essential

Run a metric-driven research loop: define a metric, evaluator, baseline, and target; evolve candidate solutions; score and adversarially verify them; then mutate survivors until the target, budget, or plateau ends the run. Use for measurable optimization of prompts, code, experiments, or creative variants; use deep-research for fact-finding.

Decides between researched approaches before implementation. Use when asked to choose an approach, challenge a recommendation, make an architecture decision, compare options, define rollback and falsification signals, or obtain approval; routes blindspot passes, brainstorms, interviews, references, and prototypes to essential:discover.

Conduct comprehensive multi-source research with AI-assisted analysis, adversarial claim verification, and explicit source synthesis. Use when investigating complex topics, comparing evidence, gathering current information, or producing a fact-finding report with citations and uncertainty notes. Do not use for metric-driven candidate optimization.

Discovers material unknowns before planning. Use for a blindspot pass or unknown unknowns, to brainstorm approaches from cheapest to ambitious, interview about architecture, extract reference implementation semantics, make a disposable prototype before touching the real app, or check whether discovery is ready for a decision; researched option selection belongs to essential:decide.

Check the integrity of local state and durable ADRs with the structural doctor, diagnose folder-structure and format drift against current contracts, and offer user-approved migration or repair. Use for health checks of .state/ or docs/architecture/decisions/, before resuming old work, or after suspected corruption or drift; this skill repairs records, never the work itself.

Create or execute a context-complete cross-domain plan as an orchestrator. Use when another agent must continue without prior context, or when a multi-domain plan needs coordinated execution while this skill retains decision ownership. For coding-session persistence, use essential:handover.

Persist work stream state and update the global overview, both in the default source tree's centralized .state/, so the session can pause and any later session resume from the files on disk. Use when pausing coding work; this skill records continuity and does not execute the work.

Discover, validate, stitch, and install specialist agent templates contributed by Essential, its marketplace, and explicitly trusted enabled marketplaces. Use when asked to install agents, set up subagents, refresh the agent team, or configure Claude Code, Codex, or Grok Build on a new machine.

Install bundled Claude Code output-style Markdown files into ~/.claude/output-styles. Use when asked to install, refresh, or repair a response style shipped with this skill; do not use for editing style content or project-scoped .claude/output-styles files.

Install the bundled Bullet Train statusline into ~/.claude and wire settings.json statusLine. Use when setting up Claude Code on a new machine, installing or restoring the statusline, or repairing its configuration; preserve the bundled executable and report permission or platform limitations.

Resume paused work from the default source tree's centralized .state/works/. Offer incomplete streams, settle each reviewing stream against its applicable landing evidence, resolve pending decisions, delegate planning to the relevant lead, and drive the selected stream toward its charter. Promote durable results before non-coding completion; route coding publication through the relevant change-publication capability.