LLMs: From Baby Bots to Expert Employees
A beginner’s tour of prompt engineering — from tag hacks to autonomous agents
Large Language Models (LLMs) such as ChatGPT predict the next token in a sequence. Over the last five years, users kept inventing “prompt hacks” to steer them. Each hack soon became training data for the next model — until the trick was baked in.
Glossary
1. The Highlight Era
1.1 — Why was it hard?
Early public LLMs ignored half of what we said unless we pointed out the crucial bits, much like showing bright colours and capital letters to babies.
1.2 — What prompt-engineers did?
Wrapped every must‑have fact in custom tags, then explained the tag:
<highlight>
Country: Brazil
Metric: Inflation rate
Period: 2010–2020
</highlight>INSTRUCTION
— — — — — -
Everything inside <highlight> … </highlight> is crucial context.
Use it first; do not skip or alter these details.
By bracketing key tokens with <highlight> and defining the tag’s meaning, users forced the model’s “attention” onto exactly the right words.
1.3 — What’s baked in now?
Modern models infer salience on their own; we rarely need XML or JSON scaffolding.
Just as a toddler eventually recognizes the essential parts of a sentence without the aid of flashcards, LLMs learn to identify salience during training. The highlight trick became an internal instinct — we ask a standard question, and the model automatically zeroes in on the keywords.
1.4 — Evolution
1.5 — BottomLine
Highlighting began as a survival skill; now it’s implicit. Use bold or italics for emphasis, not verbose XML.
2. Reasoning — The Chain-of-Thought (CoT) Era
2. 1 — Why was it hard?
Vague requests (“Bring me some water”) yielded messy responses — Toddler returns with a giant cup, water sloshing everywhere. So we switch to clear, ordered steps:
Go to the kitchen → Pick the small plastic cup → Turn on the tap just a little → Fill the cup halfway → Turn the tap off → Walk slowly and hand me the cup.
LLMs behaved the same way. A broad question like: “Who won the match and why did their strategy work?” Often produced a messy or hallucinatory answer.
2.2 — What prompt-engineers did?
Prefix complex questions with “Think step by step.” Then checked every line of reasoning.
2.3 — What’s baked in now?
High‑parameter models rehearse those micro‑steps silently; we just see the answer (or ask them to “show your work”).
Just as a high-school kid can now handle: “Hey, could you grab me a glass of water?” They choose a sensible cup → fill it to the right level → return without spilling. They’ve rehearsed the routine so many times that the intermediate steps run silently in the background.
The same shift happened in LLMs. Step-by-step prompting became famous as the Chain-of-Thought method, and the mental routine was folded into training.
2.4 — Evolution
2.5 — BottomLine
CoT freed us from scripting every micro‑step. We keep it handy for audit trails, but brevity is the new default.
3. Advanced Reasoning — The Tree of Thought Era
3. 1 — Why was it hard?
Models latched onto the first idea, like a student who submits the first draft of an essay.
Get Sanjay Krishna Anbalagan’s stories in your inbox
Join Medium for free to get updates from this writer.
LLMs behaved similarly to high school students; they tended to latch onto the first solution that came to mind. “Plan a 3-day Tokyo trip on a $600 budget.” The model selects the first itinerary it thinks of, but it costs $890.
3.2 — What prompt-engineers did?
Mini decision trees: Sometimes the best answer isn’t linear — it’s the best among many possibilities. Prompt-hackers responded with mini decision trees inside the prompt:
Think in a tree of options:
1️⃣ Plan A ( ¥60 000) … 2️⃣ Plan B (¥80 000) … 3️⃣ Plan C (¥55 000) …
Pick the plan under $600 with the least daily travel time.Return only that plan.
The model now explored several branches and delivered a $585 itinerary that fit.
3.3 — What’s baked in now?
Tree-of-Thought (ToT): Millions of generate‑compare‑select traces taught newer models to branch, rank, and prune by themselves.
3.4 — Evolution
3.5 — Why did ToT become feasible?
3.6 — Bottom line:
Today’s LLMs explore a forest of possibilities, then serve the best branch — no spilled logic.
4. The Orchestration Era
4. 1 — Why was it hard?
A request like “Turn meeting minutes into a polished slide deck” crashed unless you also scripted every tool call. Imagine giving a college intern this request: They’ll probably ask: “Which minutes? What template? Do we have a design tool?” Unless you also hand them a dependency graph…
1. Fetch minutes from `/docs/meetings/Q3‐review.md` 2.Summarise into five bullet points 3. Load the “Q3 Template.pptx” 4. Insert each bullet onto its own slide 5. Export as PDF
…the intern can’t finish without more help.
4.2 — What prompt-engineers did?
Wrote explicit dependency graphs
Plan:
1. getFile(“Q3_review.md”) 2. summarise → 5 bullets 3. createSlides(template=”Q3 Template”) 4. insert each bullet 5.exportPDF()
Run the plan.
4.3 — What’s baked in now?
Trained models decompose → select tool → execute → merge entirely on their own.
4.4 — Evolution
4.5 — Why this leap became possible?
4.6 — Bottom Line
Chain-of-Thought taught LLMs how to think.
Tree-of-Thought taught them how to explore options.
Orchestration now teaches them whom to call.
Conclusion
Prompt-engineering tricks do not die; they evolve.
- <highlight> → inferred salience.
- “Think step by step” → silent reasoning.
- Mini decision trees → internal branching.
- Tool graphs → autonomous agents.
We still use the same strategies — just sparingly and declaratively.
In other words, the strategies survive — just in a minimal, declarative form instead of a verbose, procedural one.
Your turn
The pattern is clear: once a community-driven hack gains traction, it soon becomes a built-in feature.
Which prompt hack do you believe will be abstracted away next — and why?
Drop your prediction in the comments. Let’s crowd-forecast the next baked-in super-power!
