Changelog
Every change on modernEngineer: new pieces and updated pieces.
- Which Claude model should you use: Fable 5.2, Opus 5.5, Sonnet 5.5 or Haiku 5.5?
Start on Opus 5.5 at medium effort, drop to Sonnet or Haiku when the work is clear and easy to check, and try Fable 5.2 when Opus at high still falls short. Includes pricing tables and ways to mix models.
- Connect modernEngineer to Claude and ChatGPT, no coding needed
Add modernEngineer to Claude or ChatGPT in a few minutes: paste one address, sign in, choose Allow. Then ask in plain words and the assistant finds and reads posts for you. No API key, no config file.
- Cognitive surrender: keeping your judgment when agents do the work
When should you check an agent's output closely, how do you check the evidence, and which judgment stays yours? An analysis of Addy Osmani's take on cognitive surrender and accountability when you delegate.
- ak-dogfood: turn a session's stumbles into skill fixes
ak-dogfood reviews the session you just ran, finds where a skill cost you time, tokens or a redo, then proposes the smallest fix to that exact skill, with evidence and risk.
- FDE or rebranded consulting? How to tell before you hire an AI deployment team
Gartner warns of "FDE washing" while a16z calls FDE a way to build a moat. How to tell a real FDE from rebranded consulting, and what to settle before you sign.
- What is a Forward Deployed Engineer (FDE)? The role bringing AI into businesses
An FDE goes into a business, finds the real problem and gets a solution into daily use. How the role differs from consulting, and why it took off in 2026.
- A path to becoming an FDE with AI coding: the skills you need and your first trial project
Already using coding agents and want to become an FDE? Five skill groups to fill in, and how to run a trial project that gives you a portfolio with measured results.
- agent-imagegen: hand your agent a brief, get a WebP in the style you saved
An image skill for any agent that can run shell commands. Codex or agy generates through the subscription you already have; the skill picks a fitting direction, checks the text and delivers WebP.
- Herdr Cook Plan: a fresh worker for every phase, the coordinator only herds
The Herdr variant of Orca Cook Plan: runs an AgentKit plan you already have, one new agent per phase in its own pane. The coordinator dispatches, verifies and commits, and never implements.
- Designing a URL shortener: generating short codes and handling collisions
From the simplest way to generate codes (hash the URL and truncate) to random codes with UNIQUE, a unique ID encoded in base62 and a pre-generated key pool. Each step shows why the previous one falls short, with collision odds.
- ak-ab-test-setup: plan an A/B test before you run it
ak-ab-test-setup helps you write the hypothesis, pick metrics, compute sample size, split traffic and lock the stopping rule before an A/B test runs, then helps you read the results when it is done.
- ak-ads-management: run ad accounts within the limits you allow
ak-ads-management reads reports, builds campaigns, changes budgets and pauses or enables ads through the Google Ads or Meta Ads API, only within the limits you allow, checking receipts and reconciling before any retry.
- ak-advise: ask until it gets you right, then advise straight
ak-advise interviews you one question at a time until you both picture the same outcome, draws a preview for you to confirm, then gives candid advice with a handoff brief.
- ak-affiliate-marketing: partner programs built from unit economics
ak-affiliate-marketing builds affiliate, KOL and KOC programs starting from unit economics, then picks the commission model, platform, partner recruiting and fraud controls, keeping partner emails and payouts as drafts.
- ak-agent-browser: drive a separate browser to test and take screenshots
ak-agent-browser lets the AI open pages, click, fill forms, take screenshots and hunt bugs in a separate browser through the agent-browser CLI, without touching your everyday Chrome.
- ak-agentize: wrap existing code as a CLI or MCP for agents
ak-agentize reads your existing code, picks a CLI, an MCP server or both, and wraps it so AI agents can call it without changing current behavior. It only publishes when you ask.
- ak-ai-artist: make images from a library of 129 prompt templates
ak-ai-artist finds the closest of 129 prompt templates, puts your subject in, and generates the image with Nano Banana in 3 modes: follow a template, blend templates, or remix.
- ak-ai-multimodal: read and generate images, audio and video
ak-ai-multimodal lets the running model read images and documents first, calls the Multix CLI only to generate images, video, voice, music or transcripts, and stops when the right API key is missing.
- ak-analytics: read marketing data correctly before concluding
ak-analytics settles the source and time frame, defines every metric, keeps missing cells as unknown, compares equivalent periods, then produces a report with charts, prioritized recommendations and owned next steps.
- ak-ask: architecture advice grounded in your real repo
ak-ask answers technical and architecture questions like a system architect, based on your real docs and code. It only advises; it does not change code.
- ak-assets-organizing: settle where marketing files go and what they are called
ak-assets-organizing settles the path and file name for marketing content in order: the place you name, the repo's convention, then the fallback assets/ tree, and tidies messy folders without renaming existing files on its own.
- ak-atoz: hand off a task once, it runs to done
ak-atoz settles the scope with you through an interview, asks for everything it needs in one pass, then runs on its own until every acceptance item is verified.
- ak-autoresearch: route iterative work to the right skill
ak-autoresearch reads a goal you want to improve over several rounds, picks the skill that owns that result, and keeps the shared conditions: a metric, a round limit, a stop condition.
- ak-backend-development: write and fix backend code in your existing stack
ak-backend-development reads your project to learn the real stack, then writes endpoints, services, data access and tests in Node.js, Python or Go without switching framework, auth or database.
- ak-banner-design: design banners at the right platform size
ak-banner-design builds banners, cover images, heroes and print banners at the right platform size. It reuses your brief, asks only for what is missing, and checks text, safe zones and logo before delivery.
- ak-better-auth: add Better Auth sign-in to a TypeScript app
ak-better-auth inspects your project first, then sets up or extends Better Auth sign-in: email/password, OAuth, 2FA, passkeys, sessions, organizations. When done, it checks the failure paths too.
- ak-bootstrap: build a new project up to a working version
ak-bootstrap settles the contract for a new project, then hands it to ak-cook to build and test, ak-code-review to catch bugs, ak-enhance-ux-ax for the web part, and ships it if you ask.
- ak-brand: build, update, review and apply your brand
ak-brand creates a brand identity, updates exactly the part you ask for, reviews assets against the standard and applies the brand to a deliverable. It always edits the source of truth and reports the files it changed.
- ak-bro: restate the last answer in plain words
ak-bro rewrites the AI's last message so it is short, clear and free of jargon. It keeps the meaning, warnings and doubts; it adds no new ideas and runs nothing.
- ak-campaign: plan, track and measure a marketing campaign
ak-campaign builds a campaign kit, reports progress with the age of its numbers, writes performance reports and drafts email sequences. Planning never turns on an ad.
- ak-chrome-profile: let your agent work in your real Chrome
ak-chrome-profile opens a URL in the right real Chrome profile, then attaches your agent to that exact tab through Chrome DevTools MCP, so it can read and act with your signed-in accounts.
- ak-cip-design: build brand identity mockups
ak-cip-design writes an identity brief, then renders mockups of business cards, signage and uniforms for exactly the items you name, using your real logo when the brand has one.
- ak-code-review: find bugs and regressions in code, with evidence
ak-code-review reviews uncommitted changes, a PR, a commit or the whole codebase, checks the spec first and then quality, and backs every finding with a file, line and evidence.
- ak-codex-goal: write long-running Codex goals with a checkable stop point
ak-codex-goal checks whether a long task suits Codex /goal, then writes a goal contract with a check command after each milestone and a measurable stop point. It only creates a goal when you ask for one.
- ak-coding-level: tune explanation depth to your coding experience
ak-coding-level sets your coding experience level from 0 to 5 so the AI explains things at the right depth. The level only changes how it talks, not the quality of the work.
- ak-competitor: competitor analysis and sourced comparison pages
ak-competitor analyzes a competitor's positioning, pricing, content and SEO, and writes comparison pages in 4 formats from one shared competitor profile, with every pricing and feature claim carrying its source, date and confidence.
- ak-content-hub: a local asset gallery with brand context
ak-content-hub opens a gallery on your machine that shows the files in assets/ as a thumbnail grid with a brand color and voice sidebar, for filtering, searching, previewing and editing text files, uploading nothing to the cloud.
- ak-content-marketing: pillars, editorial calendar and content audits
ak-content-marketing builds content pillars and topic clusters, an editorial calendar, blog briefs, audits of old content and repurposing of existing posts, reusing approved material and only publishing when you allow that exact channel.
- ak-context-engineering: spend less context and tokens, still finish the job
ak-context-engineering reads your runtime's measurements, separates context, total tokens, usage limits and money, then helps you checkpoint at the right time and pick models by real results.
- ak-cook: turn a clear request into checked code
ak-cook reads the relevant code, writes just enough plan, edits the code, runs risk-based checks, reviews itself, then closes out the plan and docs. By default it runs straight through and stops only when you need to decide.
- ak-copywriting: conversion copy in your brand voice, backed by evidence
ak-copywriting writes headlines, emails, landing pages, CTAs and social posts from your brief, uses formulas such as AIDA or PAS as a frame while letting the brand voice win, and never makes up figures or customer quotes.
- ak-creativity: build a creative brief for ads and videos
ak-creativity searches a library of styles, platforms, voiceovers, music and campaign types, then returns a creative brief or a single pick. It does not make images or videos.
- ak-cti-expert: passive OSINT lookups packaged as intel reports
ak-cti-expert looks up your assets in public sources, rates confidence for each finding, then packages it all as a report. It does not scan ports or try passwords.
- ak-databases: schemas, queries and migrations for MongoDB, PostgreSQL
ak-databases handles schemas, queries, indexes, migrations and backups for the MongoDB or PostgreSQL you already run, and only calls a task done when it has evidence that fits the operation.
- ak-debug: prove the root cause before fixing a bug
ak-debug reads the error, reproduces it, tests one hypothesis at a time, and only reports done after running a verification command. After 3 failed fixes it stops to revisit the architecture.
- ak-deep-swe: score a coding model on DeepSWE
ak-deep-swe runs a model through OpenRouter on the DeepSWE task set with Pier, goes from 1 task to a small sample, runs the full set only when you agree, then reports the score with its measurement conditions.
- ak-deploy: ship your app and check it really runs
ak-deploy finds the right platform and environment, runs the deploy, checks for proof the app really runs, then records how to roll back.
- ak-design: make logos, brand identity, posters and banners
ak-design makes logos, brand identity, posters, banners and social images. It shows its assumptions first, works in rounds and scores the final piece; it does not write UI code.
- ak-design-system: build tokens, component specs and on-brand slides
ak-design-system organizes colors, spacing and font sizes into three-layer tokens, writes state specs for components and builds HTML slides that use only the project's tokens.
- ak-devops: deploy and run infrastructure on the platform you already use
ak-devops reads the deploy config in your repo, keeps your current provider and works at the level you pick: local, preview or production, on Cloudflare, Docker, GCP, Kubernetes.
- ak-diagram: interactive system diagrams, laid out for you
ak-diagram writes your system description as typed JSON IR, then compiles it offline into an interactive HTML reader, an SVG or a fragment. You never place coordinates; the layout is automatic.
- ak-docs: keep docs and agent context files accurate and lean
ak-docs reads your repo, then writes, updates, summarizes or trims docs and CLAUDE.md/AGENTS.md, keeping only what the code cannot say: why and where.
- ak-docs-seeker: find current library docs through llms.txt
ak-docs-seeker finds current docs for a library or framework through llms.txt on context7.com, reads an official link directly if you have one, and falls back to reading the repo when there is no llms.txt.
- ak-document-skills: read, create and edit Word, PDF, PowerPoint, Excel
ak-document-skills picks the right process for docx, pdf, pptx and xlsx, reads, creates or edits the file, keeps the template and untouched parts as they were, and checks the output before handing it over.
- ak-elevenlabs: write scripts and make ElevenLabs voice audio
ak-elevenlabs writes and directs voice scripts for the ElevenLabs model, then makes speech, multi-speaker dialogue, consented voice clones or sound effects, and listens back to the real file.
- ak-email: newsletters, sequences and automated flows, drafts by default
ak-email writes newsletters, cold emails, email sequences and automated flow designs from what you already said, saves them as draft files by default, and only sends or schedules through an authorized tool, then checks the receipt.
- ak-enhance-ux-ax: review the experience for people and for AI
ak-enhance-ux-ax reviews a site or app for both users (UX) and AI and answer engines (AX), then writes a report with ranked proposals and a DONE contract. By default it only writes the report.
- ak-excalidraw: editable Excalidraw diagrams from a description or code
ak-excalidraw draws architecture, data flow and workflow diagrams on a live Excalidraw canvas or as a .excalidraw file with a PNG, reads your repo itself and reviews the image before handing it over.
- ak-explain: explain code, systems and errors with evidence
ak-explain explains concepts, code, architecture or errors with a 5-part model, keeping what it checked apart from what it infers. It can use analogies or build an HTML page with diagrams; it does not fix code.
- ak-fable-thinking: make the AI check evidence before it answers
ak-fable-thinking is a set of checks for the AI: every claim names its source, output constraints are checked by machine, and the answer leads with the result. Use it when being right beats being fast.
- ak-filmmaking: make long AI films as a production graph
ak-filmmaking builds 5 to 30 minute AI films as a production graph: script, shots, identity references, master dialogue, animatic, tiered renders, per-shot QC and a final cut with separate stems.
- ak-find-skills: find the right skill, installed or new
ak-find-skills picks one main skill from the AgentKit skills you have installed, or finds an outside skill with the Skills CLI and installs it only when you agree.
- ak-fix: find the real cause of a bug and fix it at the root
ak-fix proves the root cause before fixing, fixes the right spot, reruns the exact command that failed to prove it is gone, and stops after 3 failed fixes.
- ak-folder-context: write local context for one subfolder
ak-folder-context reads one subfolder, writes a short CLAUDE.md that holds only what has evidence, and adds an AGENTS.md that points to it. If the folder needs nothing, it creates nothing.
- ak-form-cro: get more qualified form completions
ak-form-cro reviews a contact, demo, quote, survey or checkout form and returns a prioritized list of changes, each with its evidence and how to measure it.
- ak-free-tool-strategy: plan and score a free marketing tool
ak-free-tool-strategy helps you pick a free tool idea that brings in leads, scores it on 8 factors, checks whether it pays off, then writes a strategy doc. It does not write code.
- ak-frontend-design: build tasteful UI and grade the real page
ak-frontend-design writes a 7-line design brief before any code, builds from tokens, then grades the rendered page at 3 screen sizes. Given a live URL, it audits the UX and returns a report with screenshots as evidence.
- ak-frontend-development: React/TypeScript that follows your project's stack
ak-frontend-development reads your project first, writes React/TypeScript components and pages with the router, libraries and styling already in use, then checks build, types and interactions.
- ak-gamification-marketing: design gamification around a useful behavior
ak-gamification-marketing picks game mechanics from one useful behavior and evidence about your users, then builds the campaign, the implementation and the measurement, always next to a no-game option.
- ak-git: commit, push, open prs and merge by the book
ak-git packs commit, push, opening a PR, merging a branch, merging a PR then watching CI, and stacked PRs into 6 subcommands, with standard messages and a stop on secrets.
- ak-github: GitHub work through gh, with evidence and no duplicates
ak-github handles issues, PRs, CI and repo admin through gh: it reads the real state first, checks for duplicates before creating, and backs every change with a URL and evidence.
- ak-gkg: find definitions and every usage before refactoring
ak-gkg builds a code graph with GitLab Knowledge Graph to find definitions, every usage of a symbol, and a repo map. It works best with Ruby, Java and Kotlin.
- ak-goal-warmup: prepare properly before a long /goal run
ak-goal-warmup settles an outcome contract with you, plans against it, checks what every phase still lacks, then returns Ready, Blocked or Decision required. It never starts /goal itself.
- ak-google-adk-python: build Google ADK agents on your installed SDK
ak-google-adk-python writes Google ADK agents in Python (tools, multi-agent, MCP, A2A, deploy) against the google-adk version your project already has, starting small and adding a task or eval to check it.
- ak-graphify: build a knowledge graph from code and docs
ak-graphify turns code, docs, papers and images into a saved knowledge graph, so you can ask about module relationships again and again instead of grepping file by file.
- ak-handoff: record unfinished work so a fresh agent can continue
ak-handoff writes a redacted 9-section Markdown file so a new session, model or runtime can pick up unfinished work without guessing.
- ak-help: find the installed skill or ak command for the job
ak-help reads the skills actually installed, summarizes only the ones that match your question and points to the most specific one once the job is clear. For an ak command, it reads that command's current help.
- ak-hyperframes: write short videos in HTML, render to MP4
ak-hyperframes lets you write a short video as one HTML file, then uses the version-pinned hyperframes CLI to preview, check, render to MP4 and confirm the result with ffprobe.
- ak-init: learn the marketing context, ask only what is missing
ak-init reads the existing docs of a marketing project, asks only what the repo cannot answer, then writes the context into the file that already holds each piece.
- ak-interview-docs: interview you, write your words into docs
ak-interview-docs asks you one turn at a time and writes each answer straight into your README, decision file or principles doc. The content is your words: it does not invent ideas.
- ak-issue-to-plan: turn a GitHub issue into a checked plan
ak-issue-to-plan reads a GitHub issue, scouts the code, passes the issue through an ak-brainstorm gate, and only then writes a plan, checks it and pushes it to its own branch. It stops before any code is written.
- ak-journal: turn a work session into a technical journal
ak-journal gathers the key events of a session, writes a short journal entry and saves it to plans/journals/. It only posts to social media when you ask for it explicitly.
- ak-launch-strategy: plan a product or feature launch
ak-launch-strategy works out the launch type, picks the stages and channels you need, then gives a plan with owners, dates, metrics and stop conditions. It posts and sends nothing.
- ak-llms: build llms.txt and Markdown copies of your docs
ak-llms picks your public docs pages, summarizes them, groups them and writes llms.txt per the llmstxt.org spec, plus llms-full.txt when you need full text, so AI reads your site cleanly.
- ak-logo-design: design a logo from a brand brief
ak-logo-design generates logos that follow your brief, only looks up the style or colors still open, then checks every file before handing it over. Output is a raster image, not vector.
- ak-loop: change and measure until the number improves
ak-loop runs many rounds, one small change per round: it commits, re-measures with a command that prints a number, keeps the change if it is better and reverts it if not.
- ak-markdown-novel-viewer: read long markdown in a book-like page
ak-markdown-novel-viewer runs a small server on your machine, opens a markdown file or folder in a book-like reader, returns the exact URL the server prints, and only stops the server it started.
- ak-marketing-dashboard: open a local dashboard for assets and brand
ak-marketing-dashboard starts, reuses or stops a dashboard running on your machine, with 3 areas (Assets, Brand, Settings), and only stops the processes it started.
- ak-marketing-ideas: pick a few marketing ideas that fit your product
ak-marketing-ideas searches a catalog of marketing ideas across 17 channel groups, ranks them by product, stage and resources, then gives a few ideas with test and stop rules.
- ak-marketing-planning: turn a marketing goal into a campaign plan
ak-marketing-planning reads your brand context, reuses research that is still fresh, looks up only what is missing, then writes a campaign plan folder. It only plans; it does not run the campaign.
- ak-marketing-psychology: find the psychology behind customer behavior, with a test
ak-marketing-psychology picks 2 to 3 psychological mechanisms that explain a puzzling customer behavior, each with one intervention, a metric and a harm guardrail.
- ak-marketing-research: find sourced evidence for a marketing decision
ak-marketing-research settles the scope, reuses research you already have, finds primary sources for the gaps, then writes a report where every number has a source and a date.
- ak-mcp-builder: build MCP servers so AI can use your service
ak-mcp-builder builds MCP servers on SDK v2: it finishes one tool, tests it with a real client, and only then adds more. It does not configure clients or call existing servers.
- ak-media-processing: process video, images and backgrounds from the command line
ak-media-processing picks FFmpeg for video and audio, ImageMagick for images and RMBG for background removal, keeps originals unless you allow overwriting, and checks the output before reporting done.
- ak-mermaidjs-v11: draw Mermaid v11 diagrams inside Markdown
ak-mermaidjs-v11 writes diagrams in Mermaid v11 syntax: process flows, API call sequences, schemas, project timelines. By default it only gives you a mermaid block for Markdown, and exports images only when you need them.
- ak-mintlify: edit and check Mintlify docs sites
ak-mintlify reads your existing Mintlify docs site, edits the right area, then runs the check commands the project already has. It does not install the CLI globally or scaffold a new site just to edit one page.
- ak-mobile-development: build iOS and Android apps on your existing stack
ak-mobile-development reads your real project to learn the framework, SDKs and devices, then builds iOS and Android work on that stack. It compares frameworks only when you have not picked one.
- ak-motion-design: settle motion principles before writing code
ak-motion-design picks the emotion, personality, duration, easing and order for motion, then hands back a brief. It writes no code, makes no assets and renders nothing.
- ak-motion-graphics: route video work to the right skill and check renders
ak-motion-graphics hands video and animation requests to an installed skill that can do them, only prints an install command for an outside pack when no skill has the template, and checks rendered files against spec.
- ak-motion-video: beat-synced motion graphic videos with voice-over
ak-motion-video builds 1080p motion graphic videos with HyperFrames: cuts on the beat, text timed to the voice-over, Vietnamese karaoke captions, about 70 styles, an AI voice or your own recording.
- ak-onboarding-cro: fix onboarding so new users activate sooner
ak-onboarding-cro reviews your post-signup flow, settles one activation action that predicts retention, then gives a prioritized list of changes, each with a metric and a guardrail.
- ak-orchestrate: split work into sessions with a final arbiter
ak-orchestrate runs each part of the work as an independent session in stages, saves state so you can resume, then has an arbiter check every result before reporting done.
- ak-page-builder: turn your existing components into a page builder
ak-page-builder turns your project's existing components into blocks, wires up the editor, drafts, preview, publish and agent interfaces, then creates a project operator skill for fast page edits later.
- ak-paid-ads: strategy, copy and optimization for paid ads
ak-paid-ads turns your brief and numbers into a strategy, ad copy, audiences and optimization advice. It only advises and never changes anything in the ad account itself.
- ak-payment-integration: add payments with safe webhooks
ak-payment-integration guides integrations with 9 payment providers, from VietQR to Stripe, along one shared flow and the same safety gates: only a signature-checked webhook changes an order's state.
- ak-plan: split clear work into a multi-phase plan
ak-plan reads your real code, picks the right depth, then writes a plan made of plan.md and phase files with steps and checks, ready to hand to ak-cook.
- ak-plans-kanban: open the plans kanban in your browser
ak-plans-kanban opens the /plans page of the AgentKit dashboard so you can see your plans as a kanban, timeline and progress. It starts the dashboard if needed; it edits no plan.
- ak-play: run a marketing playbook step by step, with approval gates
ak-play builds a marketing playbook from a template, tracks the state of every step, hands each ready step to the right skill or agent, and stops at gates for your approval.
- ak-predict: five experts review a change before you code
ak-predict has 5 personas analyze a large or risky change on their own, compares where they agree and clash, then gives a GO, CAUTION or STOP verdict with risks. It writes no code.
- ak-preview: view Markdown and build explainers, diagrams, slides
ak-preview opens Markdown in a browser reader, or builds an explainer, diagram, slides, diff review page, plan-versus-code check or project recap as Markdown or a self-contained HTML file.
- ak-pricing-strategy: choose billing metric, tiers and prices from data
ak-pricing-strategy proposes a billing metric, tiers and prices with the math from the data you give it, or a research plan when evidence is thin. It never publishes prices itself.
- ak-problem-solving: pick the right way to rethink when stuck
ak-problem-solving reads how you are stuck, picks 1 of 5 techniques to rethink the problem, applies it to your case and stops once there is a step you can take right away.
- ak-project-management: track plan progress and carry work across sessions
ak-project-management reads your plan files, counts the real checkboxes, updates status, writes short reports and rebuilds the task list so you can pick up in the next session.
- ak-project-organization: put files and folders in the right place, with the right names
ak-project-organization scans the folders you name, proposes a from-to table, then moves, renames and checks the result. Your repo's existing conventions always win.
- ak-react-best-practices: fix the right slow spot in React and Next.js apps
ak-react-best-practices picks, from Vercel Engineering's 45 performance rules, the rule that matches the slow spot you measured, fixes it with the rule's correct example, then compares before and after numbers.
- ak-referral-program-building: design a customer referral program
ak-referral-program-building helps you pick a reward structure, a platform or a custom build, tracking and fraud controls, then measure by customer cohort so your referral program does not lose money.
- ak-remotion: make videos in React with Remotion
ak-remotion carries 29 Remotion rule topics, reads only the part your task needs, writes frame-driven React video code, and checks the real video file instead of trusting the preview.
- ak-repomix: pack just the repo parts an AI needs into one file
ak-repomix uses Repomix to pack a repo or part of one into a single file for an AI to read, picks the smallest scope that keeps dependencies, checks tokens and secret warnings, then reports what it packed.
- ak-research: search, cross-check sources and write a technical report
ak-research settles the scope of your question, searches the web, cross-checks several independent sources, then saves a cited report with next steps. It does not write code.
- ak-research-prompt: write a research brief clear enough to hand off
ak-research-prompt writes a self-contained research brief: context, one main question tied to one decision, a source hierarchy and how to return results. It does not do the research.
- ak-retro: turn git history into a retro report
ak-retro reads one period of git history, measures pace, code health and plan progress, then writes a retro report with 3 to 5 recommendations based on real numbers.
- ak-review-pr: review, fix, post and merge GitHub PRs
ak-review-pr reads a PR's description, diff and CI, rates findings on 3 levels and gives one verdict. By default it only prints to chat; fixing, posting or merging needs a flag.
- ak-scenario: find edge cases and test scenarios before you code
ak-scenario examines a code file or a feature description across 12 dimensions, lists concrete scenarios with severity, then hands off to ak-test, ak-plan or ak-predict.
- ak-scout: find the files you need to touch in a codebase
ak-scout searches your repo and returns a list of relevant files, each with a one-line description, plus the open questions. It only searches and reads, it does not change code.
- ak-security: scan secrets, dependencies and code vulnerabilities
ak-security scans for secrets, vulnerable dependencies and dangerous code patterns, then analyzes threats with STRIDE and OWASP Top 10. By default it only reports and fixes only when you add --fix.
- ak-seo: check and improve SEO and GEO with sourced numbers
ak-seo audits sites, checks GEO, researches keywords, and writes schema and bulk page templates. Every number names its source, and missing data is marked unknown.
- ak-shader: write GLSL fragment shaders that run in your real app
ak-shader checks how your project runs shaders, then writes a GLSL fragment shader for that exact environment: procedural textures, SDF shapes, repeating patterns, animated backgrounds, verified by a real render.
- ak-ship: take a finished branch through tests and review to a PR
ak-ship takes a finished branch through tests, review, version, changelog, commit and push, then opens a PR. Add a flag and it goes on to a reviewed merge with green CI.
- ak-shopify: build Shopify apps, extensions and themes for your versions
ak-shopify decides whether the work is an app, extension or theme, reads the project's Shopify CLI and API version, writes code within the safety rails, then tests on a dev store or preview theme.
- ak-show-off: an HTML showcase page for the work you just did
ak-show-off turns a description or a markdown file into a self-contained multi-section HTML page, screenshots each section in 3 ratios, publishes the page and writes it in two languages, remembering your settings for next time.
- ak-slides: build persuasive HTML presentation decks
ak-slides builds one HTML presentation file with navigation, responsive layouts and Chart.js charts from real figures, says who the deck is for before it starts, and checks 5 criteria before handing it over.
- ak-social: draft social posts and publish when you allow it
ak-social drafts posts in each platform's format, plans a posting calendar, and when you ask, publishes or schedules for real through Typefully, Postiz or Buffer, then checks the receipt.
- ak-sowat: point out what is worth doing next
ak-sowat looks at recent work, an issue or a PR like a product owner, says what it means for users, and points out up to 3 next steps. It does not change code.
- ak-stitch: design UI with Google Stitch, then hand off to code
ak-stitch has Google Stitch draw a screen from your words, lets you review the image, then exports Tailwind HTML, a PNG and DESIGN.md for the skill that builds the code.
- ak-storage: put assets in S3 storage and check every file
ak-storage uploads, downloads, lists or syncs assets with S3-style storage such as Cloudflare R2, AWS S3 and MinIO, then reads each file back to be sure it landed. Without config, it says plainly nothing was uploaded.
- ak-sumup: recap work you just finished, no rereading
ak-sumup gathers the evidence at hand and recaps work you just finished: outcome, failures and fixes, decisions, usage, what remains. It does not edit files or carry on.
- ak-tanstack: build TanStack Start, Form and AI for your installed versions
ak-tanstack reads the TanStack packages installed in your project, then writes routes, server functions, validated forms or AI chat against those exact versions' types and checks with the project's scripts.
- ak-team: run a parallel agent team on big work
ak-team sets up an Agent Team: a lead splits the work, teammates run in parallel on their own files, and the lead merges and checks the result before reporting done.
- ak-tech-graph: draw publish-grade SVG and PNG technical diagrams
ak-tech-graph gathers components and relations from a description or code, picks a diagram type and 1 of 7 styles, draws SVG, exports a 1920px PNG, then reviews the image to catch overlaps.
- ak-test: run the right tests, report what was checked
ak-test picks the smallest trustworthy check to prove a change works, investigates every failing test and states exactly what was checked. It also creates, optimizes and audits test suites.
- ak-threejs: build Three.js 3D scenes for your installed version
ak-threejs reads the project's Three.js version and renderer, looks up examples and APIs in its bundled data, writes a 3D scene for that exact version, and counts it done only when the image shows and interaction works.
- ak-ui-styling: style shadcn/ui and Tailwind apps, keep the design
ak-ui-styling adds components, changes themes, turns on dark mode, fixes accessibility and builds responsive layouts for apps on shadcn/ui and Tailwind. It reads the current setup first and changes only what you ask.
- ak-ui-ux-pro-max: pick style, color, fonts and review UX
ak-ui-ux-pro-max searches a design dataset to pick style, color and fonts for web and mobile interfaces, then builds the UI or reviews its UX against 10 rule groups.
- ak-vibe: take an issue or request to a ready PR
ak-vibe chains worktree, plan, cook or fix, code-review, ship and review-pr into one pipeline that takes an issue or request to a reviewed PR labeled ready.
- ak-video: script, storyboard or real video, just the stage you need
ak-video does exactly the marketing video stage you need: script, storyboard, rendered video or platform repurposing, reuses approved work and only reports done after checking the real file.
- ak-watzup: sum up work in flight and rank next steps
ak-watzup scans branches, worktrees, unfinished plans and roadmaps, then writes a short report with 5 to 6 ranked next steps. It only reads and changes nothing in the repo.
- ak-web-frameworks: build Next.js and Turborepo on the stack you have
ak-web-frameworks is a knowledge skill for Next.js and Turborepo: it reads the project config first, opens the right guide for rendering, data, optimization, SEO/GEO or monorepos, then checks with build, curl and a browser.
- ak-web-testing: pick the right web test tool and prove it with real runs
ak-web-testing checks which runner your repo uses, picks exactly one tool for your web testing job (Vitest, Playwright, k6, axe, Lighthouse) and draws conclusions only from real runs.
- ak-webmcp: turn your existing pages and forms into WebMCP tools
ak-webmcp designs and registers WebMCP tools through document.modelContext for your existing pages, forms and features, adds safety annotations, then checks them with the validator, Inspector and evals.
- ak-worktree: open a separate worktree for each task
ak-worktree creates a git worktree with its own branch for each task, names the branch, picks the base branch and folder, then hands off to ak-cook or ak-fix. It also checks status and cleans up old worktrees.
- ak-write: write blogs and copy, fix CRO, review before posting
ak-write is the single entry point for marketing writing: blog posts, quick or thorough copy, formula copy, CRO fixes, scoring, and preparing a ready-to-post version.
- ak-xia: bring a feature from another repo into your project
ak-xia reads a feature in another repo, maps it against your code, challenges the assumptions, then hands ak-plan a plan with rollback. The --compare mode only writes a comparison report.
- ak-youtube: turn YouTube videos into blog posts, infographics and social posts
ak-youtube pulls a YouTube video's captions, comments and info through the VidCap API, then rewrites them into a blog post, an infographic or social posts, with quotes tied to timestamps.
- ak-youtube-thumbnail-design: readable YouTube thumbnails that meet the specs
ak-youtube-thumbnail-design makes 1280x720 YouTube thumbnails from your video brief, brand and approved images. Each variant comes with its design reasoning and one CTR hypothesis to test.
- ak-brainstorm: settle what to build and how before you code
ak-brainstorm reads your real code, settles the outcome, constraints, non-goals and acceptance, compares up to 3 approaches, then hands off to the next skill.
- From ak-plan to ak-cook: Hand Off Work So the Agent Does It Right
Turn a title-search request into a clear handoff with scope and verification criteria, using a six-stage demo for new AgentKit users.
- Is Pi right for you? Pi compared with Oh-my-pi and Claude Code
Pi is a minimal coding agent: four default tools, and no subagents, plan mode or per-tool-call approval in the core. Criteria for deciding whether to use Pi, and how it differs from Oh-my-pi (OMP) and Claude Code.
- Your first day with Pi: from install to a real task in your repo
Install Pi, decide on project trust, sign in to a model, hand it a small task in your repo and resume the session the next day. Each step has a check, plus how to run Pi in Docker for safety.
- Self-hosting Cloud Harness MCP for ChatGPT (1/3): Overview and prerequisites
Part 1 of 3: what Cloud Harness MCP is, why it needs a VPS, Cloudflare Access and your own domain, the prerequisites, and the end-to-end order.
- Self-hosting Cloud Harness MCP for ChatGPT (2/3): Deploy, Cloudflare Access and the ChatGPT connector
Part 2 of 3: DNS, the VPS deploy, Cloudflare Access, Managed OAuth, the AUTH_MODE cutover, the ChatGPT connector and the GitHub App. Includes real screenshots with identifying data blurred.
- Self-hosting Cloud Harness MCP for ChatGPT (3/3): Hand the setup to an AI agent
Part 3 of 3: how to split the Cloud Harness MCP setup between you and an AI agent, a sample prompt for the VPS work, and delegation patterns to avoid.
- Advisor in OMP, Claude Code, and AgentKit: three kinds of second opinion
Advisor in Oh-my-pi (OMP) is a review loop running beside the main agent for the whole session. This guide takes apart its severity levels and note delivery, then compares it with Advisor in Claude Code and AgentKit.
- Archify with AI agents: choosing and using ten diagram types
How to choose and use the ten diagram types in Archify 3.1 with an AI agent, based on the reader's question and the input source (code, a requirements brief, logs, traces), with the Cloud Bakery example and ready-to-use prompts.
- OMP Plan mode vs AgentKit ak-plan: when to use which
Oh-my-pi (OMP) /plan locks the session read-only, AgentKit ak-plan writes the plan to files. When to use which, why not to nest the two commands, and how to run them in sequence.
- Model routing in Oh-my-pi (OMP): roles, fallbacks, and how to tell which model is running
Roles in Oh-my-pi (OMP) are routing aliases, not agents. How to assign models by workload, set fallbacks for when a model fails, and read back which model is actually running in a session.
- Run Pi and Oh-my-pi (OMP) side by side with three isolated Mise profiles
Three commands, pi-dev, pi-ak, and pi-omp, each with its own settings, login, and sessions through Mise. Try Pi, Oh-my-pi (OMP), or AgentKit without breaking the setup you use every day.
- The cmux skill: coordinate several agents in your terminal without stealing focus
The cmux skill routes every request to the right companion skill, checks before it changes anything, opens every new pane in the background, and helps you coordinate several agents across a product lifecycle.
- Use the Codex app to coordinate several agents in parallel
Open a coordinator thread, pin it, and let it give each task its own worktree thread that reports status in a table. Step-by-step from project setup to cleanup, with prompts to paste into the app.
- Herdr: a runtime for coding agents you can drive from your phone
Herdr keeps coding agents running on your machine even after you close the terminal. This guide covers install, a tidy config, and driving your agents from your phone over SSH and Tailscale.
- Herdr part 2: let agents orchestrate other agents in parallel
Let agents drive agents through the Herdr CLI: check the target, send and wait, read results, one worktree per writing agent, plus sample workflows, the herdr-agent skill and the Vercel Sandbox plugin.
- UI Craft: from install to your first prompt
UI Craft helps an agent keep one design direction from idea to code and screenshot checks. This guide covers install, bundle validation, modes, prompts, and safe upgrades.
- Fable 5 and the unknowns problem: fix the map before you code
The stronger the model, the costlier a vague prompt. How I put Fable 5 on finding unknowns and making decisions, place risk-scaled gates around ak-cook, and hand execution to cheaper models.
- Guardrails in Claude Code and AgentKit (2/2): each guard layer, its gaps and good habits
Part 2/2: each AgentKit guard layer taken apart, from scout-block, privacy-block and simplify-gate to hard-gates and guard skills, with the known gaps, habits worth keeping and a quick glossary.
- Guardrails in Claude Code and AgentKit (1/2): the hook mechanism and config layers
Part 1/2: a prompt instruction is only a suggestion. A real guardrail is a hook in the Claude Code harness that blocks a tool call with exit 2. The hook mechanism, AgentKit's seven guardrail groups and its settings layers.
- When to turn on --deep and --tdd for /ak-plan
Use --deep when scope is wide, add --tdd when old behavior must survive, and turn on both only when both risks apply. A billing extraction example shows when each option is worth the cost.
- Getting started with AgentKit for Claude Code: setup and first commands
AgentKit (formerly ClaudeKit) gives Claude Code a brainstorm, plan, cook, test workflow. How to install it, tune CLAUDE.md and rules, the core commands, and the habits that matter for beginners.