Yesterday-AI

Establishes how agents find and invoke ytstack skills. Requires Skill-tool invocation before any response when a ytstack skill applies. Auto-injected by the SessionStart hook. Not meant for direct user invocation.

Run verification commands and confirm output before claiming work is complete. Evidence before assertions. Thin ytstack wrapper around the vendored superpowers verification-before-completion procedure, injecting the current task's verification command. Use before commit, before PR, before summarize-task, before any 'done' claim.

After a slice completes, review whether the milestone roadmap still fits reality. Reads recent summaries, surfaces deviations, prompts for roadmap changes (add/split/reorder/remove slices). Run at slice boundaries, not mid-slice.

Audit ytstack internal consistency. Compares README (source of truth) against docs/concept.md, .ytstack/DECISIONS.md, and actual plugin content (skills/, hooks/, agents/, artifacts). Reports drift: missing or extra skills, workflow mentions that do not resolve to real skills, trigger-map gaps, hook-list mismatches, skill-count drift, banned-words hits. Run before approving doc changes or after a v

Atomic Design methodology for structuring React/Next.js component libraries. Defines five levels (atoms, molecules, organisms, templates, pages) with hard import rules, classification tests, and anti-patterns. Use when creating, reviewing, or restructuring frontend components.

Use when picking technologies for German public-sector projects (Bund/Länder/Kommunen), EU digital sovereignty contexts, or when user mentions Deutschland-Stack / DS-konform / Verwaltungs-IT. Biases tech choices toward the 128 graduated open-source standards catalogued by the German government. Fetches live data from technologie.deutschland-stack.gov.de.

Use when picking software for EU customers, GDPR-driven replacements, digital-sovereignty contexts, or when user asks for the European / open-source version of a US product (Google Drive, Dropbox, Slack, …). Biases tech choices toward the 289 community-curated alternatives at european-alternatives.cloud. Fetches live data. Complements deutschland-stack-api (community-curated vs. government-graduat

User-triggered session handoff. Writes a detailed HANDOFF.md with current state, open items, pending decisions, and recommended resume prompt. Use before switching machines, before long breaks, or when you need to clear context intentionally. Parallel to the automatic pre-compact hook.

Scaffold a ytstack project's `.ytstack/` directory with the 6 core artifacts (PROJECT.md, DECISIONS.md, KNOWLEDGE.md, RUNTIME.md, STATE.md, PREFERENCES.md). Use AFTER a pitch has been validated via office-hours (and optionally plan-ceo-review / plan-eng-review in concept-mode) -- init-project reads OFFICE-HOURS.md frontmatter to populate PROJECT.md name + one-liner. Do NOT use as the greenfield en

Use as the first step when a new project, feature, or initiative has not yet been validated, or whenever the premise needs a forcing-question test. Six forcing questions expose demand reality, status quo, desperate specificity, narrowest wedge, observation, and future-fit. Produces OFFICE-HOURS.md pitch artifact with name + one-liner frontmatter that downstream skills (plan-ceo-review concept-mode

Scaffold, polish, and publish open-source projects with AI-first documentation, governance, and CI. Use when: Starting a new OSS package, polishing/publishing an existing repo, setting up AGENTS.md / CONTEXT.md / DECISIONS.md / ROADMAP.md, onboarding AI agents to a codebase, or creating standard community files (README, LICENSE, CONTRIBUTING, CHANGELOG, CI pipelines, issue templates). Outputs repo

Challenge premise, scope, and ambition via four modes: SCOPE EXPANSION, SELECTIVE EXPANSION, HOLD SCOPE, SCOPE REDUCTION. Auto-detects mode: concept-mode reviews a pitch (OFFICE-HOURS.md) before any scaffolding exists; milestone-mode reviews a committed milestone (M###-CONTEXT.md + M###-ROADMAP.md) before slicing. Use concept-mode right after office-hours to stress-test the pitch; use milestone-mo

Review execution approach for feasibility, architecture, data flow, edge cases, test coverage, performance, and security risks. Auto-detects mode: concept-mode reviews a pitch (OFFICE-HOURS.md) for technical feasibility before any scaffolding; milestone-mode reviews slice-plans (M###-S##-PLAN.md) before plan-task. Use concept-mode optionally after plan-ceo-review to sanity-check the pitched approa

Plan the next milestone: goal, exit criteria, rough size. Creates M###-CONTEXT.md and M###-ROADMAP.md. Run after init-project has scaffolded `.ytstack/`, before slice-milestone. Draws the goal from a validated pitch when available (OFFICE-HOURS.md or prior milestones); if no pitch exists yet in greenfield, prefer office-hours first so the milestone is grounded in real demand. Updates STATE.md with

Flesh out a single task within the current slice. Reads M###-S##-PLAN.md, picks the next unplanned task, writes M###-S##-T##-PLAN.md with exact file paths + verification steps. Enforces fits-one-context-window rule. Run after slice-milestone, before executing the task.

Reconstruct working context at the start of a new session. Reads STATE.md + HANDOFF.md (if present) + latest 3 T##-SUMMARY.md files and synthesizes a 'where we are' briefing. Run at session start before doing anything else.

Break the current milestone into slices. Reads M###-ROADMAP.md, asks for slice content until all slices are filled. Creates M###-S##-PLAN.md per slice. Enforces 1-7 tasks per slice (iron rule). Run after plan-milestone, before plan-task.

Clean code, robust architecture, no test slop. Foundational skill for high-quality software engineering, code structure, and reviews.

Dispatch the current milestone's slices to a Claude Code Agent Team. Each teammate works on a slice in parallel with its own 200k context (replaces GSD's fresh-subprocess-per-task pattern). Requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 and v2.1.32+. Run after slice-milestone + plan-eng-review.

After executing a task, write its outcome summary. Reads M###-S##-T##-PLAN.md, prompts for what happened, writes M###-S##-T##-SUMMARY.md, flips checkbox in slice-plan, updates STATE.md. Optionally stages git. Run once per task completion.

Root-cause debugging in four phases: investigate, analyze, hypothesize, implement. Iron Law -- no fixes without root cause. Thin ytstack wrapper around the vendored superpowers systematic-debugging procedure, routing findings into KNOWLEDGE.md / DECISIONS.md. Use when hitting a bug, test failure, or unexpected behavior.

RED-GREEN-REFACTOR: write a failing test, watch it fail, write minimal code, watch it pass, commit. Thin ytstack wrapper around the vendored superpowers test-driven-development procedure, injecting the current task's plan as context. Use during task execution.