Workflows · for students and engineering leads

My Workflows

I build with AI agents every day, and what decides whether it works is not the prompt. It is what each agent sees, who delegates to whom, and where the project's memory lives. Four short chapters, each with something to play with.

01

The context window

Before the first prompt, your tool has already filled part of the window. What you add next decides whether the agent starts informed or starts guessing. Switch the layers on and off and watch what the window is made of.

Playground · a session start

What your agent sees when a session starts.

Three layers fill the window before the first prompt: documents, memory and knowledge. Your tool takes its own share. Play the two approaches, switch items on and off, or change the window size.

Window
Your context window at the start of a session · to scale≈79.8k in use · 40%
0100k200k tokens
The curated part, enlargedto its own scale · ≈29.8k · 14.9% of the window
Build plan 3.5k
Architecture 6k
CLAUDE.md · AGENTS.md 2.5k
Second brain 9k
Shared brain 6k
Documents · foundationsMemory · working stateKnowledge · second and shared brainYour tool's shareUncurated
1DocumentsFoundations: the canonical documents the project is built against
2MemoryWorking state: what carries from one session to the next
3KnowledgeCompounded knowledge, pulled by relevance
4Your tool's shareSystem prompt and tools of the harness. Always there
×Without curationWhat gets pasted in when nobody curates
79.8ktokens in the window
40%of the window used
100%of what you added is curated
0.9×relative cost per call

Curated and lean

The agent reads the build plan, the standing brief and the latest handoff first. It knows the goal, the rules and where the last session stopped.

Illustrative numbers, close to a real project in The Curator. The tool's share is an estimate: no tool reports it.See The Curator →
Tools usedThe Curator
Go deeper
02

Orchestration

My main session plans, delegates, syncs state and audits. It does not do the bulk work. Each task goes to the cheapest model that can do it well, and nothing is accepted until it has been checked.

Step 0 of 9
  1. Rename files to the new conventionQueued
  2. Build the Library pageQueued
  3. Design the entitlement schemaQueued
  4. Audit a diffQueued
Orchestrator

Plans, routes, syncs state, audits. Reads summaries, not workings.

Orchestrator context12%
Current worker context0%
A worker can fill its window; it is thrown away after the audit.
  • Small modelSimple, mechanical worktoday: Haiku
  • Mid modelFront-end worktoday: Sonnet
  • Top modelComplex back-end worktoday: Opus

Four tasks are queued. In orchestrate mode each one is routed by complexity; in one long session the same agent does all of them.

Today: Haiku / Sonnet / Opus. The tiers are the idea; the model names change every few months. Illustrative percentages of one context window, to show the shape, not to measure a tool.

Tools usedThe CuratorConduit and Atlas
Go deeper
03

Multi-harness work

I move between coding tools and between computers on the same project. That only works if the project's state does not belong to any one tool. Step through a session that changes tool and machine and still picks up where it stopped.

Step 1 of 7
Machine A
Claude Codescope: claude-code
opencodescope: opencode
Machine B
Antigravityscope: antigravity
A Curator project · a folder you ownplain markdown
  • Foundationscanonical documentsreading
  • Standing briefyour standing ordersreading
  • Handoffscope: claude-codenot saved yet
  • Handoffscope: antigravitynot saved yet
  • Handoffscope: opencodenot saved yet
  • Journalrecent sessionsreading

Session starts. Claude Code opens the project and makes one call to read its context: foundations, the standing brief and the latest handoff.

All of this is plain markdown in a folder you own. It is portable, and it is not locked to any one vendor, tool or computer.

Tools usedThe CuratorAgentLink
Go deeper
04

Memory that compounds

Not all context is the same, so it should not be written the same way. Documents are kept verbatim, working state is replaced at each save, and knowledge accumulates. Select a tier to see what it holds and who writes it.

Project contextWhat an agent reads at the start of a session
KnowledgeCompounded, pulled by relevance

Agents read this at session start and write it as they work.

Foundations
Holds
The documents the project is built against: plan, architecture, decisions. Some are read first, some on request.
Written by
You, or an agent you commission. A model never rewrites one.
Changes
Rarely. When one changes, the document is replaced whole.
Rule
Read verbatim

Knowledge accumulates. State supersedes. A canonical document is replaced whole and read verbatim.

All of it is plain markdown on your own computer, read and written by agents through The Curator's built-in MCP server. It runs on your own API key, from one of three providers.

The Curator at mycurator.xyz⭐︎ Star The Curator on GitHub
Tools usedThe Curator
Go deeper