Install a real AI agent, finish one real task, and feel "AI can actually do things for me."
Already using an agent? Skip ahead β go straight to the diagnosis
Web ChatGPT
Chats only β gives advice, can't actually do things
A real agent
Actually gets the work done
Most of the 70% gap between you and SOTA isn't a technique problem β it's that you haven't crossed the "install an agent" threshold. Install it, use it once, and you'll jump to mid-tier in the first week.
Only tools with full MCP + Skill support β install one and you can plug LoreJump's diagnosis Skill and SOTA knowledge source straight in
Claude Code
Anthropic
Coding / CLI
Fits: Technical users / terminal-comfortable
Doesn't fit: Non-technical users (see Cowork below)
30 seconds
β Harness framework β swap models: when your quota gets tight, point ANTHROPIC_BASE_URL at z.ai and run GLM Coding to save money
Install docs βCodex CLI
OpenAI
Coding / CLI / ChatGPT users
Fits: ChatGPT Plus users / second-opinion agent / cross-model verification
Doesn't fit: No paid plan (see Gemini CLI)
30 seconds
β No conflict with Claude Code (writes to .codex/skills/) β can coexist in the same repo for cross-check
GitHub βGemini CLI
Coding / CLI / big free tier
Fits: No paid plan / heavy basic-task load / repetitive template generation
Doesn't fit: Hard tasks (multi-step reasoning / tool-call-heavy β use Claude / Codex)
5 minutes
β Completely free (just a Google account) β fine for basic tasks, upgrade for harder ones
GitHub βCursor
Anysphere
Coding / IDE
Fits: Coders / VS Code crowd
Doesn't fit: Non-coders
5 minutes
Download βClaude Cowork
Anthropic
General knowledge work / non-coders
Fits: Research / analysis / operations / legal / finance
Doesn't fit: Heavy coding (use Claude Code or Cursor)
Desktop app
β Requires a paid Claude plan ($20/mo and up)
Get access βTrae
ByteDance
Coding / Cursor-style
Fits: Coders / China-region Cursor alternative
Doesn't fit: Need to run commands in a terminal (not a CLI)
5 minutes
Download βQoderWork
Alibaba
General knowledge work / China-region desktop agent
Fits: Research / analysis / operations / legal / finance / non-coders
Doesn't fit: Heavy coding (use Trae or the overseas CLIs above)
5 minutes (desktop app / Win+Mac)
β LoreJump install command: lorejump install --tool=qoder (writes to ~/.qoder/skills/* + ~/.qoder.json)
Download βCherry Studio
CherryHQ (open source)
Multi-model workspace / China-region open source
Fits: Cross-model comparison / privacy-sensitive local runs / heavy knowledge management
Doesn't fit: Want one-click install with no model-key setup (use QoderWork)
5 minutes (Win+Mac+Linux)
β LoreJump skill: lorejump install --tool=claude-code --scope=project; add MCP in the GUI panel as https://mcp.lorejump.com/mcp
Download βPick one of three β for your first time, A is recommended (easiest to feel "it's actually doing things")
Your prompt
Please clean up the files in ~/Downloads (or your target folder). Sort them into five subfolders β Documents / Images / Videos / Installers / Other β and rename each file to something meaningful based on its content. List the plan first for me to confirm, then execute.
What the agent does
Reads the directory β generates a categorization plan by type + content β waits for your confirmation β executes mv operations
Done signal
Original folder cleared / files sorted into 5 subfolders / filenames readable
Your prompt
Please read this file [drop a recent report / PDF / long email]. Output: 1) a 200-word structured summary; 2) 3 places where data conflicts or is inconsistent; 3) 3 improvement suggestions (on depth / data backing / readability). Each suggestion must cite a specific paragraph or number.
What the agent does
Reads the file β parses structure β produces summary + consistency check + concrete suggestions
Done signal
You get three structured outputs / at least 1 suggestion you hadn't considered / references specific source paragraphs
Your prompt
Here's what I did this week (scattered): [paste your notes / email subjects / task list β unsorted is fine]. Turn it into a status report with sections: Done this week / In progress / Next week / Blockers and asks. 3-5 bullets per section, concise β no filler.
What the agent does
Parses notes β categorizes by timeline and priority β generates the 4-section structure β you review
Done signal
You get a status-report draft ready to send to your lead / fewer than 5 local edits needed
Now that you've felt "AI can do things," benchmark yourself against SOTA
Come back for the diagnosis β see your gap to SOTAOr see the full SOTA for your group first: