Dad always said:
Let the tools do the work #
Uses #
- Writing code
- Writing configuration
- Debugging
- Interact with Github
- One-off or supporting tools
- Introspecting environments
- Making graphs
- Explaining things
- Refactoring
- Batch / tedious / repetitive work
- Non-tech tasks like doc organization, research
What I've been up to #
[[nu-workflow/Projects and Experiments/Documatic|Documatic]] #
AppPack RAG app / chat with AppPack.io docs
What I've been up to #
[[Band Set List Manager]] #
What I've been up to #
[[nu-workflow/Projects and Experiments/Band Website and Branding|Band Website and Branding]] #
What I've been up to #
Open AI Image Generator #
I Built This #
Mail My AI #
![[logo-orig.png]]
Coding Agents #
- Claude Code (Anthropic)
- Codex (Open AI)
- Aider (open source)
- Jules (Google)
- Cursor, Windsurf, CoPilot
Critiques #
- Too fast to keep up with
- Code styles inconsistent
- Can get wildly off-track for simple fixes
- I still want to write the code
- Struggles with linting and code standards having to make several passes
Why Agents Are Bad Pair Programmers
I read of all Cloudflare's generated commits
Approach #
You are in control #
So many tasks can be aided by agents that are beyond "just writing code" and can make lots of tedious work commoditized
Don't get hynoptized by the genie
Approach #
Don't worry about The Vibe #
AI-assisted coding for teams that can't get away with vibes
- AI is a multiplier
- Use concrete steps and lots of communication
- Use AI to both generate and refine strategies at each step
Approach #
Good Engineering = Good Results #
A system in which AI thrives is one with markers of a high quality team and codebase. These are:
- Good test coverage, with useful assertions
- Automated linting, formatting and test checks before code merges
- Continuous integration and deployment
- Well documented changes, tech specs, ADRs with good commit messages
Approach #
Good Engineering = Good Results #
(continued)
- Consistent styles and patterns, enforced through a formatter
- Simple, concise, well-organised code
- Clearly defined features, broken down into multiple small story cards
Communication is Key #
- Be thorough and explicit with prompts
- Reference files and entities directly
- Provide links to docs or other references
- Be explicit about goals
Tools and Configs #
The agent can search the web, read and edit files, run commands
CLAUDE.md aka Memory #
- Like a "system" prompt
- Meant to enforce preferences and standards
- Wide range of uses and interpretations
https://docs.anthropic.com/en/docs/claude-code/memory
Level Up #
Stage One: Commands #
Like an alias for a prompt
Project / Slash Commands #
- start new project with idea
- product research
- implement feature
- debug or test
- triage
- document
- diagram
Project Commands #
A Community list for inspiration
Awesome Claude Code - Slash Commands
Level Up #
Stage Two: Parallel Agents #
- Let them work while you review and plan
- Get multiple options
- Code is cheap! Throw it away and try again
Parallel Agents #
- Claude Squad: https://github.com/smtg-ai/claude-squad
- Claude Swarm: https://github.com/parruda/claude-swarm
- IndyDevDan video: https://www.youtube.com/watch?v=f8RnRuaxee8
- Tasks tool
Parallel Agents #
- Agent controls multi-tasking vs. task per worktree?
- https://www.reddit.com/r/ClaudeAI/comments/1l4jocr/psa_claude_code_can_parallelize_agents/
- Different types of agents per task? (swarm)
Git Worktrees #
Little known Git feature for isolating branches
Level Up #
Stage Three: MCP #
Add MCP servers to enhance agent capabilities
- Streamline execution, use less tokens
- Build knowledge bases from your work
MCP Servers #
- AWS MCP: docs, service interaction
- Cloudflare MCP: docs, service interaction
- Serena: semantic code / LSP support
- Basic Memory / Revect Knowledge base management
Tips #
- use --resume to get back in sessions or reference work