Personal projects
Side projects &
how they connect.
Trust infrastructure for agentic development, one layer at a time — from not trusting a single risky action, up through whether shipped code actually matched the plan.
Don't trust one action, one session, or unbounded concurrency.
Don't trust a plan nobody checked against.
The same discipline production systems demand — pointed at agents instead of servers.
Shadow Mode
A Claude Code Plugin that intercepts risky tool calls — Bash, Write, Edit — and logs exactly what would have happened instead of executing it. A /shadow-mode command toggles it live; a second, model-invoked skill summarizes a session's shadow log in plain English. Installable from a real self-hosted marketplace, validated against Anthropic's own plugin review checks.
- Python
- Claude Code Hooks
- Claude Code Skills
- MCP
Install /plugin marketplace add uanandu/shadow-mode-marketplace
Backlog Match
A Claude Code skill that turns a casual, half-remembered mention of backlog work — "let's do the meal tracking one" — into a formal OpenSpec proposal, no exact BACKLOG.md row name required. Matches the phrase against a row's Change and What it does columns, checks its Depends on column against the other rows' Status before proceeding, and hands off to /opsx:propose once confirmed.
- Claude Code Skills
- OpenSpec
Install /plugin marketplace add uanandu/backlog-match