"Advanced Prompting (2026) — Roles, Examples, and Chaining in Practice"

The 4-block role formula, four chaining patterns, and a working self-critique loop

ํ•ต์‹ฌ ์š”์•ฝ

  • Audience: Anyone who's already comfortable with the five-pattern basics and wants to push output quality with multi-step prompting.
  • What you'll get: 1) The 4-block role formula, 2) what makes few-shot examples actually work (and fail), 3) the four standard chaining patterns, 4) how to wire up a self-critique loop, 5) meta-prompting — letting the AI write your prompts.
  • Key shift: Basics was "ask well in one shot." Advanced is "break it into steps" + "let the model refine its own answer."

1. Role prompting — the 4-block formula

You are an expert alone is too thin. Use four blocks.

Block 1: Role

Who you are + what you're great at.

"You are a contract lawyer with 12 years of experience specializing in IT-startup SaaS and B2B disputes, with a focus on protecting suppliers (the service-provider side)."

Block 2: Audience

Who the answer is for.

"Your audience is a solo founder with a technical background but limited legal training. Briefly define legal terms on first appearance."

Block 3: Task

Exactly what to produce.

"Review the SaaS contract draft below and produce: 1) clauses risky to us, 2) ambiguous language, 3) protective clauses we should add."

Block 4: Constraints

What not to do + format.

"Don't speculate. Cite the contract clause for each claim. Output in plain English, Markdown table, 1–2 lines per row."

Save this 4-block as a system prompt and you can reuse the persona across many questions: ChatGPT Custom GPTs, Claude Projects, Gemini Gems.

2. Few-shot — what actually makes examples work

The basics piece established that 1–2 examples can lift accuracy ~7.48%. Advanced is about example quality.

2.1 Diversity beats quantity

Five similar examples < two distinct ones. The model latches onto whatever all examples share.

Bad set (all alike)

Five "diary → tweet" examples, all about food and waiting in line.

Good set (varied)

1) Food/queue, 2) Movie/emotion, 3) Workout/pain, 4) Commute/exhaustion.

2.2 Contrastive (X vs Y) examples

Showing what to do and what to avoid sharpens the signal.

"Of the two responses below, follow style B. Response A (avoid): formal, long, stiff Response B (target): casual, short, point-first"

2.3 The trap: too-strong examples kill diversity

If your task itself requires diverse output (brainstorming, ideation), use zero or one example. Examples = answer keys = the model converges.

3. Prompt chaining — break it up instead of all-in-one

Complex tasks split into 3–5 steps gain accuracy and become much easier to debug (Prompt chaining 2026 guide).

3.1 The four standard patterns

Pattern Steps Best for
Analyze → Plan → Draft → Refine Analyze → plan → draft → polish Writing, presentations
Extract → Transform → Summarize Extract → transform → summarize Data and document processing
Classify → Route → Generate Classify → branch → generate Triage, customer support
Generate → Critique → Improve Generate → critique → refine Code, design, copy

3.2 Walk-through — blog post (Analyze → Plan → Draft → Refine)

[Step 1 — Analyze]
"My topic is 'AI automation for beginners.'
Profile the readers, their search intent,
and what existing posts get right vs. miss."

[Step 2 — Plan]
"Using that analysis, draft an outline
with 7 H2 sections. For each, state the core
message and target word count."

[Step 3 — Draft]
"Write section 3 from the outline. 
Markdown, ~700 words."

[Step 4 — Refine]
"From the previous draft:
1) Remove vague phrases ('many people')
2) Mark numbers needing sources with '[source needed]'
3) Strengthen the opening sentence."

Output of one step is the input of the next. Even web UIs flow this naturally inside a single chat.

3.3 Mix models per step

A subtle benefit: different model per step. Cheap model (Haiku, GPT-4o-mini) for extraction; reasoning model (Sonnet, Opus) for analysis; creative model for generation. Documented savings of up to ~70% vs. running the strongest model end-to-end (SurePrompts 2026).

4. Self-critique loops

Generate → Critique → Improve, but folded into a single ask.

4.1 Single-prompt version

"Do this in three steps: Step 1: Produce a first draft of [task]. Step 2: As a critic, name three weaknesses. Step 3: Rewrite the draft addressing those weaknesses.

Task: 30-second ad copy for an eco-friendly tumbler."

The model drafts, criticizes itself, and improves — all in one response.

4.2 Multi-call version (stronger)

Splitting the steps into separate calls produces deeper critique. Boost it further by giving the critic a different persona.

"[Call 1] You are a copywriter. Write the ad." "[Call 2] You are a marketing director. Identify three weaknesses in the ad above." "[Call 3] You are the copywriter again. Rewrite v2 addressing those weaknesses."

4.3 Self-Consistency

Sample N answers with temperature ≥ 0.7 and pick the majority (Wang et al., 2022, arXiv:2203.11171). Strong on simple reasoning but multiplies cost. Use only on critical decisions.

5. Meta-prompting — let the AI write your prompts

The laziest and often most effective: ask the AI for better prompts.

"Generate three prompt versions for the task below. Each takes a different angle: (1) maximally specific, (2) example-rich, (3) role-based.

Task: Reply to a customer complaint email."

Pick the best output, or merge the best parts. Anthropic Console's Prompt Generator does exactly this (official).

When meta-prompting really shines

  • A task you'll repeat 100+ times (automation, batch jobs)
  • System prompts going into production
  • Domains you don't deeply know (legal, medical, finance personas)

6. Summary — when to reach for what

Situation First try Add if needed
One-off question Five basic patterns "Think step by step"
Long writing or research 4-block role Analyze → Plan → Draft → Refine chaining
Tight reasoning Chain of thought Self-Consistency (cost permitting)
Code review / debug Generate → Critique → Improve Multi-persona critique
Automation / batch System prompt + few-shot Meta-prompting for the system prompt

Developer notes

For API users:

  1. Chaining = N× tokens. A 4-step chain = 4 calls, with previous outputs feeding the next. Use prompt caching (Anthropic, OpenAI, Google all support it) aggressively.
  2. Chaining ≠ agents. You define the steps; agents (ReAct, function-calling loops) let the model decide. Start with chains; graduate to agents.
  3. Self-Consistency implementation: call N times at temperature ≥ 0.7, then majority-vote — or use an LLM-as-judge to pick the best.
  4. Mix-model orchestration: LangChain or LlamaIndex routers let you assign models per step explicitly.
  5. Meta-prompting at scale: Anthropic Prompt Improver returns improved versions of an existing prompt.
  6. Evals are non-negotiable. Chaining and meta-prompting introduce regressions easily. Maintain a 10–20-item golden set + LLM-judge eval pipeline.

References

This is part 3-2 of 11 in the AI Basics series. Next: AI Learning Tools — NotebookLM + Perplexity in practice.

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks ๐Ÿ“š Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… ๐Ÿ“š Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent ๐Ÿ“š Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »