Multi-tier deep-research skill with intensity levels and a quality gate

A Claude Code skill that exposes three intensity tiers — Light, Medium, Heavy — and a 5-phase pipeline (planning → parallel research → synthesis → quality gate → report). Light is one agent, one pass. Heavy spawns parallel aspect-researchers (each doing dual-source search), synthesizes across them, and runs a quality-gate that scores saturation/diversity/tier-quality with a PASS | WARN | FAIL verdict before letting the report-generator finalize.

Without explicit intensity choice, deep-research skills either over-spend on trivial queries or under-deliver on heavy ones. The PASS/WARN/FAIL gate replaces a human "is this enough?" judgment call with a measurable verdict, which makes the skill safe to run unattended (e.g., in a nightly research cron). Used in practice for investment DD, content research, and founder strategy work; pairs with (the broader multi-provider deep-research aggregator) as the intensity-tier wrapper around it.

1. Install the skill into ~/.claude/skills/Research/ — the public reference structure is in the cybos repo at github.com/Gerstep/cybos.
2. On invocation, pick a tier (Light for a single quick question, Medium for a structured topic, Heavy for DD-quality multi-aspect work).
3. The research-planner agent decomposes the topic into 3-7 aspects and writes plan.yaml.
4. For Medium/Heavy: parallel aspect-researcher-v2 agents run dual-source search (e.g. Exa + Brave) and write aspects/{id}.yaml.
5. synthesis aggregates results into synthesis.yaml with cross-aspect insights and evidence links.
6. quality-gate scores saturation/diversity/tier-quality and emits PASS | WARN | FAIL.
7. On PASS, report-generator writes FINAL_REPORT.md. On WARN/FAIL, route back for a gap-fill round.

• The original skill is missing a dedicated "phase 2b net-new gap fill" round and a separate auditor agent — author flagged both as gaps. Without them, WARN verdicts often just produce a longer report rather than actually filling the missing aspects. Add an explicit auditor pass if you rely on this unattended.