Loading

← Claude Operator Track

Projects, memory, and durable context

> Stop re-explaining your company every chat — put standing instructions where Claude can reuse them.

Canonical: https://www.brainyxai.co.za/education/claude/projects-and-context

Markdown: https://www.brainyxai.co.za/md/education/claude/projects-and-context.md

Course: Claude Operator Track

Lesson: 3 of 9

Minutes: 14

Author: Brainyx AI

What you will be able to do

  • Decide what belongs in a Project vs a one-off chat
  • Write a short standing instruction block without secrets
  • Know when to reset context vs continue a long thread

Why Projects exist

Long threads rot: the model loses the plot, you forget what was decided, and secrets sneak into the scrollback. Projects (and similar project/workspace features in the Claude product) exist so standing context lives outside a single conversation.

Put in Project knowledge / instructions:

  • Product one-pager and glossary
  • Brand voice (short)
  • “How we write support replies”
  • Architecture sketch (non-secret)
  • Definition of done for common tasks

Never put in Project files you’ll share carelessly:

  • API keys, customer PII, payroll, credentials
  • Anything your compliance policy forbids in third-party AI tools

Memory vs instructions

  • Instructions — how Claude should behave every time
  • Memory / recalled prefs — durable facts about you (review and prune)
  • Thread context — ephemeral; compact or start fresh when quality drops

Operator habit: after a messy thread, start a new chat and paste only the decisions that matter.

CLAUDE.md foreshadowing

If you use Claude Code, you’ll meet `CLAUDE.md` — repo-level standing rules. Same idea as Project instructions, but for engineering. We’ll cover it in the Claude Code lesson.

Brainyx AI production note

For client systems we externalize context into owned stores (RAG, runbooks, policy packs) with access control — not only a chat Project. Use Projects to learn the habit; graduate important knowledge into systems you control.

Context lab

1. Create or open a Claude Project for one real initiative.

2. Add a ≤400-word instruction block: goal, voice, must-not-do, output defaults.

3. Run the same brief in a blank chat vs the Project — note what improved.

Checkpoints

  • Standing context lives outside one thread
  • No secrets in Project uploads

Deciding what earns a place in standing context

Standing context is expensive: it is loaded every time, and stale entries quietly degrade every answer. Apply a simple test — does this fact change how the model should behave in most conversations for this initiative? If yes, it belongs in the Project. If it matters for one task, paste it into that chat.

Good standing context is short and decisive:

  • A glossary of internal terms an outsider would misread
  • Your voice rules stated as constraints, not adjectives ("no exclamation marks, no 'delve', name the customer's problem in the first sentence")
  • Definition of done for the two or three artefacts you produce most
  • What is out of scope, so the model stops offering it

Bad standing context is an archive. Uploading forty documents "so it has everything" reliably makes retrieval worse, not better.

Recognising context rot

Long threads degrade in a predictable pattern: the model starts referring to superseded decisions, re-raises settled questions, and hedges more. When you notice it, do not fight the thread. Open a new chat and paste a short decisions block — what was agreed, what is still open. You will get a sharper answer in one turn than in five turns of correction.

A useful habit: at the end of any long working session, ask for a decisions-and-open-questions summary, and keep that rather than the transcript.

Memory hygiene

Durable memory is convenient and quietly accumulates wrong facts — a role you no longer hold, a client you no longer serve, a preference you changed. Schedule a review. If a stored fact would embarrass you in an output next month, remove it now.

POPIA and Project uploads

This is where South African teams most often create risk without noticing. Project knowledge feels internal because it sits under your account, but it is still a third-party processor holding whatever you uploaded. Before you upload, confirm the file contains no personal information you have not cleared for that purpose, and no credentials at all.

For anything you genuinely need the model to reason over at scale — policy libraries, contracts, historical tickets — the right answer is retrieval over a store you control, not a chat Project. Brainyx AI builds those as [RAG knowledge systems](https://www.brainyxai.co.za/services/ai-implementation) so access control and audit stay yours.

Mini-FAQ

A: Projects for stable behaviour rules that fit in a page. RAG when the knowledge is large, changes often, or needs access control and citations.

A: Short enough that you would actually reread it. Beyond roughly 400 words, most teams are describing rather than instructing.

A: Yes, if their context differs materially. Cross-contamination between client contexts is a real and embarrassing failure mode.

Next lesson

Continue to [Chat vs Cowork](https://www.brainyxai.co.za/education/claude/chat-vs-cowork).

Official reference: https://claude.ai

Course hub: https://www.brainyxai.co.za/education/claude · Previous: https://www.brainyxai.co.za/education/claude/briefs-and-prompts · Next: https://www.brainyxai.co.za/education/claude/chat-vs-cowork

Return to the course hub, or explore Brainyx AI services to put the skill into a live system. Markdown: /md/education/claude/projects-and-context.md

Book a consultation · joshua@brainyxai.co.za · Markdown mirrors