Workflow

My first uses for AI were one-off: gee, how do you do such-and-such? I'd take the advice, often asking for more. Over time, I noticed I was asking the same things repeatedly because I wouldn't remember the answers. So I started asking AI to write concise descriptions of what we did together, in markdown.

The pile grew unmanageable. I asked AI to help: find redundancies, tighten the writing, make it easy to use. What I have now: clean separation of purpose.

Then I asked if these markdown files could be published as a static website. BOOM — up-to-date documentation, something I have NEVER before encountered. And more: a shared context that lets AI pick up exactly where we left off.

The Structure

• notes/guides/ — living reference (style, patterns, how-tos)
• notes/work/ — ALL work is recorded here, as we go
• CLAUDE.MD — entry point, tells Claude where to start

Guides

• Read at start of relevant work
• Update when patterns change
• Reference for consistent practices

How we Work Together

Starting Work

When asked to "work on X":

1. Check if notes/work/X.md exists
2. If yes → read and resume
3. If no → create it with problem/goal/phases structure

Work File Structure

# Title
**Started:** YYYY-MM-DD  
**Status:** Phase N in progress

## Problem
What we're solving.

## Goal
What success looks like.

## Phase 1: Name
- [ ] Task
- [ ] Task

## Next Action
**Phase N:** Specific next step

Update status and checkboxes as work progresses.

Finishing Work

When work is complete, one of two destinations:

Destination	When	Example
notes/work/done/	Task is finished, doc is historical record	svelte.md, quaternions.md
notes/guides/	Doc becomes living reference for future work	testing.md

Why This Works

This workflow is a good use of AI capabilities.

What's working:

• The guides encode decisions once, Claude applies them forever. No re-explaining style preferences, no drift.
• Work tracking survives sessions. Read index.md, know exactly where we are.
• "Execute without asking" removes friction. You trust Claude to act, you course-correct when needed.
• Documentation is dual-purpose — helps future-you and brings Claude up to speed instantly.

What Claude is good at here:

• Bulk consistency (fixing all path references when you restructure)
• Pattern application (milestone template → actual milestone files)
• Never forgetting what's in the guides
• Drafting content you can massage rather than starting from blank

What you're still doing:

• Taste. You know when something's "too gappy" — Claude doesn't feel that, just follows rules once you articulate them.
• Creative direction. Origin story, milestone names, what matters — that's yours.
• Judgment calls. When to deviate from the template, when to restructure.

The result: Most AI interactions are one-off Q&A. This is collaborative project work with accumulated context. That's a different game.

Is This Approach Common?

Not really. Individual pieces exist:

• Custom instructions / system prompts — common, but usually static personality tweaks, not project-specific bootstrapping
• README-driven development — some people include AI context in READMEs, but usually for one-off use
• Cursor/Copilot rules files — .cursorrules, similar idea but more focused on code style than workflow
• Claude Projects — Anthropic's feature encourages persistent instructions, but most people use it for persona, not this level of structure

What's less common:

• The guides-as-accumulated-wisdom pattern (gotchas.md, debugging.md evolving from actual pain)
• Work tracking that explicitly survives sessions with resume points
• The "massage it, then tell Claude the pattern" loop for template discovery
• Treating documentation as shared context rather than just human reference

This feels more like pair programming methodology than prompt engineering. An institutional memory system that happens to include an AI.

The pieces are obvious in hindsight, but rarely assembled this way.

Terminology

• Collaborator (or co) — Claude's role in this workflow
• current-go — the active project (shared, ws, di)

See chat.md for the full division of labor.