Sonith ManakkatThere are thousands of AI tools, and the list changes every month, so a list of names goes out of date quickly. A better map is depth: how much a tool can do on its own, and how much of your work and data it can touch. This article is organised by skill depth and by the job you want done, not by product. Facts about Claude, Claude Code and MCP were checked against official pages on 2026-10-21 and are listed at the end. Opinions are labelled.
Level 1: beginner, above the water
- Ask and draftA chat assistant answers questions and writes first drafts
- ResearchFind and summarise material, with sources you then open yourself
- NotesTurn messy notes into a tidy summary
- SlidesDraft an outline, then design it by hand
- DesignGenerate rough ideas for a poster or logo
Level 2: intermediate, just under the surface
- VoiceDictate, or have text read aloud
- Video and avatarsShort clips from a script
- RepurposingOne article becomes a summary, a post and a checklist
- AutomationConnect two apps so a step happens by itself
- No-code app buildersDescribe a simple tool and get a working draft
- Prompt libraries and projectsReusable instructions and background files
An automation you can picture
- 1TriggerA new form entry, email or file arrives
- 2AI stepSummarise it or draft a reply
- 3Human approvalYou read it and approve, edit or reject
- 4ActionThe reply is sent or the record is saved
Takeaway. Put the approval step before the action, especially when the action cannot be undone.
How to choose a tool
- 1Do you just need an answer or a draft?Yes Use a chat assistant and check the resultNo keep going
- 2Is it the same task every week?Yes Save the prompt, or automate it with an approval stepNo keep going
- 3Does it need your files or tools?Yes Connect them with limited, read-only access firstNo keep going
- 4Must it act across several steps on its own?Yes Use an agent with permissions, checks and a way to undoNo keep going
- 5OtherwiseStart with a chat assistant and learn what it gets wrong
Takeaway. Go only as deep as the job needs.
The everyday loop
Level 3: advanced, deep water
- Coding agentsTools that read your files, run commands and edit code
- Tool connectionsMCP lets AI applications reach files, databases and services
- Workflow platformsChains of steps that run on a schedule or a trigger
- Scraping and data pipelinesCollect, clean and use data at scale
- Agent frameworksCode that lets AI plan, call tools and loop
- Backends and paymentsReal users, real data and real money
What an agent actually does
What you let AI touch
- It only readsAnswers from what you paste in
- It draftsYou copy and decide
- It edits with approvalYou see each change first
- It uses connected toolsStart read-only, approve writes
- It acts within limitsSmall scope, logs, a way to undo
Cost and control
| Depth | Effort to set up | How much you control | What can go wrong |
|---|---|---|---|
| Beginner chat and writing | Very low | High: you copy and decide | Wrong or invented facts |
| Intermediate automation | Low to medium | Medium: you design the steps | A wrong action runs repeatedly |
| Advanced agents and tools | Medium to high | Depends on your limits | Changes to files or systems you did not intend |
| Backends and payments | High | Needs real engineering | Data leaks, charges, outages |
Expert viewEvaluations, guardrails and cost control
Once something runs without you, add three things. An evaluation is a small set of example tasks with known good answers that you rerun when you change a prompt or a model. A guardrail is a rule outside the model, such as an allowed-actions list, a limit on spending, or an approval step. Cost control means measuring usage and capping it. Claude Code shows token use with a usage command and documents ways to reduce it. Build these before scaling, not after a surprise.
Expert viewScraping and data pipelines
Gathering data from websites raises questions of terms of service, privacy law and fairness to the sites you load. My view: check a site's rules and use an official data feed when one exists, keep personal data out of pipelines unless you have a clear reason, and log what you collect. Treat anything fetched from the web as untrusted text, because it can contain instructions meant to steer an AI.
Expert viewAgent frameworks, backends and payments
An agent framework is code that lets a model plan, call tools and loop. It is useful when you need custom behaviour, but it adds moving parts: your own tools, your own limits and your own tests. Backends and payments mean real users and real money, so keep keys out of code and prompts, test with sandbox or test modes, and have an experienced person review sign-in, data storage and payment code before launch. See Vibe coding explained for a safety checklist.
A 30-day plan for a beginner
- Week 1: one chat tool
Use it daily for drafts and questions, and check every fact
- Week 2: save your prompts
Keep three prompts you reuse, and try a project or notes
- Week 3: one automation
Pick a weekly chore and add a human approval step
- Week 4: try one deeper thing
Read-only, on a copy, with a way to undo
The goal is not to use the deepest tool. It is to use the shallowest one that does the job well.
Frequently asked questions
Which tool is best?
There is no honest answer without testing on your own tasks, and products change quickly. Pick by the job and the depth you need.
Do I need to code?
Not for levels 1 and 2. Level 3 is easier with some engineering knowledge, and payments and user data need real expertise.
Is deeper always better?
No. Deeper tools can do more and break more. Use the shallowest tool that does the job.
Where do I start?
One chat assistant, one weekly habit and the habit of checking its answers.