If someone asks me how a normal person should understand Claude Code and other coding agents, I would not start with code.
I would start with a much simpler picture: put a pile of PDFs, screenshots, spreadsheets, meeting notes, and transcript files into a folder, then ask the agent to read them, write scripts where needed, and produce an HTML report you can actually open and inspect.
That does not sound like “programming.”
But it may be the most useful way for non-programmers to use a coding agent.
Thariq from the Claude Code team recently put it bluntly: the basic trick for using Claude Code for non-technical work is to put a bunch of files in a folder and tell it that it can write scripts and make HTML. The post got 2,192 likes. In a follow-up, he listed examples: image or video editing, finance and tax work, medical documents, paperwork, reports, and planning.
I think that framing is right.
Most people underestimate how much useful context is already sitting in their files. The problem is not always that the model is too weak. Very often, the problem is that we give it almost no working context. A one-line request in a chat box forces the model to guess. A folder full of real materials lets it work.
That difference matters.
This is not about making everyone write code
When non-programmers hear “Claude Code,” many immediately assume it is not for them.
Fair reaction. Also slightly wrong.
In this workflow, code is not the product. Code is the shovel.
Imagine you have a pile of receipts, bank exports, and invoice PDFs. What you want is not a Python script. What you want is a clean expense report. Traditionally, you would open files one by one, copy values, paste them into a spreadsheet, reconcile totals, and then write up the summary.
A coding agent can take a different path:
- inspect the folder
- write a small script to extract structured data
- clean the data
- find suspicious or missing items
- generate an HTML report
- keep the script and intermediate files so you can check the work
The important shift is not that the AI “knows Python.” The shift is that you no longer have to do every boring intermediate step by hand.
You define the goal, provide the materials, and review the output.
That is a much better division of labor.
A folder is the simplest context layer
Most AI workflows still begin in a chat box. Chat is convenient, but it is also thin.
You type: “Help me write a report.”
The model asks for background. You add a little. It asks for the target audience. You add a bit more. After a few turns, the context is still scattered, and a lot of it is probably missing.
A folder is different.
A folder is a plain, underrated context layer. You do not need a database. You do not need a knowledge graph. You just put the relevant materials in one place:
- project notes
- meeting transcripts
- screenshots
- spreadsheets
- PDFs
- saved web pages
- previous drafts
- a short brief explaining what you want
Then you tell the agent: these are the materials, do not invent facts, read first, process second, and produce something I can verify.
That is a very different starting point.
Many AI failures happen because the model has no evidence and tries to fill the gaps with fluent language. Once the files are present, the task becomes less like improvisation and more like desk work.
What a practical workflow looks like
Suppose you want to review a month of content performance.
You have:
- exports from Xiaohongshu, WeChat, and video platforms
- a few high-performing posts
- a few underperforming posts
- screenshots of comments
- your own notes
Create a folder like this:
content-review/
├── data/
│ ├── xiaohongshu.csv
│ ├── wechat.csv
│ └── video.csv
├── posts/
│ ├── good-post-1.md
│ ├── good-post-2.md
│ └── weak-post-1.md
├── screenshots/
│ └── comments.png
└── brief.md
In brief.md, write something concrete:
Analyze this month’s content performance:
1. Find the top 5 posts and their common patterns
2. Identify likely problems in weak posts
3. Suggest 3 topics to test next month
4. Generate an HTML report that a non-technical operator can read
5. Do not fabricate data. Every conclusion should trace back to a source file
Then run Claude Code in that directory.
A useful agent should inspect the files first. If there are CSVs, it may write a script to process them. If the data is messy, it should clean it. If a report is requested, it can generate HTML. If something is missing, it should say so instead of pretending the evidence exists.
That is far more reliable than asking a chat box to “review my content strategy” with no data attached.
You are no longer giving the model a wish. You are giving it a workbench.
Why HTML is a surprisingly good output
Thariq specifically mentioned “write scripts and make HTML.” That detail is not accidental.
HTML is a good output format for this kind of work.
First, agents are good at generating it. HTML, CSS, and JavaScript sit squarely inside the skill set of coding models.
Second, it is easy to inspect. You can open the file in a browser. No special software. No formatting drama.
Third, it can combine text, tables, charts, references, and collapsible details in one place. That makes it better than a raw Markdown file for many reports.
Fourth, it can preserve traceability. You can ask the agent to attach source filenames, rows, or excerpts under each major claim.
That is why I think HTML reports may become one of the most practical outputs for non-technical agent workflows. They are readable, portable, and much less fragile than a long chat transcript.
Do not romanticize it
This workflow is useful, but it is not magic.
There are a few traps worth saying out loud.
First, the messier the material, the clearer the brief must be. The agent can organize files, but it cannot decide what a “good report” means for you. You still need to define the audience, scope, constraints, and output format.
Second, be careful with sensitive data. Finance, medical records, customer information, and private company documents need privacy judgment before they go anywhere. Use local tools when possible. Redact when needed. Saving thirty minutes is not worth leaking something you should not have uploaded.
Third, keep the work auditable. Ask the agent to preserve scripts, intermediate files, and source references. A beautiful report is not enough. A believable report is the goal.
Fourth, start small. Do not begin with your most important financial audit or a complex legal document. Start with public materials, a small content review, a trip plan, or a low-risk research task. Learn the workflow before trusting it.
The fastest way to ruin an agent workflow is to expect full automation on day one.
A small template to try
If you want to test this, start with a tiny folder structure:
my-ai-report/
├── materials/ # source files
├── output/ # generated results
└── brief.md # task description
Put this in brief.md:
Goal: turn the files in materials into an HTML report.
Requirements:
1. First list the files you found
2. Identify which files are useful and which are missing context
3. If spreadsheets need processing, write scripts and keep them
4. The report must include: summary, key findings, evidence, and recommendations
5. Every key conclusion should cite the source file
6. Mark uncertainty clearly. Do not invent missing facts
That is enough to begin.
The point is not to discover a magic prompt. The point is to package a task so the agent has materials, instructions, and a concrete deliverable.
That is when it stops floating.
The real shift is who does the middle work
I increasingly think the next useful phase of AI is not “give me the perfect answer.” It is “do the middle work for me.”
Extract the data. Normalize the format. Compare the files. Make the chart. Build the report. Check what is missing.
A lot of modern work is made of these small, dull steps. They take time, but they are not where human judgment is most valuable.
Claude Code and similar coding agents are good candidates for that middle layer. Humans still decide what matters. Humans still check whether the conclusion is believable. Humans still choose what to do next.
But humans should not have to copy, paste, rename, reformat, and assemble every report by hand forever.
That is the practical promise here.
If the old AI chat box felt like an intern who could talk, the folder-plus-script-plus-HTML workflow feels like giving that intern a desk, a stack of documents, and a specific assignment.
It will not be perfect every time.
But at least it starts to look like work.