Building content with Claude

Building content with Claude

Five ways to work. They differ in where the files live and how much you watch, and they all end at the same place: a branch, a passing gate, and a pull request somebody reviews.

Everything on this page was checked against the official Claude Code documentation on 10 September 2026. Where a version or a command matters, the source is linked.

---

Pick a lane in ten seconds

flowchart TB
    S["What are you about to do?"]
    S --> A{"Is it a whole<br/>day pack?"}
    A -->|yes| W["<b>Claude Code on the web</b><br/>a long run you check on"]
    A -->|no| B{"Are you arguing about<br/>shape, or writing files?"}
    B -->|arguing| P["<b>Claude Project only</b><br/>think first, no files yet"]
    B -->|writing| C{"Do you want to watch<br/>the diff as it happens?"}
    C -->|yes| D["<b>Desktop or terminal</b><br/>fast loop, local files"]
    C -->|no| W
    P --> BL["<b>The blended workflow</b><br/>settle the spine, then build"]

Loading

Lane	Where files live	Watch it?	Best for
Claude Code on the web	A cloud VM, pushed to a branch	Check back	A full day pack, the default
Claude Code on desktop	Your machine	Yes, continuously	Fixing one artifact, a tight loop, anything visual
Claude Project only	Nowhere, it is a conversation	Yes	Shaping a week, arguing about a spine, drafting before commitment
The blended workflow	Both, in sequence	Both	What most real work turns out to be
Manual steps and checks	Your eyes	Always	The things no tool does for you

---

What is the same in every lane

Four things do not change, and they are the reason a session written by somebody else still lands.

flowchart LR
    A["<b>CLAUDE.md</b><br/>loaded at the start<br/>of every session"] --> B["<b>the spine gate</b><br/>a one-screen plan<br/>you approve or reject"]
    B --> C["<b>passes, one family<br/>at a time</b><br/>deck, notebook, exercises"]
    C --> D["<b>verify.py</b><br/>the six proofs,<br/>one command"]
    D --> E["<b>a branch and a PR</b><br/>never main"]

Loading

1. CLAUDE.md is read at the start of every session, on every surface. It is the file that carries the naming rules, the folder layout, the banned words and the ground truth order. If a session is producing the wrong shape of thing, the fix usually belongs in CLAUDE.md rather than in the prompt.
2. The spine gate is not optional. A session states the envelope, the continuity and a one-screen spine, then stops. You approve or you reject. Nothing past the spine gets built without an explicit yes, because rejecting a spine costs a minute and rejecting a finished pack costs an afternoon.
3. One artifact family per pass. Decks, then notebooks, then the activity, then exercises with solutions, then the take-home, then the Kahoot pack, then study notes with the cheat sheet and the pre-read.
4. python3 scripts/verify.py content/W01/D3 is the gate. One command runs all six proofs. A pack that has not passed is not done, and it is not reviewable.

---

The one rule about sessions

One session per day pack. Do not continue yesterday's session for today's pack.

The reason is drift. A long session carries yesterday's decisions as context, and the second pack starts inheriting the first pack's shape without anybody choosing that. Start fresh from prompts/day_pack_prompt.md.

If you ask a session to build two days at once it should refuse and say why. That refusal is correct behaviour, not a malfunction.

---

What a good first message looks like

The worst opening is "build Week 4 Day 2". It gives the session nothing to check itself against. A good opening does four things:

Build the day pack for W04/D2.

Envelope: who this is for, what slot it fills, how much of their
effort it can absorb, what runs before it and what runs after.

Sources: the links I have already verified are below. Use these as
the lock rather than as a starting point.
  - <link>  checked <date>
  - <link>  checked <date>

Continuity: what landed on D1 that this day builds on, and what
D3 needs this day to leave behind.

Stop at the spine. I will approve it before you build anything.

The fourth line is the one people drop, and it is the one that saves the afternoon.

---

What the artifacts look like when it works

A cheat sheet, authored as markdown and printed by scripts/build_cheatsheet.py. Panel one is the anchor and carries the day's picture.	A slide, authored as markdown and built by scripts/build_deck.py through the shared layout in scripts/deck_layout.py.

Both are built from markdown by a script. Neither is edited by hand in its final form, because a .pptx older than its markdown is stale, and the gate fails it.

---

Where to go next

You want to	Read
Run a full day pack from a browser	Claude Code on the web
Fix one artifact quickly on your own machine	Claude Code on desktop
Think about a week before anything is a file	Claude Project only
Do both, in the order that works	The blended workflow
Know what you still have to do yourself	Manual steps and checks