Agent Framework

Liberal arts majors can hand their work to AI without writing code: designing your own Agent framework

My approach is simple: use files and folders to organize, and use natural language to spell out the rules clearly. It is the same idea as writing a script for actors to perform, except the script is now the rules and the actors are the AI.

Published 2026-04-20 | Last updated 2026-04-20

What is this article talking about?

Engineers look especially impressive at using AI, and many people assume it's because they understand programming. The real key is elsewhere: code is highly structured text, and AI reads it especially smoothly. Flip that around and it becomes the opportunity for people from a humanities background: as long as you organize your documents and knowledge to be as structured as code, AI can read and apply them well. And you already know the tools for that: folders handle categorization, and documents spell out the rules in natural language, and the AI follows them once it has them. This article gathers the full content of a general-education class, from Harness Engineering to LLM Wiki and Tag Wiki.

Who this is for

· Content workers, teachers, consultants, and anyone with a heavy document workload
· People who want AI to take over their work but want to close the tab the moment they see the word "engineering"
· People already using ChatGPT or Claude who want to move toward Agents and a personal work system

What you will take away

· The three things a non-engineer can do within Harness Engineering
· Three concrete ways to help AI understand the relationships between your documents (timestamps, link syntax, cross-system references)
· How to build a personal profile, anthropomorphic skill packages, and three kinds of journals

The whole class in one sentenceAI tools are smart now, so let them learn you instead of wearing yourself out learning them. Build one document that clearly describes your system, your processes, and how you think, and let AI read it. Whichever AI you're using today, that's the one that reads it.

Harness Engineering: three things a non-engineer can do

A harness is a set of reins, and Harness Engineering is, at heart, about steering the AI. OpenAI's 2026 definition of an Agent boils down to a formula: an Agent equals the AI model (the brain) plus the engineering by which we steer that brain (Harness Engineering).

Which company's brain it is, neither you nor I get to decide, and there's no keeping up with all of them. Steering, the other half, is what we can do. And most of that work needs no code at all:

First itemDesign Environment

Plan the folder structure and file placement logic so that AI will know where things are as soon as it comes in.

Second itemDefine intent

Use a document to clearly describe the purpose of each folder, the rules and restricted areas of each process.

The third itemCreate a loop

Let AI automatically run along the written path, check and write back the rules after running.

The quickest way to grasp it is through scriptwriting: designing the environment is building the set, defining intent is writing the script, and building the loop is having the actors follow that script through the scene. People from a humanities background already train in exactly this: reading, categorizing, interpreting, and laying out context clearly. The only difference is that you used to write for humans; now you write one more copy for the AI.

Keep one underlying fact in mind: if you don't tell the AI your special logic, it will always follow the default. Given three documents A, B, and C, the AI defaults to reading them in ABC order; if your actual process goes A, then C, and only after C back to B, you have to write that down in advance for it to run ACB. AI keeps getting smarter, but your special logic will always be yours to define.

Loop Engineering: do it once, check, write the rule back

Harness Engineering handles "how to get AI to work within controllable bounds," while Loop Engineering handles "how to make it do a little better each time" (the full method is in What Is Loop Engineering; for how the two fit together, see Loop Guardrails). Do it once, check it, write the experience back into a rule, and make fewer mistakes next time. Long-term work runs on this cycle; a single prompt can't carry it.

The most practical tool for writing rules back is three kinds of journals. Everything else you can have AI organize, but the journals only you can keep, because AI can't read your mind.

  • Work log: every time you finish a process, especially one that took several rounds to get right, your next instruction is "write up what we just did as a work log." Next time you do something similar, have the AI read the previous log first. Even if the AI gets it right in one go this time, still write it up, because AI has an element of randomness, and smooth this time doesn't mean smooth next time.
  • Insight journal: when you notice an insight that cuts across systems or documents, write it up on its own rather than stuffing it into the original document. That keeps the logic clean when you look at the original system, and gives you somewhere to look when you want to compare across systems.
  • Mood journal: record your emotional state, the points of inner friction, and where you got stuck. Emotional cost is a cost too. I especially recommend this one for solo founders.
Work log example: Originally I thought the process was ABC, but this time when an unexpected situation occurred, it actually became ACB. If you encounter this situation in the future, remember that the process is ACB. (The next time the AI reads it, it will know how to deal with it. There is no need to teach it again.)

LLM Wiki: Knowledge map for model reading

The core of LLM Wiki proposed by Karpathy is to allow the knowledge base to be continuously accumulated, automatically linked, and automatically updated. AI needs a readable, referenceable, and traceable knowledge structure to know what are the core concepts, what are the related concepts, and how different pages are connected together.

Brought down to a personal knowledge base, there are three concrete ways to write it:

1. Timestamp (prompt): Add YYYY-MM-DD-HHMM to the file name, and the AI will know the order when it sees the time. 2. Connection syntax (specified, more precise than timestamp): Simply write "A followed by C" in the document. 3. Cross-system reference: Write at the beginning of file B "Please refer to C in another system. The underlying logic is almost the same. Can be used as a reference between systems. "

A good practice is to write both timestamp and link syntax. When a new document is added to the knowledge base, AI can automatically determine the association and establish a link. In special cases, it can be manually filled in. This is the knowledge graph: directional links between documents, allowing the knowledge base to be scaled up while retaining room for manual intervention.

Tag Wiki: Personal work semantic system

LLM Wiki manages the links between documents, while my own Tag Wiki is more focused on personal work semantics: using tags to connect data classification, task scenarios, roles, processes and output forms. If the labels are clearly defined, AI will know which workflow a certain note should enter.

The minimum starting point for novices is to first define a few common tags, such as data, process, task, finished product, and pending. Only after the labels are stable can the Agent help you organize and recall them.

Two immediately available accessories: personal profile and anthropomorphic skill package

Personal Profile: Resume for AI

A personal profile is a document that allows any AI to quickly learn about you. AI will understand you after reading this one document, so there is no need to crawl through hundreds of documents each time. The method is very simple. Open your most commonly used AI and issue this command:

Help me analyze the content of our recent conversation, Organize my work patterns, interests, preferences, and processes.

Review the draft, fill in the gaps, and save it as a fixed document. The content covers the core of the brand, who you want to serve, work logic and strategy, personal preferences and processes. The key point is that this file should be readable by any AI company, not bound to the platform, and updated regularly as working methods change.

Personification Skill Package: Memorizing people is easier than memorizing tools

With too many tools, you will forget which function is placed where. The solution is to personify the skill packages: give each skill package a name, write a character description, and clearly explain what this character is responsible for and what situation he is called in. For example, my "Chief Editor Chun" is responsible for copywriting, marketing, and content creation. As long as it is related to copywriting, I will directly call Editor Chun. The larger the Agent team is, the more necessary it is to personify it, and the persona should be specific to "what situation should I find him in?"

The right pace for adopting a new toolA hot tool has the most bugs right when it blows up, so hold off at first. Wait a week to confirm downloads keep climbing and nobody is calling it a scam; wait another week or two until the vendor or the community has a fix for the bugs, then start learning it. The more permissions a desktop Agent has, the more careful your homework needs to be: before installing, ask an AI you already use to check the risks and how to uninstall it. And just take the strongest concept out of a new tool and absorb it into your own system; you'll never finish trying to move your whole setup over.

How you can practice: Write an Agent rule card

For someone from a humanities background, designing an Agent framework starts with a single rule card. It condenses the three parts of Harness Engineering onto one page.

  1. What material I have: list your main folders and what each one is for.
  2. What tasks I do often: pick three recurring jobs and write down how you do each in three to five sentences.
  3. What must never run automatically: sending things externally, deleting files, touching sensitive data. Mark the no-go zones first.
  4. How to verify when done: for each task, write one "definition of done" so the AI can check its own answer.

This card is the starting point of your personal Agent system. After every round of work, write the experience back with a work log, and the rule card grows on its own. AI comes and goes, knowledge endures: the tools will keep changing, and the system you've organized is the asset that stays.

Agent FrameworkHarness EngineeringLoop EngineeringLLM WikiTag WikiLiberal arts backgroundKnowledge Management