AI Agents I Built (1/7) — Writing Director: The Full Pipeline from Research to Auto-Publish

Designing a content-dedicated pipeline from insight input to automated Blogger publishing

Key Summary

  • The content writing pipeline consists of 7 stages: insight input → classification → interview → research → compile → verify → publish
  • Simple requests (1–2 stages) are handled directly by the orchestrator; complex requests (3+ stages) run the full pipeline — when ambiguous, classify as complex
  • The "My Thoughts / Perspective" section in source files records only user-provided insights; the AI does not generate viewpoints arbitrarily

Pipeline Background

Completing a single blog post requires multiple stages: research, structuring, verification, writing, and publishing. Without clearly defining what each stage must produce, output quality varies with every run.

The Writing Director resolves this with a gated stage pipeline. Each stage is isolated as an independent skill (SKILL.md), and the output of each stage becomes the input of the next.

Body

1. Full Pipeline Structure

Insight input → Classification (simple / complex)
  ├─ Simple → research → source file → verify → publish
  └─ Complex → deep-interview → research → compile → verify → publish

Each stage is an independent skill (SKILL.md). One skill owns one responsibility.

Stage Skill Role
Clarification deep-interview Convert ambiguous requests into scoped specifications
Investigation research Systematic research, source collection
Structuring compile Author source file (.md)
Verification verify 4-stage fact-check
Blog publish publish Post via Blogger API
Twitter publish twitter-draft Deliver draft to Discord

2. Classification — Simple vs. Complex

Every request passes through a classification stage first.

Simple (1–2 stages): "Give me a quick summary of this topic." The orchestrator handles research and source file authoring directly. This path is taken when the overhead of invoking a sub-agent exceeds the task volume.

Complex (3+ stages): Requests requiring deep analysis. Execution starts from the interview stage and proceeds through the full pipeline sequentially.

Core principle: when ambiguous, classify as complex. Over-classification is safer than under-classification. Classifying a request as simple and discovering mid-execution that it requires more depth forces a full restart. Classifying it as complex when it turns out to be simple means the intermediate stages pass quickly.

3. Ambiguous Request Gate

There is one gate before classification. If a request contains no concrete signal — no keyword, scope, or intended use — deep-interview is forced before research begins.

"Write something about AI" is ambiguous. Which aspect of AI? For a blog? For Twitter? Who is the reader? Starting research under these conditions produces directionless material.

deep-interview computes an ambiguity score:

Dimension Weight Measurement
Topic 40% What is being investigated?
Scope / constraints 30% How far? Exclusions? Intended use?
Completion criterion 30% What collected material constitutes done?

Research does not start until the weighted average ambiguity drops to 20% or below. One question is asked per turn, targeting the highest-ambiguity dimension first. The force: prefix bypasses this gate.

4. Research — Iterative Exploration Loop

Research is not a single search pass. It operates as a repeated cycle: explore → read → evaluate → re-explore.

Round 1: Initial exploration. Target: secure at least 3 independent sources. Each source is tagged with a credibility label — primary source, expert interpretation, secondary summary, unverified.

Round 2+: Evaluate collected results. Which questions remain unanswered? Which claims rely on a single source? If further exploration is warranted, execute the next round.

Exit conditions: The core question is supported by 3+ independent sources; unresolved points are classified as inaccessible information; or the maximum of 3 rounds is reached.

Collected information is classified by evidence type:

Type Meaning Tag
Fact Verified data, official announcements ✅ Fact
Interpretation Analysis from experts or institutions ๐Ÿ” Interpretation
Estimate Reasoned inference with supporting basis ⚠️ Estimate
Opinion Personal view, unverified claim ๐Ÿ’ญ Opinion

5. Compile — Source File Structure

Research output is structured into a source file. Filename format: YYMMDD_topic_source.md, stored under sources/.

---
์ œ๋ชฉ: [topic]
๋‚ ์งœ: [YYYY-MM-DD]
ํƒœ๊ทธ: [related keywords]
์šฉ๋„: [blog / twitter / both]
์ƒํƒœ: [collecting / draft-complete / published]
---

## Key Summary
- (3 lines max)

## Collected Material
- Organized with sources

## My Thoughts / Perspective
- (User insights only — AI-generated viewpoints prohibited)

## Blog Notes
- (Points to expand into long-form)

## Twitter Notes
- (High-impact points for short-form)

Rule for "My Thoughts / Perspective": This section records only insights provided by the user. The AI does not generate "I think…" content. If no user insight has been provided, the field reads "(awaiting user insight)". The perspective published to the blog must be the user's.

6. Verify — 4-Stage Verification Process

A completed source file must pass 4-stage verification before proceeding.

Stage Verification Pass Criterion
1. Source check Does every item have a source? Zero unsourced items
2. Evidence classification Are fact / interpretation / estimate / opinion labeled? Every item tagged
3. Counter-scenario Does the file include a "what would make this wrong?" review? Minimum 1 counterargument
4. Applicability Does the content deliver real value to the user? Consistent with key summary

Failure at stage 1 blocks execution of subsequent stages. After verification passes, a Red Team self-attack is run: "Where could a factual error occur?", "Are the sources biased?", "Is any information time-sensitive and likely to become invalid?"

7. Publish — Two-Track Delivery

Source files that pass verification are delivered through two paths.

Blog: A blog post is authored from the source file and confirmation is requested from the user. Upon approval, the post is published automatically via Blogger API v3. Publishing without user approval does not occur.

Twitter: A Twitter draft is authored and delivered to Discord. The user posts directly.

Key Design Decisions

Why a gated pipeline structure: The inability to skip stages is the core constraint. Research output must be structured into a source file; the source file must pass verification before the publish stage opens. The publish path does not unlock without verification.

Basis for simple vs. complex classification: In early operation, classifying most requests as simple caused complexity to surface mid-research, triggering rework. "When ambiguous, classify as complex" is set because in the complex pipeline, a genuinely simple request passes intermediate stages quickly; the reverse direction requires a full restart from the beginning.

Prohibition on AI-generated viewpoints: If the AI generates plausible-sounding opinions for the "My Thoughts / Perspective" section, the blog publishes the AI's reasoning rather than the user's perspective. This is enforced as an explicit rule: user insights only, AI-generated viewpoints prohibited.

Model routing: Models are assigned by task character. File search and simple formatting route to a lightweight model. Content writing, research compilation, and draft editing route to a mid-tier model. Strategy design, review synthesis, and fact-check arbitration route to a high-capability model. Each stage has different precision requirements.

Closing

The core of the writing pipeline is that each stage is independent while being sequentially linked. Research output is structured into source files; source files must pass verification before publishing. Stages cannot be skipped.

The problem this pipeline solves is quality consistency. Define clear input/output and pass criteria for each stage, and reproducible quality follows.

In content automation, the most important factor is not speed — it is verification. Publishing accurate, confirmed facts protects the blog's credibility more than publishing quickly.

Series overview: Series index

๋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 »