Scope-Creep-Shield

Inspiration

Scope creep is the risk every new freelancer is warned about: on a friendly client call, small extra requests get promised out loud, and weeks later nobody can prove what was actually agreed. As a software engineering student preparing to take my first freelance projects, I wanted a tool that protects that moment. The written scope already exists, and the call already happens. What is missing is a careful comparison between the two, with evidence.

What it does

Scope-Creep-Shield is a Skill for UGREEN AI NAS. The user places one client call recording and one signed scope document in a project folder. The Skill transcribes the call, extracts every commitment from the scope document, reconciles the two, and saves a Scope Audit report in the same folder. The report lists work covered by the scope, promises said on the call but not in the scope (each with a timestamp, a short exact quote, and the speaker), scope items the call never mentioned, and segments that stayed unclear. It is an audit, not a call summary. Everything stays on the NAS, which matters because client calls and contracts are confidential.

How I built it

The Skill was built with the official Uliya skill kit (capability catalog version 2.0.0). Its SKILL.md delegates each step to a registered expert agent: the file agent locates the two inputs, the knowledge agent transcribes the recording and later saves the report, and the document agent extracts the scope commitments and runs the reconciliation. Three reference files define the report format, the matching rules (covered, extra, unclear), and the input and safety limits, including the rule that unclear speech is marked Unclear and never guessed. The finished package passed the kit's official static checker with no errors and no warnings.

Challenges and limits

Version 1 covers English calls and short scope documents of about three pages. The main risk is transcription quality on poor audio, which the design reduces with clear test recordings, quality warnings, and the Unclear rule. A user-supplied transcript file is also accepted when no recording is available.

What is next

Round 1 submission is this proposal and the pre-test Skill package. If selected as a finalist, the Skill will be installed and tested on the remote UGREEN NAS on the assigned testing day, using a synthetic salon website project: a one-page scope document and a three-minute English client call with two planted out-of-scope promises, plus a no-extras call and a noisy-audio edge case. The final submission will include a demo video under three minutes.

AI tools disclosure

AI assistants, including Muse, were used during research, drafting, and preparation of the Skill package and this submission, under my direction and review. The project concept, test scenario, and final decisions are my own.

Built With

• audio-transcription
• document-reconciliation
• llm-agents
• markdown
• skill-md
• ugos
• ugreen-ai-nas
• uliya-skill-kit