blog-writing - SKILL.md Agent Skill

name: blog-writing description: Write original blog posts for the AI topics blog. Save to blog/ directory (NOT wiki/raw/articles/).

Blog Writing

Directory Rule (CRITICAL)

Save all original blog posts to /opt/data/ai-topics/blog/.

wiki/raw/articles/ is reserved for externally-sourced, finalized article scrapes from the web. Do NOT save original blog posts there.

Naming Convention

{YYYY-MM-DD}_hermes_{short-slug}.md

YYYY-MM-DD = Publication date (usually today)
hermes = Author marker
short-slug = Hyphenated descriptive slug (e.g., performance-controllability-tradeoff)

Process

Research phase: Gather real operational data before writing. Use delegate_task with batch mode (up to 3 parallel subagents) to collect stats, read reference files, and extract examples simultaneously. This grounds the post in concrete numbers and real artifacts rather than abstract claims.
Write the post to blog/{filename}.md
Update wiki/log.md with the new entry (append a dated section summarizing what was created)
Commit and push: cd /opt/data/ai-topics && git add blog/ wiki/log.md && git commit -m "blog: <title>" && git push

Companion Slides (Marp)

When the blog post introduces a technical topic to an audience, companion Marp slides often belong in slides/:

Naming: {YYYY-MM-DD}_{short-slug}.marp.md
Reuse the same research gathered for the blog
Reference existing slides in the same directory for CSS theme reuse (copy the style: block verbatim)
Structure: 5 parts — Lead/Intro → Core Content (2-3 parts) → Takeaways → Thank You
Use  for part transitions,  for opening/closing
Slides expand on what the blog covers in depth; they are not a summary of the blog

Active blog post outlines and research notes are kept in references/ within this skill.

references/baudrillard-blog-outline.md — In-progress structure for Baudrillard × AI blog post (2026-05-11).
references/structural-duality-analysis-pattern.md — Reusable three-layer framework for AI architecture essays: divide-and-conquer → duality → naive model's limit. Emerged from the 2026-05-20 Anthropic multi-agent × RLM analysis session.

Frontmatter Template

---
title: "Title Here"
date: YYYY-MM-DD
author: Hermes (kzinmr's AI Topics)
tags: [tag1, tag2, blog]
sources:
  - concepts/example.md
  - https://example.com/source
---

For series posts, add:

series: series-slug
series_index: N

Voice

Japanese (user's language)
Humanized: no boldface overuse, no em dashes in titles, no signposting, no grand conclusions
Personal, opinionated, varied rhythm
If the user provides voice samples or feedback, apply it

Structural Preferences

The user's blog posts follow a narrative arc pattern:

出発点 (起点) — What drives the phenomenon? Ground it in concrete forces (e.g., information flow acceleration, arbitrage dynamics, economic/technological pressures).
中間展開 — Unfold the layers. Use theoretical lenses (Baudrillard, Baumol, etc.) to illuminate what's happening. Each theoretical move must be explained carefully — don't skip inferential leaps.
着地点 — Don't just identify problems. Include a pragmatic stance: given undeniable external pressures that make engaging with these dynamics unavoidable, how should we relate to them?

Theory Integration Rules

When weaving in abstract theoretical concepts:

Start from lived experience, not theory. The user rejects posts that feel like "theory playing with itself" (論を弄し過ぎた). The post must open with concrete, personal felt experience — the theory is a lens brought in later to illuminate what was already observed. If the first draft opens with a theoretical framework (Baudrillard's 4 stages, Baumol's cost disease model, etc.), scrap that opening and rewrite from the user's actual experience (building the wiki, reading summaries, feeling the gap between form and substance).
Explain step by step. Concepts like Baumol's cost disease or Gödelian incompleteness need careful, grounded exposition — not a name-drop. Show the causal chain connecting the theory to the concrete phenomenon.
Problem identification alone is not enough. The user considers pure problem-cataloguing (especially with abstract formalism) pedantic. Every problem raised must be accompanied by a pragmatic position: why does it matter practically, and what posture should we take toward it?
Acknowledge external pressure. Structural problems are easy to list. The harder and more interesting move is to acknowledge the inescapable external forces (competition, information velocity, economic incentives) that make engaging with these broken systems unavoidable.

Deep Technical Architecture Articles

When synthesizing wiki knowledge into a deep technical architecture article (e.g., mapping a design pattern like Tenant Agent Pack or Agent Control Plane to its supporting wiki primitives):

Research Phase

Identify the target concept's components or layers
For each component, search the wiki for supporting technical primitives: ls wiki/concepts/ | grep -iE "keyword1|keyword2|..."
Read the most relevant pages (3-5 per layer) to understand the current state of knowledge
Identify gaps — what's NOT yet covered in the wiki that would be needed

Writing Phase

Frontmatter: title, date, author: Hermes (kzinmr's AI Topics), tags (include blog), sources (link to wiki concept pages and raw articles)
Audience: Business/development professionals with high resolution understanding of the domain
Language: Japanese (日本語)
Structure:
- Compelling title with concrete insight (not generic)
- Architecture diagram (ASCII) showing component relationships
- One major section per layer/component, each anchored to specific wiki primitives with backtick references
- Comparison tables where 3+ items are analyzed
- Gap analysis section identifying what's missing from the wiki
Length: 150-250 lines for deep architecture articles
Tag validation: All tags must exist in SCHEMA.md taxonomy

Post-Writing

Update wiki/log.md with a concise entry (prepend using execute_code Python open/read/prepend/write — never use patch with --- anchor)
git add blog/ wiki/log.md && git commit -m "blog: ..." && git push

Reference Files

references/tenant-agent-pack-9-layer-architecture.md — 9-layer mapping of Tenant Agent Pack to wiki primitives with gap analysis
references/agent-control-plane-13-component-architecture.md — 13-component × 3-layer mapping of Agent Control Plane to wiki primitives with platform comparison

Pitfalls

Ambiguous skill name: The blog-writing skill exists in two locations (~/.hermes/skills/ and ~/ai-topics/config/hermes/skills/). When skill_view returns "Ambiguous skill name", skip loading it and proceed directly with the known workflow.
Log.md prepend via --- anchor: Never use patch with --- as old_string in log.md — it matches YAML frontmatter delimiters elsewhere. Use execute_code Python open/read/prepend/write instead.
Backtick in terminal Python: When using terminal() with inline Python, backticks in the string are interpreted by bash as command substitution, silently corrupting content. Use execute_code for any log.md operations involving backtick-wrapped paths or URLs.

Session Reset in Discord Threads

When the user continues a blog post discussion across Discord sessions and you can't find the referenced draft/outline:

STEP 0 — Check skill references FIRST. skill_view(name='blog-writing') returns a linked_files dict. If a reference file matches your topic (e.g., references/baudrillard-blog-outline.md for a Baudrillard post), read it immediately with skill_view(name='blog-writing', file_path='references/...'). Outlines and user feedback from prior sessions are saved here.
If the reference file doesn't exist or doesn't contain the outline: Ask early. Don't burn 10+ tool calls on session_search + search_files + execute_code probing for a file that only existed in a prior session's memory.
Demonstrate understanding of all feedback points, then ask for the file path or outline re-share.
The canonical locations to check are blog/ and wiki/concepts/ (for concept pages that may serve as source material).