Project Recall: Round 1 proposal
1. One sentence: What does the Skill do?
Project Recall generates a cited Markdown wiki from a NAS production archive, enabling owners to resume dormant projects and spot decisions whose supporting evidence or linked assets require verification.
2. Who is the target user and problem?
Video creators managing NAS archives that store readable scripts, briefs, or approval logs alongside media files. Returning after months requires piecing together project goals, past decisions, and open questions. While practitioner reports reflect this friction, customer demand remains unvalidated.
3. What is your proposed solution?
Maintain three Markdown files inside a dedicated wiki directory: project context, an evidence index, and an update log. Track each decision’s source version, scope, and linked deliverable while preserving prior decisions and human annotations. Source files remain unmodified.
4. What is the Core User Flow?
Prompt: “Help me resume last year’s campaign.” The system confirms target folders and whether material is synthetic or confidential, parses documents, drafts the wiki, and saves after user agreement and platform approval. In later sessions, Recall checks saved evidence and linked files; modified file bytes flag a review without revoking prior approvals, though old sign-offs cannot certify replacement bytes.
5. What substantive difference does your idea offer compared with existing tools or conventional AI solutions?
Existing wiki Skills and ArgusNAS overlap with this project. Our approach separates documented sign-offs from physical file observations and re-evaluates their link upon return. Text edits and binary file updates carry distinct implications—neither a "final" filename nor an "Approvals" directory guarantees valid authorization.
6. Which UGREEN AI NAS capabilities will be used, and at which step in the workflow?
Setup and Recall: file-agent enumerates scoped directories, inspecting metadata and file hashes. Update: file-agent parses plain text, while document-agent extracts citations and coverage from PDFs and office files. Following user consent and platform approval, file-agent writes and verifies the wiki. Large linked media files undergo metadata-only checks.
7. What was your technical approach?
Inputs comprise distinct source/wiki paths and project intent. Pure Markdown instructions orchestrate registered agents without custom scripts. The workflow retains original QA prompts, verifies visible quotes, logs actions prior to writing, and halts on partial errors. Confidential data requires local-model confirmation before ingestion. Configured GPT 6.1 Sol assisted naming, scoping, development and drafting; Claude Opus 5.5 reviewed; Aside browsed. Upstream identities unverified.
8. What is the current development status?
The workflow and six reference files are implemented, passing static checks (0/0/0). A 47-minute simulation finished build, update, and recall phases—though its initial Ask lacked fresh file verification—and a denial/recovery test succeeded. Earlier 35-minute runs timed out. Citation accuracy remains partially unverified, with no live NAS runtime, user validation, or benchmark superiority claimed.
9. What is your final validation plan?
Core 1: Build and recall in a new session, rejecting final99 as valid proof of approval. Core 2: Process an amendment and swap linked bytes while retaining history, pending approvals, and user notes. Edge: Deny a write after the Started log; report unfinished work and resume without duplicating IDs.
10. Are there any known limitations of your Project?
Extraction may miss evidence or misstate citation locations, and absent records stay unresolved. Static metadata on large files cannot guarantee byte integrity. The Skill lacks video analysis, NLE project recovery, rights verification, and backup certification. Writes are non-atomic, single-writer, and lack background polling. Local routing relies on platform configuration and user confirmation. The Skill requires pre-existing readable companion documents.
Third-party license information
UGREEN Uliya Skill Kit 0.7.7-en and capability catalog 2.2.0: authoring and checking references under UGREEN's terms; not redistributed. Concept reference: Andrej Karpathy, “LLM Wiki,” https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f . No gist text or code, third-party executable libraries or model weights are incorporated.
AI tools and processing disclosure
The entrant supplied the NAS LLM Wiki idea and revised the proposal. Configured GPT 6.1 Sol through cliproxy/Pi assisted planning, naming, scope, code, synthetic fixtures, simulation and drafting/editing. Configured Claude Opus 5.5 provided completed independent reviews. Aside, using its configured model, retrieved public sources and assisted submission; requested model names are not independently verified upstream identities. Development-only Python, ReportLab and PDF tools are not included in the Skill ZIP. Runtime uses Uliya's registered file/document agents, with platform-controlled model routing, not a custom external API or upload service. Real confidential source and wiki content requires the user's confirmation of the NAS local model before reading; cloud consent does not bypass this condition. No real client data was used in the local simulator. The Skill does not independently enforce model routing. Synthetic media/editor files were non-playable placeholders.
References
- Andrej Karpathy, LLM Wiki: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f (concept inspiration, no copied gist code/text).
- http://aaronwilliams.tv/2013/10/making-sense-of-file-organization/
- https://moshermedia.com/blog/organizing-footage-and-other-assets-for-editing Dated practitioner/company workflow descriptions, not customer interviews or adoption validation.
Built With
- markdown
- ugos
- ugreen-ai-nas
- uliya