Build Your Own AI Agent — Assemble Once, Stop Re-Explaining

OUTCOME — What you'll have afterward

Following this tutorial leaves you with three things.

What gets built

• One personalized AI agent of your own — running inside ChatGPT, Claude, or Gemini, with name, role, and voice configured
• One system prompt inside that agent — the document that carries your identity and rules
• At least one knowledge file — a PDF, text file, or document you refer to often

📖 Term: Custom GPT, Claude Projects, and Gemini Gems All three are personalized AI agents layered on top of a base AI, carrying your own role, knowledge, and rules. The platforms differ; the principle is identical. OpenAI's Custom GPTs — now simply called "GPTs" — launched in 2023 and have a mature ecosystem. Anthropic's Claude Projects is stronger on long-document handling and natural prose; as of 2026 it also offers "Claude Skills," reusable custom agent behaviors you can save and remix. Google's Gemini Gems connects directly to Workspace (Gmail, Docs, Drive). Whichever you choose, the same assembly process applies.

Tutorial metadata

Item	Value
Time	40 min (fast: 25, unhurried: 60)
Level	Intermediate (assumes you finished the first AI tool tutorial)
Tools	ChatGPT Plus ($20/mo), Claude Pro ($20/mo), or Google One AI Premium ($19.99/mo)
Coding knowledge	None required
Terminal	None (all in the browser)
Payment	Paid subscription required — all three platforms gate this feature behind paid tiers

Why this, framed as ending repetition

Anyone who uses ChatGPT for more than a month eventually hits the same wall: you find yourself pasting the same background paragraph every time. "I'm a university professor teaching e-commerce, please answer in Korean, use formal tone, give Korean company examples..." If you're copying that paragraph every morning, you are not using the AI as a tool — you are bending yourself to fit the AI's format. A personalized AI agent ends that loop with one single act of setup.

Where you are now: you've confirmed what this tutorial builds. Next, verify that your prerequisites are actually in place.

---

PREREQUISITES — What you need

Three things, nothing more.

Required

• One paid subscription: ChatGPT Plus ($20/month), Claude Pro ($20/month), or Google One AI Premium with Gemini Advanced ($19.99/month). All three platforms gate custom agents behind paid tiers
• A logged-in state with one AI tool: finishing the first AI tool tutorial is enough
• One knowledge file: PDF, TXT, or MD — your resume, your company's one-pager, your favorite prompt library, a book draft, anything

Not required

• Programming knowledge, API keys, a server
• "Actions" or MCP connections (the advanced integrations) — this tutorial deliberately skips them
• A plan to publish to the GPT Store — keep it private, just for you

📖 Term: What is a system prompt? A system prompt is the paragraph that tells the AI "you are this kind of entity." Regular questions you type into the chat box are called "user messages." The system prompt sits one layer above them as a standing instruction. No matter what the user asks, the AI obeys the system prompt first. In Custom GPT it's the "Instructions" field; in Claude Projects it's "Custom instructions"; in Gemini Gems it's "Instructions." That single page is the soul of the agent.

Environment check (2 min)

Platform	How to confirm
ChatGPT	chatgpt.com → "GPTs" or "Explore GPTs" visible in left sidebar
Claude	claude.ai → "Projects" tab visible at top
Gemini	gemini.google.com → "Gems" in left menu

If none of these appear, you're on the free tier. Upgrade the subscription first.

Where you are now: paid subscription and one knowledge file are ready. Next we cover why a personalized AI agent is different from "just saving a long prompt" before moving to actual assembly.

---

PLATFORM — Which one to choose

Before building, pick one platform. Here's how they compare in 2026.

	ChatGPT (GPTs)	Claude (Projects)	Gemini (Gems)
Price	Plus $20/mo	Pro $20/mo	One AI Premium $19.99/mo
Knowledge files	Up to 20 files	Exceptional on long docs	Google Drive direct link
Korean	Good	Very natural	Good
Strength	Mature GPT ecosystem	Long context, literary writing	Workspace integration
Best for	Existing ChatGPT users	Writing & document analysis	Gmail/Docs-heavy workflows

Don't overthink it. If you already use one platform, start there. The principles transfer across platforms in under ten minutes once you've learned them once.

---

WHY — Why build a personalized agent

"Can't I just save a long prompt in a notes app and paste it when needed?" Fair question. Three differences matter.

First reason — persistence.

Copy-paste only works when you remember to do it. Open a new chat, forget to paste, and that conversation starts without context. A personalized agent is different. The moment you open it, the system prompt is already active, and whatever you ask lands inside your context. Memory cost is zero.

Second reason — knowledge files.

Pasting a long PDF into a regular chat makes it available only for that one conversation. Open a new chat and you paste it again. In a personalized agent you upload the file once, and it is referenced across every conversation with that agent. Your entire book manuscript, your company's full FAQ set, twenty academic papers — upload them once and be done. Claude Projects is particularly strong here: its context window reaches 200K tokens, large enough to hold an entire thick manuscript without any retrieval chunking.

Third reason — sharing and duplication.

A well-built agent is an asset. You can share it with teammates, or keep several specialized ones and swap as needed: a "writing polish agent," a "meeting notes agent," a "lecture slide agent." A long prompt in a notes app can't do any of that.

📖 Term: Knowledge base and the context window Uploading a knowledge file doesn't mean the AI memorizes the whole thing. It retrieves the relevant portion on each turn. For this, the AI uses a "context window" — the maximum amount of text it can process in a single pass. ChatGPT holds roughly 120K words, Claude reaches 200K tokens, Gemini scales up to 1M tokens in some configurations. If your knowledge files exceed the window, they get chunked and searched automatically. The implication: structure matters more than file count. A well-titled PDF with clean headers retrieves far better than a wall of text with no headings.

Where you are now: you understand that a personalized agent is not prompt storage but context automation. Next, we build one. Five steps total, about 35 minutes combined.

---

STEPS — 5 steps, done

Step 1. Lock the purpose to one sentence (3 min)

Before touching any field, write a single sentence answering "what does this agent do?" Without this sentence, your system prompt will drift in every direction.

Good purpose sentences:

• "A TA that fills my e-commerce lecture slides with Korean company case studies"
• "An editor that rewrites my blog drafts in my voice and proposes five alternative titles"
• "A first-line support agent that answers customer questions grounded in our 20 FAQs"

Bad purpose sentences:

• "A multipurpose assistant that helps with everything" (too broad)
• "Like ChatGPT but smarter somehow" (no differentiator)

Pick one purpose only. Don't get greedy. Stuffing "writing + translation + research + planning" into one agent makes all four mediocre. Build a single-purpose agent first; add more later.

Step 2. Open the creation screen (2 min)

ChatGPT

• In chatgpt.com, click "Explore GPTs" in the left sidebar
• Click "+ Create" in the top right
• Go straight to the "Configure" tab. The "Create" tab lets you build through conversation, which is actually confusing for your first one. Configure is a form you fill in directly, giving you clean control

Claude

• In claude.ai, click the "Projects" tab at the top
• Click "+ New Project" → enter a name → click "Set instructions"
• Claude Projects has a relaxed length limit for instructions, so longer system prompts are fine

Gemini

• In gemini.google.com, click "Gems" in the left menu
• Click "New Gem" → enter a name and instructions
• You can link a Google Drive folder directly as a knowledge source — no file upload needed

All three platforms accept the same five-block system prompt you'll write in Step 3.

Step 3. Write the system prompt (15 min — the core)

This step is the heart of the tutorial. The other four take five minutes combined; this one deserves patience.

A system prompt has five blocks. Fill each one like a template.

Block 1 — Identity

Start with "You are [role]." Nail down who the agent is before anything else.

You are a teaching assistant for Professor MR_5PM's e-business course.
The class targets 2nd- and 3rd-year undergraduates; lectures are in Korean.

Block 2 — Purpose

"You do X. You do not do Y." Define scope explicitly.

Your role is to fill draft lecture slides with Korean company case studies.
General Q&A, assignment grading, and student counseling are not your role.
If such a request arrives, reply: "That is outside this agent's scope."

Block 3 — Voice

How it speaks. This is where personality lives.

Voice rules:
- Formal Korean, ~습니다 form by default
- Concise — aim for sentences under 40 characters
- Analogies grounded in Korean context (prefer Samsung, Naver, Coupang over foreign examples)
- No emotional modifiers ("truly amazing", "stunning", etc. are banned)
- If uncertain, do not guess — say "I am not certain"

Block 4 — Format

How the output is structured. Without this block the AI returns a different shape every time.

Default output format:
- Use markdown
- Core answer in the first 3 lines
- Supporting detail below
- End with "3 follow-up questions worth asking"

For slide requests:
- Each slide = title + 3–5 bullets + 1 real Korean company case

Block 5 — Edge cases

Pre-decide the tricky situations. As you use the agent, this block grows longest.

Edge cases:
- If a question requires current info, first note "This requires a web search"
- When citing numbers (revenue, market size), always include source and year
- If no Korean company example comes to mind, do not fabricate — suggest "An international case can substitute"
- Never assert facts without a source

Paste the five blocks in order into the Instructions field with one blank line between each. The blank lines help the model lock onto the structure.

📖 Term: Prompt engineering — instruct, don't describe A good system prompt is written as instructions, not descriptions. "This agent uses a friendly tone" is a description. "Answer in a friendly tone. Use formal Korean, but keep it warm" is an instruction. Instructions are obeyed more reliably. Also, negative rules ("do not do X") are weaker than positive rules ("do X"). Instead of "do not use emotional modifiers," prefer "state facts plainly and dryly." You can use both.

Step 4. Upload knowledge files (5 min)

In the Knowledge section of your platform, upload your files. Following these rules sharply improves quality:

• Structure beats count: three well-titled PDFs with clear headings beat twenty unstructured TXT files
• Remove duplicates: if the same content appears across multiple files, the AI will contradict itself. Version-manage cleanly
• File naming: use names that reveal content, like "2024-curriculum.pdf" or "case-library-korea.pdf"
• Limits: ChatGPT allows up to 20 files (512MB each); Claude Projects fits within its 200K-token context; Gemini Gems can link a Google Drive folder directly

When uploads finish, test with "Please read me the first line of one of my uploaded files" in the preview pane. If a sensible answer comes back, it's working.

Step 5. Run three test conversations (5 min)

Before saving, send three test prompts and verify each response.

1. Intended-purpose question: "Draft five week-2 lecture slides using Korean company cases." Does it respond the way Step 1 defined?
2. Out-of-scope question: "What should I eat for dinner tonight?" Does Block 2's "That is outside this agent's scope" trigger?
3. Knowledge reference question: "According to the 2024 curriculum I uploaded, what's the week-3 topic?" Is it actually reading the knowledge file?

When all three pass, save the agent. Private (Only me) is the right default for a first build.

Where you are now: one personalized AI agent is saved and pinned in your sidebar. Next we verify it's "actually done" before you start using it daily.

---

VERIFICATION — What counts as success

If you can answer "yes" to all five of these, this tutorial is complete.

1. [ ] The new agent is pinned in your sidebar. You can open it with one click even after navigating away
2. [ ] Four conversation starters appear when the agent opens. If not, set them. About 30% of perceived quality comes from here
3. [ ] Intended-purpose questions return output that follows the format defined in your system prompt. Voice, structure, and the three follow-up questions at the end — all respected?
4. [ ] Out-of-scope questions get declined with the "outside scope" response. If they don't, Block 2 is too weak. Strengthen it
5. [ ] The agent actually reads knowledge files and cites them. Responses starting with "According to your uploaded file..." indicate success

If any answer is "no," go back to that step and reinforce. The most common failure is number 3: the format isn't obeyed. In that case, rewrite Block 4 of the system prompt to be more specific and more forceful. "Use markdown" is weak. "Every answer must follow ## heading + 3 bullets + 1 example at the end, always" is strong.

---

VARIATIONS — Situational variants

Variant 1 — Claude Projects for writing and long documents

If your primary use is book-manuscript polishing, academic paper analysis, or Korean prose editing, Claude Projects is my preferred choice. Write the same five blocks in Custom instructions. As of 2026, Claude also offers "Claude Skills" — reusable bundles of custom behaviors (e.g., "always cite sources," "end with three follow-up questions") that you can save and apply across projects.

Variant 2 — A writing-style agent

Upload 10 of your past blog posts as knowledge files. In the system prompt write: "Learn this writing style and rewrite drafts in the same voice." Your own old writing becomes a voice trainer. I use this exact technique to polish long essay drafts in my actual voice.

Variant 3 — An FAQ support agent

Compile 20–50 frequently asked questions into one PDF and upload as the knowledge file. The system prompt: "Ground every answer strictly in this FAQ document. If the answer is not there, do not guess — reply 'Please contact a human agent.'" That's a first-line support agent, done.

Variant 4 — Gemini Gems for Google Workspace

If Gmail, Google Docs, and Drive are your daily environment, Gemini Gems is the most frictionless option. You can link a Drive folder directly as the knowledge source — no upload/download cycle, always pointing to the latest version of your documents.

Variant 5 — A personal assistant agent (caution)

The "all-knowing assistant that remembers my schedule, tastes, and values" is tempting but not recommended. Too broad a scope collapses quality everywhere. Instead, build 3–4 purpose-specific agents and switch between them. Much more practical.

---

TROUBLE — Where people get stuck

❓ "I don't see the GPTs / Projects / Gems menu" You're on the free tier. All three platforms gate custom agents behind paid subscriptions. Check subscription status in account settings.

❓ "I hit an error when saving" Most often: the Instructions field exceeds the character limit (ChatGPT is around 8,000 characters), or special characters (emoji, unusual Unicode) are mixed in. Cut the content in half and narrow down which section causes it.

❓ "The agent doesn't seem to read my knowledge files" Three common causes. First, your filenames are too generic ("document.pdf"). Give them meaningful names. Second, the file itself has no headings or structure, so retrieval quality drops. Third, the system prompt doesn't explicitly tell the AI to reference the knowledge files. Add a line requiring that reference.

❓ "Responses come back in English" Explicitly write "All responses must be in Korean with formal politeness" in the system prompt. The base model sometimes defaults to English even for Korean input because of training data patterns.

❓ "The same question gets different answers each time" AI is probabilistic by design — identical responses cannot be guaranteed. But by strengthening Block 4 (Format), you can keep the shape consistent even when wording varies. If you need strict content determinism, you'd move to the API, which is outside this tutorial's scope.

❓ "I want to publish it to the GPT Store" Possible, but not recommended for a first agent. Use it yourself for at least a week and iterate the system prompt five or more times before publishing. A rushed public agent is an embarrassing asset.

---

NEXT — What comes after

Forty minutes ago you may have been pasting the same background paragraph into a fresh chat every morning. Now you have that paragraph assembled once and kept on standby. The gap looks small, but after a week of use it compounds into real hours saved.

This tutorial is the second chapter of Part 4 (Productize) in the SPIN series. The arc:

1. Writing prompts ✓ (Part 1 complete)
2. Picking your first AI tool ✓ (Part 2 complete)
3. Making your own CLAUDE.md ✓ (Part 3 complete)
4. Building your own AI agent ✓ (now complete)
5. Putting your agent to real work — automation pipelines (next tutorial, in progress)

The best way to use this tutorial is to spend the next two weeks using only the agent you just built. Even if new use cases come up, resist creating a second agent immediately. Instead, keep refining the system prompt of the first one. Refining one agent five times is always a better investment than building five new agents. Once the first one is running cleanly, the second will be much faster.

당신은 MR_5PM 교수의 e-비즈니스 수업 조교입니다.
대학교 2~3학년 대상 수업이며, 교수는 한국어로 강의합니다.

당신의 역할은 수업 슬라이드 초안을 한국 기업 사례로 채우는 것입니다.
일반 질문 응답, 과제 채점, 학생 상담은 당신의 역할이 아닙니다.
그런 요청이 들어오면 "이 에이전트의 범위 밖입니다"라고 답하세요.

말투 규칙:
- 존댓말, 습니다체 기본
- 간결하게, 한 문장 40자 이내 지향
- 비유는 한국 상황 기반 (삼성·네이버·쿠팡 등 국내 기업 우선)
- 감정적 수식어 금지 ("정말 훌륭한", "놀라운" 같은 표현 사용 금지)
- 확신 없으면 추측하지 말고 "확실하지 않습니다"라고 쓸 것

기본 출력 형식:
- 마크다운 사용
- 핵심 답은 첫 3줄 안에 제시
- 그 아래에 근거 또는 상세 설명
- 끝에 "다음으로 물어볼 만한 질문 3개" 제안

슬라이드 요청이면:
- 한 슬라이드당 제목 + 불릿 3~5개 + 국내 기업 실제 사례 1개 구조

예외 처리:
- 최신 정보가 필요한 질문이면 먼저 "웹 검색이 필요한 질문입니다"라고 안내
- 숫자(매출·시장규모 등)를 쓸 때는 반드시 출처와 연도 병기
- 한국 기업 사례가 떠오르지 않으면 억지로 만들지 말고 "해외 사례로 대체 가능합니다"라고 제안
- 출처 없이 사실을 단언하지 말 것