jasonm4130-claude-skills

Configurable adversarial panel review for any artefact — plans, code, design docs, prose, model outputs.

Model-tiered, adversarially-verified multi-source research via parallel sub-agents and synthesis with citations. Cheaper alternative to the built-in all-Opus deep-research.

Context-fill-triggered handoff skill — writes a structured resume doc when context fills, auto-loads it on next session.

Interactive session retrospectives — deterministic Stop/PreCompact triggers, diff-driven interview, native memory entries. No external services.

Stop-hook gate against trail-off: detects unpushed commits at turn end and injects an agent-directed nudge to review and finish the branch.

PreToolUse hook that nudges Claude to tier models in high-fan-out Workflow scripts — denies a Workflow call that spawns many agents with no per-agent model: override.

Writes a durable, committed Markdown ADR/plan as the source-of-truth record, and — only when the content needs visuals markdown can't express (wireframes, annotated split-diffs, before/after columns) — also emits a self-contained rich plan.html to /tmp. Markdown canonical, HTML disposable; no backend.

Deterministic workflow-driven subagent development loop — plan/confirm in the controller, then a background Workflow enforces per-task implement/review/fix with tiered models and codified ponytail discipline.

Turns an intent into a grounded, cited, build-ready ADR with load-bearing decisions surfaced to the human, then hands it to the subagent-driven-development loop. Front-end for adr → sdd.