shuaeb6

Sponsordesk — AI skill for Claude Code

AI community

Human-approved AI triage for a 401(k) service desk: PII tokenization, call-type classification, plan and payroll record checks, routing, and a code-enforced close guard.

How to install Sponsordesk

This entry records only its repository, not the path inside it, so there is no exact command to give. Open shuaeb6/sponsordesk and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Sponsordesk does

Human-approved AI triage for a 401(k) service desk: PII tokenization, call-type classification, plan and payroll record checks, routing, and a code-enforced close guard. MCP servers and a an AI agent(Claude).

Alternatives in AI

Tokenization — ai-research-skills HuggingFace Tokenizers, SentencePiece 5.4k ★
Sentrux — Real-time architectural sensor that helps AI agents close the feedback loop, enabling recursive self-improveme 2.9k ★
Claw Empire — Command Your AI Agent Empire from the CEO Desk — A local-first AI agent office simulator that orchestrates CLI 1.1k ★

README

SponsorDesk

**Human-approved AI triage for a retirement plan service desk.**

Employers who sponsor 401(k) and 403(b) plans send a steady stream of tickets: rejected payroll files, missing contributions, enrollment questions, loans, distributions. Each one means reading the ticket, finding the plan record and payroll files, choosing a team, and writing a reply, all while the text is full of SSNs and account numbers.

SponsorDesk does the groundwork and leaves every decision to a person:

Redacts PII into reversible tokens before any model sees the text.
Classifies the request into one of eight call types, with a confidence score.
Checks the records: plan rules, recent payroll files, funding status, and loan limits, to find the likely root cause.
Routes and drafts: recommends a team with a reason and writes a plain-language reply.
Waits for a reviewer to approve, edit, or reject. Nothing is sent automatically.
Guards the close: a code-level rule (not a prompt) refuses to close any ticket whose resolution criteria are not confirmed.
Audits everything: every proposal, decision, and blocked close is logged as structured JSON.

flowchart LR
    T[Sponsor ticket] --> R[piiguard
tokenize PII]
    R --> C[Classifier
call type + confidence]
    C --> X[Record checks
plan, payroll, funding]
    X --> P[Route + draft]
    P --> H{Reviewer}
    H -- approve / edit --> A[Routed to team]
    H -- reject --> T
    A --> G[Close guard]
    G -- criteria met --> D[Closed]
    G -- blocked --> L[(Audit log)]
    P --> L
    H --> L

Results on the held-out set

Check	Result
Raw PII values reaching the model	0 of 1,399
Routing accuracy	93%
Unresolved tickets that could be closed	0 (enforced by tests)
Call-type accuracy (TF-IDF + logistic regression)	100%, an upper bound on templated tickets; harder paraphrased tickets are on the roadmap