Escape from Theory Land · Bonus pack #1

Beyond the Chat

A practitioner's guide to long projects with Claude, for non-developers

Give the project a memory that you own and that Claude can read.

A chat is a good place to think and a poor place to keep a project.

One long chat Decision Older messages: summarised, detail lost
A few weeks later, the chat began to forget.
Prologue

Three books, one AI, and a chat that forgot

I wrote this guide because of a problem I did not see coming.

Over several weeks, I worked with Claude on a series of three practical books about AI in organisations. Claude helped me structure, draft, check facts and lay out the pages. The first two volumes are finished, at around a hundred pages each. The third was in progress when I wrote this guide.

At first, everything happened in one chat. It went very well, so I kept going.

A few weeks later, the chat began to forget. A decision we had made disappeared. An idea I had dropped came back as if it were still valid. I explained the same background again and again. I knew we had solved a problem before; I could not find where.

The AI had not become less intelligent. The project had.

Each time something broke, I added a piece: a document, a rule, a new way of organising the chats. Some pieces failed and were replaced. What remained is the method in this guide.

Who this guide is for

Anyone who uses Claude for work that lasts longer than one conversation: a book, a strategy, a research topic, a product, a transformation programme. You do not need to be technical. There is no programming, no version-control tool and no special software.

In the guidePrologue · Three books, one AI, and a chat that forgotp. 6

Stage 0One long chat

A chat is a place to think, not a place to store.

The problem

A chat is a good place to think and a poor place to keep a project. After a few weeks, a long chat loses details. Decisions get argued twice. Two chats contradict each other, and neither knows.

For now, everything lives in one chat: drafts, decisions and questions. When it grows too long, older messages are summarised and detail is lost.

The answer. Give the project a memory that you own and that Claude can read. Then add structure only when a real problem asks for it.

The seven additions. Each one answers a problem I actually hit while writing a book series with Claude.

Try this. Open your longest chat and ask: "List the decisions we made in this conversation, with the date of each." Check three of them against what you remember. If one is wrong or missing, you have the problem this guide solves.

In the guideChapter 1 · A chat is a place to think, not a place to storep. 10

Stage 1Project, files

What matters lives in a document you own.

The problem

One long chat forgets what was said weeks ago

What I added

A project, with instructions and files

The chats now sit inside a Claude project, with instructions, a knowledge base of files, and Claude's memory beside them. The state of the project still lives in a static file.

Try this. Rewrite your project instructions in under 150 words. Keep only what every chat must know before doing anything. If you need more, you need the next chapter.

In the guideChapter 2 · Give the project a homep. 15

Stage 2The hub

Every chat reads the hub first.

The problem

The files go stale, and Claude cannot update them

What I added

The hub: one living doc that every chat reads first

The project instructions shrank to a few lines. Their main job became one sentence: read the hub first.

What it does not hold matters as much. There are no to-do items, no drafts and no chapter text. Those live in the pending lists and in the documents themselves.

HubA living document with revisions

Example

Read this first in every new chat about the series. It replaces the handover file. To-do items stay in each volume's pending list, not here.

4. Status
Volume 1 · The Map. Fully laid out: 101 boards on the canvas. Frozen since
7 October; the canvas is the version of record. Anything still open is in
its pending list.

Blank

Hub · [project]
Read this first in every new chat. To-do items stay in the pending lists, not here.
Revision: [R1] · [date]

1. Where everything lives
   [Document] · [link] · [role] · [master or mirror]
2. Rules
   Content: [how the work is written]
   Working: [how every chat works]
3. Decisions
   Settled positions: [views the project rests on]
   Log, newest first: [date] · [decision] · [source]
4. Status
   [part] · [where it stands] · [link to its pending list]
Project instructionsPlain text, under 150 words

Blank

This project holds [your project]. Every new chat first reads the hub: [link].
It holds the links, rules, decisions and status.
Do not rely on older summaries or on memory for status.
If this chat has been summarised, re-read the hub.
Then read the pending list of the part you are working on.
Only one chat edits a document at a time. Re-read a document before changing it.
Claude proposes closing items; only I close them.
Every turn ends with an alignment report.

Try this. Create a doc called "Hub" with four headings: where everything lives, rules, decisions, status. Then replace your project instructions with the eight lines above. The skeleton is in appendix A.

In the guideChapter 3 · One page to read first: the hubp. 22

Stage 3Pending lists

Nothing stays buried in a chat you will never reopen.

The problem

Open questions are scattered across chats

What I added

A pending list for each part of the work

Each part of the work now has its own pending list. Like the hub, each list is a Claude doc, and the hub's table of links points to it.

Pending list entryLiving documents

Example

J1 · Sustainability chapter
Type: decision needed · Status: open

Question: Volume 1 promises a topic on the environmental cost
of AI. Volume 3 has no chapter on it. Add one, or drop the promise?
Context: volume 1, page 96, "What comes next".

Decided on 7 October: its own short chapter, between software
development and measuring.
Applied in the doc on [date]: chapter outline added.
Still to do: write the chapter.

Blank

[ID] · [Subject]
Type: [question / decision needed / action / risk / idea / waiting]
Status: [open / next / in progress / waiting / proposed for closing / closed]

Question: [written so that it reads on its own]
Context: [where in the work, with page or section]

Decided on [date]: [decision, in the author's words]
Applied in the doc on [date]: [what changed]
Still to do: [what remains]

Try this. Ask Claude: "Read the comments in my documents and our recent chats. Build a pending list: one entry per subject, each question written so that I understand it without reopening the work. Do not answer any question yourself."

In the guideChapter 4 · Everything still open, in one place: the pending listp. 28

Stage 4Decisions log

A decision is never edited. It is replaced, and the old one stays visible.

The problem

Decisions get reopened or quietly contradicted

What I added

A decisions log, where decisions are replaced, never edited

A proposal becomes a decision only when you say yes. After that, the decision is never edited. If its meaning has to change, a new decision replaces it. The old one stays visible, marked as replaced, so you can see what governed the work and when.

7 OctThe volume 3 memo has no "To:" line, like the memos in volumes 1 and 2.
7 OctThe volume 3 memo would have one.replaced
Decision

Example

7 Oct 2026 · The volume 3 memo has no "To:" line, like the memos in volumes 1 and 2. This replaces the decision taken earlier the same day, which rested on a wrong reading of those volumes. · Volume 3 pending list, A15

Blank

[Date] · [Decision, one or two sentences]
Rests on: [the fact or reading it depends on]
Replaces: [earlier decision, or none] · Replaced by: [later decision, or none]
Source: [who decided, and where]

Claude should never resolve a contradiction silently, not even when the contradiction comes from you. It stops and shows a conflict block:

Conflict block

Example

CONFLICT
Existing decision (7 Oct): volume 3's executive memo prints no "To:" line.
New proposal: address the memo to transformation leads.
Nothing has been changed.
Options: reject the proposal · reword the decision · replace the decision.

Blank

CONFLICT
Existing decision ([date]): [decision].
New proposal: [proposal].
Nothing has been changed.
Options: reject the proposal · reword the decision · replace the decision.

Try this. Ask Claude: "List the decisions in this project that are not in the log yet, with the date and the chat where each was made. Do not add anything until I confirm."

In the guideChapter 5 · Decisions you never argue twice: the decisions logp. 34

Stage 5One chat, one job

One chat, one job.

The problem

Chats grow too long and step on each other

What I added

One chat per job, and a coordinator chat

The single chat per volume is gone. A coordinator chat reads and proposes; working chats, one job each, write in the pending lists and chapter docs within their scope; you carry the prompts and reports between them.

Coordinatorchatedits nothingYouWorking chatone job1234
  1. 1The coordinator proposes,
  2. 2you pass the prompt on,
  3. 3the working chat does the job within its scope and reports,
  4. 4and you bring the report back.
Working chat prompt

Blank

This chat has one job: [job].
Read the hub first: [link]. Then read the pending list: [link].
In scope: [what to do]. Out of scope: [what not to touch].
The only document you may change is [document].
Before starting, estimate the work and split it into passes. Tell me the plan.
For every check, say how it was done.
Never close an item; propose closing it, with the evidence.
If you find anything else to fix, add it to the pending list instead of fixing it.
End every turn with an alignment report.
Coordinator chat prompt

Blank

This chat coordinates [part of the project]. It edits nothing.
Read the hub and the pending list first.
When I paste a report from a working chat:
1. check it against the pending list and flag anything missing or wrong;
2. give your recommendation on each open question it raises;
3. propose the next step: say which chat to use (this one, an existing one
   by name, or a new one) and give the full prompt to paste there.
Never propose two chats that change the same document at the same time.

Try this. Start your next chat with: "This chat has one job: [job]. Read the hub first. The only document you may change is [document]. If you find anything else to fix, add it to the pending list instead."

In the guideChapter 6 · One chat, one job: the coordinator and working chatsp. 40

Stage 6One writer, one master

One master for each piece of content. Everything else is a mirror.

One writer at a time. Re-read before you write.

The problem

Work gets overwritten or maintained twice

What I added

One writer at a time, and one master for each piece of content

The chapter docs are declared the master, and the layout canvas mirrors them. The hub now also records what frozen work prints and what it promises.

If an artifact changed since the last read, re-read it, keep the other changes, and reapply yours on top.

Chapter docsthe masterLayout canvasthe mirror

A mirror is never edited first.

Try this. For each kind of content in your project, write one line in the hub: "Master: [document]. Mirrors: [documents]." If you cannot say which is the master, you have two masters.

In the guideChapter 7 · Protect the work: one writer, one masterp. 46

Stage 7Honesty rules

"Verified" always says how.

The problem

Claude claims checks it did not make

What I added

Honesty rules: say how, say what was saved, report every change

This is the complete setup. Your preferences shape every chat from your account settings. Every chat opens with three lines, and every turn ends with an alignment report.

  1. the revision of the hub it read;
  2. the state of the project as it understands it;
  3. the decisions that touch today's subject.
Weak
Strong

Checked: the metaphors we removed no longer appear in volume 1.

Checked by a text search of all 101 boards: none of the removed metaphors appear.

The legal notices are identical.

The legal notices were compared by a text diff after normalising spaces: identical, word for word.

Alignment report

Example

Alignment report
- Chapter 3 doc: section 3.2 rewritten (revision 41).
- Layout: not changed; page 34 still to mirror.
- Pending list: A4 marked "applied in the doc"; proposed for closing.
- Hub: not changed.

Blank

Alignment report
- [Document]: [what changed] ([revision]).
- [Mirror]: [changed / not changed; what is still to mirror].
- Pending list: [items updated; items proposed for closing].
- Hub: [changed / not changed].
Saved: [new text only / new file / copy changed / reference document changed]

Try this. Add two lines to your project instructions: "Every turn ends with an alignment report." and "'Verified' always says how it was checked."

In the guideChapter 8 · Keep Claude honestp. 52

Reference card

The routine

Set up once

Create a Claude project. Create a Claude doc called "Hub" with four headings: where everything lives, rules, decisions, status. Write project instructions in under 150 words that point to the hub. Ask Claude to fill the hub from your existing material, without inventing anything. Add your stable sources to project knowledge, as clean .md or .pdf files.

Every chat

Give it one job and one scope. It reads the hub first, then opens with three lines: the hub revision it read, the state as it understands it, the decisions that apply. It works in passes if the job is long. It ends every turn with an alignment report.

Every decision

It becomes a decision only when you say yes, in the chat or in a comment on its entry. It goes in the log with its date, its source and what it rests on. If it contradicts an earlier decision, Claude stops and shows a conflict block. A decision is never edited; it is replaced.

End of a phase

Move closed items out of the pending lists. Freeze what is finished, and record what it prints and promises. Run a self-audit in each long chat. Start a new coordinator chat with a short hand-off note.

Chapter 9Use only as much as you need

Use only as much structure as the project needs.

The same architecture, sized to the project. Light keeps the hub and nothing more; strict uses everything from chapters 3 to 8.

Keep a project memory as soon as one of these is true:
Go strict as soon as one of these is true:

Just chat

Just chat if everything fits in one conversation and nothing will be reused.

What to add, and when
AddWhen
A pending list for each partMore than about ten open points, or more than one part of the work
A decisions log in the hubYour first decision you would not want to argue twice
One chat per jobA chat gets summarised, or runs out of actions mid-task
A coordinator chatThree or more working chats on the same project
One declared masterThe same content exists in two forms, for example text and layout
Frozen work and promisesYou deliver the work in parts, and later parts depend on earlier ones
A self-auditAt the end of each phase

The thresholds are rules of thumb from my own project, not measurements. Adjust them to yours.

Try this. Look at your current project and answer the three questions above. Then apply the matching mode, and nothing more.

In the guideChapter 9 · Use only as much as you needp. 61

The complete setup
New at this stage Added earlier To come !The problem

When things break

Start from the symptom.

01Claude forgets a decision from weeks ago

Likely cause

It lived only in a long chat, now summarised

Fix

Write it in the decisions log; start a new chat

Chapter 1Chapter 5

02A new chat works from old information

Likely cause

It read a project file that is out of date

Fix

Move changing information to the hub; keep files for stable sources

Chapter 2

03Claude misses a rule you know is written down

Likely cause

The rule is buried in a large set of files that Claude searches instead of reading

Fix

Put the rule in the hub, which is read in full

Chapter 2Chapter 3

04The same question is answered twice, differently

Likely cause

Open questions live in several places

Fix

One pending list per part, declared the single source of truth

Chapter 4

05An old idea comes back as if it were decided

Likely cause

An idea was never marked as one, or a replaced decision was erased

Fix

Mark ideas as ideas; replace decisions, never edit them

Chapter 5

06A decision turns out to rest on a mistake

Likely cause

Nobody wrote down what it rested on

Fix

Add "rests on" to every decision; review those that depend on the mistake

Chapter 5

07A long operation stops halfway

Likely cause

The chat ran out of actions for the turn

Fix

Estimate first; work in passes

Chapter 6

08Another chat's changes have disappeared

Likely cause

Two writers on one document, or a write from a stale copy

Fix

One writer at a time, you included; re-read before writing

Chapter 6Chapter 7

09A correction appears in the layout but not in the text

Likely cause

Two copies, no declared master

Fix

Declare the master; change it first; mirror afterwards

Chapter 7

10New work contradicts finished work

Likely cause

Nothing records what the finished work says or promises

Fix

Freeze it; record its content and its promises in the hub

Chapter 7

11"Updated", but nothing changed

Likely cause

New text was written in the chat, not in the document

Fix

Ask which of the four outcomes happened; check the revision

Chapter 8

12Claude closed an item you had not decided

Likely cause

No rule on who closes

Fix

Claude proposes closing; only you close

Chapter 4Chapter 8

13Claude's memory contradicts the hub

Likely cause

Rules or status were copied into memory

Fix

Keep memory for preferences; the hub wins

Chapter 2Chapter 8

14The project keeps growing beyond its plan

Likely cause

Every good idea gets added

Fix

A parked-ideas section and a size budget

Chapter 4Chapter 9

15You spend more time on the method than on the work

Likely cause

Too much structure for the project

Fix

Drop to the mode the three questions point to

Chapter 9