This skill should be used when the user asks for "ADHD output", "fewer output tokens", "short numbered steps", "limited working memory formatting", or explicitly invokes "adhd-output-style".
Format every response for a reader with limited working memory who needs
low-friction starts and visible progress, while still teaching. Apply to all
interactions in the current task.
Structure (ADHD)
Open with the actionable step or the answer, not context or setup.
Break multi-step work into numbered lists, one action per step.
End with a single next action that takes under two minutes.
Keep secondary issues separate; do not bundle them into the main answer.
Restate progress each turn (e.g. "step 3 of 5"); assume prior context is lost.
Use concrete time estimates ("~2 min", "3 files"), never vague ones.
State what now works in plain terms instead of burying it in a recap.
Describe errors factually: cause, then fix. No alarmed language.
Cap lists at five items; split longer ones into priority tiers.
Cut preambles, recaps, and closing pleasantries. Start at the answer, stop when done.
Exceptions: give full walkthroughs when asked; confirm before destructive
actions; pause with a diagnostic question after repeated failed debugging;
ask one clarifying question on genuine ambiguity before proceeding.
Education (Explanatory)
Before and after writing code, add a short educational note using this block:
Put depth here, not in the main answer. Prefer insights specific to this
codebase or the code just written over general programming concepts. Cap at
three points so the block stays scannable. The rest of the response stays terse.
Battle-tested Claude Code, Claude Desktop, OpenAI Codex, and Cursor setup with skills, commands, hooks, subagents, and MCP servers, plus Kimi, MiniMax, and GLM API support.
"They make wrong assumptions on your behalf and just run along with them without checking. They don't manage their confusion, they don't seek clarifications, they don't surface inconsistencies, they don't present tradeoffs. They really like to overcomplicate code and APIs, they bloat abstractions, they don't clean up dead code after themselves." -- Andrej Karpathy
This repo's guidelines are structured to fix exactly these pitfalls.
Plugins add skills, commands, and automations to your AI coding tool. Install only what you need from the plugin list below.
Prerequisites: See INSTALL.md for setup requirements.
Claude Code
# Add marketplace (one time)
claude plugin marketplace add fcakyon/claude-codex-settings
# Install any plugin by name
claude plugin install fable-advisor@claude-settings
In Claude Code desktop, select + beside the prompt, then Plugins and Add plugin.
Codex CLI
# Add marketplace (one time)
codex plugin marketplace add fcakyon/claude-codex-settings
# Install any plugin by name
codex plugin add simplify@claude-settings
In the Codex desktop app, open Plugins, find the plugin, and select Install.
Gemini CLI
# Install any plugin by name
gemini extensions install --path ./plugins/simplify
Enter /plugin, open the marketplace, and install the plugin at user or project scope. In the Cursor app, open Customize and select Install.
Want the shared agent guidance too? .claude/CLAUDE.md is the file written for your projects. Run /claude-tools:sync-claude-md to pull it into ~/.claude/CLAUDE.md, or paste it into a project as CLAUDE.md, then point the other tools at the same file:
Run /simplify to review your staged or committed diff across four angles (reuse, simplification, efficiency, altitude) with parallel agents, then apply the cleanups. It reviews changed-code quality, not correctness bugs. It brings Claude Code's built-in /simplify to Codex, Cursor, and Gemini, which don't ship it natively, so all three get the same review and apply flow.
Skills:
simplify - review the changed diff across four angles and apply the fixes
Hooks:
guard.py - Require a completed /simplify run before each Claude Code or Codex commit, unless you explicitly ask to skip it for one commit
humanize - Write naturally and block AI buzzwords in files, commits, and messages
Before a Write or Edit saves, humanize scans the new text and blocks whatever reads like a machine wrote it, each hit paired with a plain-word swap. It reads Markdown files whole and checks comments and docstrings in code files while skipping source code and recognized code-shaped references:
3 marks: the em-dash, section sign, and stray semicolon (the en-dash stays)
53 stock words like leverage, seamlessly, vibrant, game-changing, each mapped to a plain swap
16 openers and cliches like in conclusion, a testament to, aims to bridge
10 pile-up words like crucial or significant, flagged at 3+ uses in one write
Runs on Claude Code and Gemini. Unlike a markdown rule the model can ignore, the hook checks every Write or Edit and blocks each match until it is replaced. The word list draws on Wikipedia's "Signs of AI writing".
Skills (ZIP for claude.ai, ChatGPT, Claude Code, Cursor, Codex, VS Code):
Ask a Fable 5.1 reviewer to challenge a plan or conclusion before you commit. Claude Code uses the native plugin agent, while Codex, Cursor, and Gemini call the same model through an authenticated Claude Code CLI with claude -p --model claude-fable-5-1.
Agents:
fable-advisor - Native Claude Code reviewer with recent conversation context
Skills:
fable-advisor - Routes Claude Code to its native agent and other tools to claude -p
codex-advisor - Get a second opinion from GPT-6 Astra via Codex CLI before big decisions
Claude Code
Codex CLI
Gemini CLI
claude plugin install codex-advisor@claude-settings
Ask GPT-6 Astra to challenge a plan before you build it, the cross-model check the built-in advisor cannot give you. Claude Code attaches the recent conversation, so "ask the codex advisor" is enough. Codex, Cursor, and Gemini reach the same reviewer. Needs the Codex CLI signed in.
Agents:
codex-advisor - Sends the recent conversation to GPT-6 Astra and returns the verdict verbatim
Skills:
codex-advisor - Routes Claude Code to its native agent and other tools to codex exec
adhd-output-style - Use fewer output tokens with short steps, quick insights, and clear next actions
Claude Code
Codex CLI
Gemini CLI
claude plugin install adhd-output-style@claude-settings
Cuts filler and keeps output token use down with the answer or next step first, one-action-per-step lists, concrete time estimates, a single under-two-minute next action, and short teaching notes while coding. Claude Code applies the output style automatically. Codex, Cursor, and Gemini CLI expose the same instructions as an adhd-output-style skill because they do not support plugin output styles.
Output styles:
ADHD Explanatory - Low-token ADHD formatting plus educational Insight blocks while coding
Skills (ZIP for claude.ai, Claude Code, Cursor, Codex, VS Code):
When Claude Code auto-summarizes a long session, the default summary routinely drops the highest-signal facts. This plugin tells the summarizer to keep them:
File paths under investigation so the next turn doesn't re-discover where you were
Confirmed root causes so you don't re-debug what's already solved
Open questions, metrics, and IDs that prose summaries usually round away
Findings from expensive subagent runs that took minutes to gather
Runs on every /compact (manual) and every auto compaction. Claude Code only; Codex, Cursor, and Gemini CLI don't yet expose a comparable summary hook.
Verify actions, recovery, locales, accessibility, and real integrations
The locally maintained ui-regression skill checks known workflows on local, staging, and deployed UIs, including browser extensions. Use dogfood for exploratory bug discovery.
CLI Tool:agent-browser - install via npm i -g agent-browser && agent-browser install
Official Stripe agent skills for payment integration: API selection, Connect platform setup, billing/subscriptions, Treasury, and SDK upgrades. From stripe/ai.
Skills (ZIP for claude.ai, Claude Code, Cursor, Codex, VS Code):
Voice AI agent development with the LiveKit Agents SDK, from design and live testing to simulations and production. Works with LiveKit Cloud or a self-hosted server, and looks up current docs with the lk CLI when no docs server is connected. Based on livekit/agent-skills, patched so the Cloud-only parts also cover self-hosting.
Skills (ZIP for claude.ai, Claude Code, Cursor, Codex, VS Code):
Programmatic access to OpenObserve (Cloud or self-hosted) via the documented REST API. Covers:
Auth and search: HTTP Basic auth, the search/SQL endpoint with microsecond timestamps
Streams and dashboards: stream listing/schema, dashboard CRUD, the v8 panel JSON schema
Known pitfalls: the customQuery re-aggregation bug that doubles table rows when fields.y carries an aggregationFunction
Built for AI agents: uses curl only, no SDK or CLI dependency. Pairs naturally with claude-telemetry-hooks for Claude Code usage dashboards. Reference docs are mirrored from openobserve/openobserve-docs.
Skills (ZIP for claude.ai, Claude Code, Cursor, Codex, VS Code):
Hetzner Cloud infrastructure management via the hcloud CLI. Decision trees for compute, networking, storage, DNS, and common deploy workflows. Reference docs extracted from hetznercloud/cli.
Skills (ZIP for claude.ai, Claude Code, Cursor, Codex, VS Code):
phd-skills - Hypothesis design, paper review, citation checks
Academic research toolkit from fcakyon/phd-skills. Experiment design, literature review, paper writing, citation verification, and reviewer defense for PhD workflows.
Skills:
dataset-curation - Dataset preparation and annotation guidelines
Ultralytics Platform API flows and YOLO26 training know-how, plus auto-formatting hooks for Python, JavaScript, Markdown, and Bash and a guard that blocks force-push and rebase.
Skills (ZIP for claude.ai, Claude Code, Cursor, Codex, VS Code):
Official colors, Archivo typography, logo and naming rules, tone of voice, and per-format defaults from ultralytics.com/brand for PDF/PPTX/Canva decks, DOCX documents, marketing email HTML, social posts, and web designs.
Skills (ZIP for claude.ai, Claude Code, Cursor, Codex, VS Code):
Search papers, look up DOIs, download PDFs, and follow citations with Paper Search CLI. Run /paper-search-tools:setup after install. Requires uv and Python 3.10+; no Docker or MCP server.
Pull reviewer comments from an Overleaf project, locate each in your local git-tracked *.tex files, and apply edits you review. Run setup once to paste your Overleaf session cookie; sessions slide ~5 days idle. Pairs naturally with github-dev for the commit step.
Skills:
setup - Capture and refresh the Overleaf session cookie
review-overleaf - Fetch unresolved review threads and apply edits to local .tex files
seo-skills - Audit search visibility, structured data, and site architecture
SEO workflows from vercel-labs/marketing-team-eve-template. Audit pages in evidence order, plan programmatic page sets, add valid JSON-LD, and improve hierarchy and internal linking. The schema skill includes a dependency-free Node validator for local JSON-LD checks.
Skills (ZIP for claude.ai, Claude Code, Cursor, Codex, VS Code):
Auto mode: auto permission mode with a custom autoMode classifier block in .claude/settings.json - see the auto mode config reference for what each rule section does, and run claude auto-mode defaults to print the current built-in block and allow rules
Advisor: the built-in advisor tool uses Fable 5.1. For a review outside the Claude family, use codex-advisor
Plugins: codex-advisor is enabled in both Claude Code and Codex. Claude Code also enables intelligent-compact. Standalone fable-advisor stays enabled only in Codex
For Codex CLI, use Kimi K3 through the local Responses proxy in .codex/config-kimi.toml. The recipe uses Kimi Code model ID k3. Use k3-256k in both the proxy command and config when 256K context is enough.
MiniMax M3 (1M context)
Run Claude Code with MiniMax M3 via Anthropic-compatible API - source:
Model: MiniMax-M3[1m] - 1M context window (512K guaranteed minimum, [1m] unlocks the full window)
Fast model: MiniMax-M2.7-highspeed - high-throughput variant
For Codex CLI, see the recipe at .codex/config-minimax.toml. Note that Codex requires a local Responses API proxy since MiniMax only exposes chat completions.
Model: gpt-6.1-sol with model_reasoning_effort set to "medium"
Sandbox: workspace-write with network access enabled
Auto mode: the "Approve for me" option under /approvals, set by approvals_reviewer = "auto_review". A reviewer subagent judges each escalation against the [auto_review] policy block in .codex/config.toml instead of stopping to ask you, so it trades tokens for far fewer prompts. The policy mirrors the autoMode block in .claude/settings.json
Plugins: a curated set enabled from this marketplace (simplify, github-dev, python-skills, and more)
Thinking display: point claudeCode.claudeProcessWrapper at a small wrapper that forces --thinking-display summarized, so thinking tokens still show in the extension (works around 2.1.111+ dropping the flag)
~/.local/bin/claude-wrapper (chmod +x)
#!/bin/bash
# Workaround for VS Code Claude Code extension 2.1.111+ not passing
# --thinking-display to the binary on Opus 4.7 (API default changed
# to "omitted", so thinking blocks arrive empty).
exec "$@" --thinking-display summarized
TODO
Visual demos:
Add before/after comparison slider images or GIFs for each plugin showing the value visually
Zero-MCP goal:
Replace MCP-based plugins with CLI alternatives where possible (mongodb, tavily, gcloud, azure)
New plugins/skills:
Payments: Stripe best practices, Polar billing and local dev
