Agent Starter Wiki

Starter knowledge base for AI agents. These are reference documents that complement the agent-starter-skills — the skills tell the agent what to do, this wiki tells it how to do it well.

This is a starting point, not a finished product. Building and maintaining your own knowledge base is the single most impactful thing I've done to make agentic workflows actually work. More than the skills, more than the tooling, more than prompt engineering. When the agent has your taste, your style, and the lessons you've already learned written down, the output stops being generic and starts being yours.

Use these files as examples, then iterate: try the skills, notice what the agent gets wrong, and write down what you learn so it doesn't happen again.

What's Included

Writing:

• writing-with-zinsser.md — Complete Zinsser writing guide: core rules, AI workflow, taste decisions, critique patterns, structural patterns, AI tells, before/after examples
• elements-of-eloquence.md — 16 rhetorical techniques from Mark Forsyth for manual editing passes

Thinking:

• cognitive-debt.md — Mental model gap from AI-assisted coding, four phases of AI usage, Peter Naur connection

Brand:

• isaac-flath-brand.md — Positioning, marketing philosophy, visual design, email style

How to build your wiki

The starter files cover writing fundamentals. But your wiki should grow to include knowledge specific to your domain, your style, and what you've learned works. The process:

1. Use the agent skills. As you work, notice where the output is wrong or generic. That's a signal that knowledge is missing.

2. Work through problems in conversation. When the agent makes a bad choice — wrong tone, bad structure, misunderstood intent — explain what you want and why. That conversation contains the knowledge.

3. Ask the agent to write it down. After a productive session, ask: "Write up what we learned as a reference doc for future sessions." Save it to your wiki. Be specific — not "writing tips" but "why we rejected the negative-then-positive arc as a LinkedIn formula."

4. Repeat. The wiki compounds. Each captured lesson makes the next session better. After a few weeks, the agent produces good output on the first try because your preferences are written down.

With agentkb (recommended)

agentkb automates the consolidation step. It manages your wiki with git sync and provides semantic search so the agent can find relevant knowledge during sessions.

# Install
uv tool install agentkb

# Fork this repo, then configure as your wiki
agentkb settings set wiki_remote "git@github.com:youruser/agent-starter-wiki.git"
agentkb sync pull

# After working sessions, consolidate what you learned
agentkb consolidate

agentkb consolidate cross-references your recent Claude Code sessions against the wiki. It shows where you were active but haven't captured the knowledge yet. The agent reads through the sessions, extracts reusable lessons (mistakes, techniques, taste decisions, domain knowledge), and writes or updates wiki pages.

agentkb currently supports Claude Code and is open to PRs for other coding agents. But the core loop works with any agent — you don't need agentkb to build a wiki. After any hard session, just ask the agent to write up what it learned and save the file.

Setting up agentkb with Raw2Draft

Raw2Draft clones this wiki on first launch. To upgrade to agentkb-managed wiki:

# Install agentkb
uv tool install agentkb

# Fork this repo on GitHub, then configure
agentkb settings set wiki_remote "git@github.com:youruser/agent-starter-wiki.git"
agentkb settings set skills_remote "git@github.com:youruser/agent-starter-skills.git"
agentkb sync pull

# Index your Claude Code chat history (optional, enables consolidation)
agentkb chats index

# Now consolidate after working sessions
agentkb consolidate

Your wiki is now searchable via agentkb search -s wiki "topic" and syncs across machines via git.

With Claude Code directly (no agentkb)

git clone git@github.com:youruser/agent-starter-wiki.git ~/git/my-wiki
claude --add-dir ~/git/my-wiki

The agent can read any file in the wiki directory. After sessions, manually add markdown files to the directory and commit.

Format

Plain markdown with optional YAML frontmatter. No special structure required — organize by topic in whatever way makes sense for your domain. agentkb indexes all .md files for semantic search.