agent-kevin

End-to-end refresh — compile pending raw inputs, lint+fix the wiki, run a flywheel pass across active projects, surface what needs attention (including a pending plugin upgrade and any planning/review skill that's come due, with the slash command to run it), optionally chain into a morning or evening briefing, snapshot recent Claude Code sessions (where-am-i radar), then refresh both dashboards (TASKS.md + dashboard.html) last so they capture the briefing's news and the run's final state, and close with a short interview offering concrete next steps (only when something's actually surfaced) that you can act on now or queue as a task. Run anytime you want to bring Kevin's state fully current and get one consolidated update. Heavier than quick-pulse, lighter than running each skill by hand.

Guided first-run onboarding for the agent-kevin plugin. Walks through Kevin's character (SOUL), role (IDENTITY), your basics (name, timezone), an optional web pull from your blog/site/LinkedIn/etc., and communication style — then scaffolds CLAUDE.md (operating manual + @-imports), SOUL.md, IDENTITY.md, USER.md, .claude/settings.json, and seeds four system-architecture concept articles into knowledge/concepts/. If a CLAUDE.md already exists at the home directory, Kevin's version is written to CLAUDE.local.md instead. Skill packs are configured inline at the end or via /agent-kevin:configure-skills any time later. Invoke once after installing the plugin.

Create a git worktree for parallel agent work and bootstrap it so it's ready to code — copies the gitignored local files (`.env*`, `.claude/settings.local.json`, `.cursor`, `.cmux`) from the main checkout, installs dependencies, and builds the packages. Use whenever the user asks to spin up a worktree, work on a branch in parallel, set up an isolated checkout for another agent, or "make a worktree for <feature>". First pins down WHICH repo the worktree is for (the user's words, the `$KEVIN_CODE_PATH` default when they assume you know, or by asking when neither resolves), then creates the worktree as a sibling of that repo, never nested inside it, and offers to add it to a sibling `*.code-workspace` if one exists.

Tailored evening wrap — what shipped, what drafted, what stalled, goal-delta, and tomorrow's first move. Strict today-only scope; cheeky one-liner when the day is genuinely empty.

Cross-project work session. Triage active tasks, advance every project meaningfully, close what's done, log what mattered. Invoke when you have time to work across the whole portfolio rather than one focus area.

Set this month's themes and big-rocks goals — reads the full task board + last 2 weeks of sessions, then interviews you (2-3 rounds) before drafting themes. Run on the 1st of the month or whenever feels right.

Tailored morning brief — today's priorities, drafted artifacts, goals delta, per-project pulse, stale callout, signal-topic news, geopolitical news, and one concrete first move. Run when you sit down at the start of the day.

Interactive session where you and Kevin look at accumulated feedback together and decide what to change in Kevin's prompts or skills. Run when you've accumulated enough corrections in raw/user/feedback.md to be worth processing.

Set this week's goals — reads the full task board + recent sessions, then interviews you (2-3 rounds) before drafting. Writes the goals block in TASKS.md. Run on Sunday or Monday.

Show the Claude Code sessions from the last 24 hours scoped to the folder Kevin runs from (the HOME and everything beneath it), with a substantive summary of what each was about, where it left off, and the resume command. Use whenever the operator asks "where am I", "what sessions are running", "what was I working on", "which sessions are open", "I'm lost / overwhelmed", "what did I leave off on", or wants to find/resume a recent session. Also useful at the start of a day or after a break to re-orient. Accepts an optional hours window (e.g. "/agent-kevin:where-am-i 48") and "all" to include every project on the machine.

Plan the year quarter by quarter — reads the full task board, projects, and goal history, then interviews you before drafting per-quarter outcomes. Run quarterly: mid-year it shapes the remaining quarters; in Q4 it drafts next year starting from Q1.

Apply pending HOME migrations after a plugin code update. `/plugin update` refreshes plugin code (skills, hooks, MCP server, templates) but never touches a home's scaffolded files (CLAUDE.md, SOUL.md, settings, rules) or runs bun install. This skill reads the CHANGELOG's Upgrade blocks from the home's recorded baseline up to the installed version, backs up, runs bun install when needed, auto-applies functionality-critical changes, and asks before touching anything you may have personalized. Use when the SessionStart banner / dashboard shows "upgrade available" or "enable update tracking", or the user says "upgrade kevin", "apply the update", "I just ran /plugin update".

Configure Kevin's optional skill packs (SEO, Browser, Database, GitHub) or author a brand-new custom skill. The pack skills ship with the plugin and auto-load — this skill just wires up API keys, MCP server registrations, database connections, and tool permissions. Custom-authored skills land in `<HOME>/.claude/skills/<name>/`. Invoked at the end of /agent-kevin:init or any time after.

Producer-only. Cut a versioned release of the agent-kevin plugin — analyze commits since the last release, detect what consumers must do to upgrade (bun install, settings, template/HOME changes), bump plugin.json, prepend a CHANGELOG entry with a machine-actionable Upgrade block, and stage the commit + tag for approval. Use when the maintainer says "cut a release", "release kevin", "ship a new version", "bump the version and update the changelog". Not for consumers — they use /agent-kevin:upgrade.

Pull real Google Search Console data (queries, impressions, clicks, CTR, average position, indexing status) for the configured site via `bun run dispatch google-search-console`. Use this instead of guessing what users search for, whenever you need SEO or search-traffic evidence.

Retire a completed or cancelled project cleanly. Moves the project folder to the root-level archive/, strips references from active docs, heartbeat schedules, and task-management code, adds a final-thoughts banner to the project README, logs the archival in today's session file, and runs the knowledge compile to refresh the project index. Use when the user says "archive <project>", "retire <project>", or declares a project done/cancelled.

Spin up a new project under projects/<slug>/ with a README and tasks folder. The 2-letter task prefix is derived from the slug automatically — no config edit. Use when the user says "create a new project called X", "start a project for Y", or otherwise declares a new multi-artefact initiative worth its own folder.

Plan a trip end-to-end and generate a polished, interactive, print-ready HTML itinerary. Use whenever the user wants to plan a vacation, family trip, day trip, road trip, or weekend getaway, asks for an itinerary, or wants help organizing travel dates, flights, routes, or hotels, even if they never say the word "itinerary". Wizard-style: interview first, then web research (flights, drive times, prices, opening hours, halal food, weather), then render the itinerary HTML into the trip project.

Read-only SEO audit for the site configured in GSC_SITE_URL. Pulls GSC + PageSpeed + WordPress (if applicable) + on-demand SerpAPI for anomaly investigation, applies four audit rules, ranks findings by impact, writes a markdown report, and threads notes into any matching active tasks. Diagnostic-only — never publishes or edits. Invoke on demand or wire to any scheduler (cron, GitHub Action, etc.). Trigger phrases — "run the SEO audit", "audit the site", "gsc audit", "search console audit".

Query WordPress content via the public WP REST API — posts, pages, custom post types, categories, tags, media, and modified dates. Use this when auditing a WordPress site's content, cross-referencing with Search Console data, or listing products/brands from Toolset custom post types. Targets whatever site is configured via `GSC_SITE_URL` (same site used by `google-search-console`). No authentication required for published content. For SEO title/meta (The SEO Framework, Yoast, RankMath, AIOSEO), fall back to Playwright — those fields are injected at render time and not in the REST response.

Remove signs of AI-generated writing from text. Use when editing or reviewing text to make it sound more natural and human-written. Based on Wikipedia's comprehensive "Signs of AI writing" guide. Detects and fixes patterns including: inflated symbolism, promotional language, superficial -ing analyses, vague attributions, em dash overuse, rule of three, AI vocabulary words, negative parallelisms, and excessive conjunctive phrases.

Deep-dive specification writer that reads input files, interviews the user with Socratic questioning, and produces a comprehensive spec document. Use this skill whenever the user wants to plan a feature, write a spec, flesh out an idea, turn rough notes into a structured plan, or go from brainstorm to blueprint. Triggers on /plan-spec. Also trigger when users say things like "help me spec this out", "plan this feature", "turn these notes into a spec", "interview me about this project", or "I need to think through this design". Even if the user just has a vague idea and a file with some notes, this skill applies.

Review the target (script/app/area/change) and simplify it. Ensure it is elegant, simple, concise, accurate, robust, reliable, and follows best practices — not over-engineered or over-complicated. Remove dead code. Only runs when explicitly invoked via /simple-simplify.

Drive a website in a VISIBLE browser to run real, repeatable flows end to end — scrape a page into structured data, fill a form, or click through a multi-step task. Kevin opens a headed Chrome and (only when a flow needs it) waits for you to log in by hand — your real session, no API keys. Flows are pluggable, one folder each; `hacker-news` is the reference example. Manually invoked only — use /agent-kevin:browser-flows with plain instructions like "digest the top Hacker News stories" or by naming a flow.

Refresh and open the Agent OS dashboard — a static HTML mission-control page covering today's plan and activity, work across projects, sessions, Kevin's brain, reports, capabilities, and system internals. A refresh snapshots the latest Claude Code sessions (where-am-i radar) and regenerates both dashboard.html and projects/TASKS.md from current state, without the heavy sync work (no compile, flywheel, or briefing). Use when the user says "refresh the dashboard", "update the dashboard", "open the dashboard", or wants the big picture of what Kevin is and what's going on.

Quick scan — anything stalled, any recent errors, any tasks past their cadence. ~1 minute read. Use when you want a status check without a full briefing.

Compile raw session logs, feedback, and inbox items into structured wiki articles. Orchestrated in this session — MCP returns work items, Claude synthesizes, MCP confirms.

Run Google PageSpeed Insights (Lighthouse performance scores + Core Web Vitals — LCP, CLS, INP, FCP, TBT, SI) for any public URL via `bun run dispatch google-page-speed`. Use this whenever performance or Core Web Vitals evidence is needed (auditing a specific page, diagnosing ranking drops, comparing mobile vs desktop).

Free domain-authority proxy via Open Page Rank — returns a 0–10 rank per domain for trend tracking and rough competitive comparison. Use this for monthly DA-trend tracking or quick authority comparison across competitors, not as an absolute benchmark against Moz DA.

Real Google SERP data (organic results, AI overview, shopping carousel, People Also Ask, knowledge graph, related searches) via SerpAPI. Use this whenever you need ground-truth on what's ranking for a query, what SERP features Google is showing, or how the SERP differs between desktop and mobile. This is the correct tool for SERP investigation — headless playwright will CAPTCHA on Google and should not be used for SERP capture.