next
On the assigned finalist NAS, test the full registered-expert workflow with synthetic materials, verify exact source references and unchanged originals, exercise fresh transcription, and test missing files, partial reads, cancelled actions and approval resumption.
AI and third-party disclosure
Codex assisted ideation, code, tests, documentation, fictional text and submission drafting. OpenAI image generation created fictional test photos. macOS speech synthesis produced local voice samples. Local Qwen models via Ollama performed vision and writing in the separate prototype. Sources, licenses and boundaries are documented in the proposal and the ZIP's linked disclosures. No model weights, sample media, third-party libraries or executable code are included in the submitted Skill package.
