bharat3645
UserAgent harness / control-plane around a stateless LLM: deterministic policy floor on PreToolUse/PostToolUse hooks, secret redaction, injection fencing, tool-result compression, code-graph memory, orchestration. Five plugin modules for Claude Code (Tier 1) or any MCP host (Tier 2). MIT.
Categories
Indexed Skills (9)
design-critique
Pre-emit self-critique for UI/visual/design output (components, CSS, copy, layouts) — anti-slop gates, taste dials, and a design-token contract check, run BEFORE the work is shown, not after. Use whenever producing or editing anything with a visual/UX surface.
session-observer
Watch this session for corrections, repeated patterns, and judgment calls, and record them as episodic observations for later human-reviewed skill/policy improvement. Use continuously during any non-trivial session, not as a one-shot invocation — the value is cumulative.
subagent-driven-development
Execute a multi-task plan as a controller that dispatches one fresh-context subagent per task, reviews each result against the spec, loops fixes, and tracks progress in a durable ledger. Use for any non-trivial build so the controller's context stays lean and every task is independently reviewed.
using-ideal-harness
Bootstrap skill for the Ideal Harness. Explains the installed modules and when to route to them. Injected at session start so the agent knows the harness is active.
brainstorming
Turn a vague request into an agreed design before any code is written. HARD GATE — produce a design and get explicit approval first; do not implement, scaffold, or edit until the user approves. Use whenever a request is non-trivial or under-specified.
caveman
Output-side token compression. Write terse, high-signal prose — drop articles, filler, and pleasantries — while keeping all technical substance, code, and errors verbatim. Use when output token volume matters and prose polish does not.
focus
Answer-first, structured output mode — lead with the action, number multi-step work, restate state across turns, suppress tangents. Use when the user wants to act on your answer immediately, not read around it. A different axis from caveman (structure/legibility, not token count) — combine freely with it.
grill-with-docs
Before brainstorming or coding, interrogate a vague request for scope/constraints AND ground it against real, current docs (not training-data memory) for any library or API it touches. Produces a CONTEXT.md the brainstorming/planning phase consumes. Use at the very start of any non-trivial task that names a library, framework, or external API.
tdd
Write the failing test before the implementation, for a single piece of code — red, green, refactor. A finer grain than the ledger's task-level verify.command (subagent-driven-development); use this INSIDE an implementer's work on one task, not as a replacement for the ledger.
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.