bifrost

Zero-context fresh-eyes review of specs and briefs, multi-agent Terraform/Terragrunt (IaC) review, repository refactor planning, and application-logging authoring/review guidance. Maps to spec-kit's analyze phase.

31 skills

Project setup, decision records, and session priming — scaffold a dev-docs/standards structure (template/backfill/upgrade modes), record architecture decisions (ADRs) under docs/adr/, write anything an agent reads and put each line where it loads (CLAUDE.md, AGENTS.md, .claude/rules, skills, memory), review agent files and memory notes clause by clause against the agent-doc playbook (.claude/rules, CLAUDE.md, AGENTS.md, skills, memory), and assign a session persona. Maps to spec-kit's constitution phase.

31 skills

Design-token architecture — build, name, theme, and govern a design-token system the way the field's canonical sources converge on it. Covers the three-tier model (primitive/semantic/component), naming grammar, OKLCH color scales, DTCG authoring, Style Dictionary pipelines, dark-mode/multi-brand/density theming, CSS-variable distribution, WCAG contrast gates, and token semver/deprecation. Also the floorplan C4 design loop: propose a model change, check it, render it, show it in the studio, ask on the canvas with Huginn annotate, and accept it.

31 skills

Implement from an accepted spec. /stark-build runs one fresh headless session per task gated by checks the agent cannot edit, with a held-out e2e gate, one cross-vendor advisory review, and one bounded fix round for medium+ findings.

31 skills

Repository operations beyond the spec-driven pipeline — coordinate Minions with Gru, cut releases, run session start/end, hand off work across sessions, run one ticket solo and unattended with Agnes, open a Lucius brainstorm in a tab of its own, and send rich text to Slack, Jira and GitHub with the fleet commands, with no code that builds Block Kit or ADF.

31 skills

Plan work before it starts — author a human-gated spec with its task DAG, and write short tickets someone else can pick up, with no guessed claims.

31 skills

Long-form writing craft, from the author's reusable voice profile through story editing, adversarial sharpening, independent judging, and multi-model jury calibration.

31 skills

A pane beside the chat that walks you through a deck of decisions: Claude presents the deck with its present tool, you pick options or write your own words for each one, and Send returns your answers to Claude as your next message.

Claude asks you questions in a pane beside the chat and keeps working: pick an option or write your own words, and each answer reaches Claude in its running turn, or as your next message once the turn has ended.

A band above the prompt for the Needs you lines asked since your last message: answer each with one key or your own words, or take all of Claude's picks at once. Your next message sends the answers and closes the band.

The Gru pane beside Gru's chat: a calm board of the run (the epic's title and when the run started, counts, a bar graph of the tickets at each step, the tickets worth a look, a metro line per ticket with its state in UPPERCASE and its own color, the ticket list grouped in progress, to do and done, and a button to each open worker tab), a Log tab of the big steps Gru and each worker did, in plain words, a Stats tab of Gru's own counts and the whole run's (session and working time, tokens, lines, commits, PRs, review rounds, tools, prompts, messages between sessions, model and effort), a GitHub tab of each ticket's PR with its Actions runs and its commits, grouped open, merged and closed like the ticket list, read from GitHub while the tab is shown, and a Settings tab Gru reads before each launch. In each Minion and Agnes tab, the ticket pane beside the chat: the one ticket (its name, its epic's title and when it started, as the Gru pane names them, its metro line, whether it is worth a look, its activity with its last hour and the next step), a Log tab of its steps, a Stats tab of its worker's counts, and a Settings tab the worker reads before its review, its merge and its release; a progress row above the prompt stands in while it is not placed. All drawn from the state files under ~/.cache/stark-progress/, in the terminal and the desktop app.