A pattern for building persistent, context-rich AI workspaces around your projects — so the agent knows your work instead of starting from scratch every conversation.
Create a dedicated project directory on your machine. Open a Claude Code (or Codex) session in it. Every day, just talk to the agent — tell it what you're doing at work, what's coming up, what you're stuck on.
Over time, the agent builds deep understanding of your job, your responsibilities, your project's quirks. Instead of Googling or asking ChatGPT on the web (where you start from zero context every time), you ask inside your project directory — and the agent already knows 80% of the background.
mkdir -p ~/Documents/my-project
cd ~/Documents/my-project
claude # or open in Codex / Claude CodeTell it: "I'm working on [project]. Here's the high-level scope..." and let it help you organize from there.
Ask the agent to create a structure that tracks each workstream independently. Here's what a real enterprise cloud migration project looks like after 3 months of accumulated context (~600 docs):
project/
├── context/ # Background & reference material
│ ├── customer/ # Customer-specific context
│ │ ├── legacy/ # Historical docs, shared docs
│ │ └── raw/ # Raw inputs (emails, specs, dumps)
│ └── vendor/ # Vendor product docs, reference apps
│
├── work/ # Active workstreams (one dir per track)
│ ├── infrastructure/ # Infra design, cloud, observability
│ ├── auth/ # Authentication flows, SSO, identity
│ ├── cicd/ # GitOps, pipelines, Helm charts
│ └── spike-poc/ # Spike / PoC work
│
├── deliverables/ # Polished outputs for customer/mgmt
│ ├── infrastructure-design/
│ │ ├── current/ # Latest version
│ │ └── emails/ # Sent correspondence
│ └── cicd/
│ └── current/
│
├── shared-assets/ # Reusable across workstreams
│ ├── product-docs/ # Official product documentation
│ ├── helm-charts/ # Chart source for reference
│ └── templates/ # DOCX/PPTX corporate templates
│
├── memory/ # Agent's accumulated knowledge
│ ├── context/ # Project-level facts
│ ├── projects/ # Per-workstream memory
│ └── people/ # Stakeholder notes
│
├── governance/ # Index of all artifacts, decisions
│ └── ARTIFACT_INDEX.md # Master registry of every doc
│
├── scratch/ # Throwaway exploration
├── tmp/ # Build outputs, draft versions
├── tools/ # Custom CLI tools built for the project
└── .skills/ # Agent skills (project-specific)
├── branded-docx/ # Generates branded Word docs
└── excel-reports/ # Generates Excel reports
work/— Each workstream is isolated. When you switch context from auth to infrastructure, the agent switches with you. No cross-contamination.context/— Raw reference material the agent can search when answering questions. Customer specs, vendor docs, meeting dumps — all searchable.deliverables/— Clean separation between working docs and what goes to the customer.memory/— The agent's own knowledge base. Stakeholder preferences, decisions made, project-specific terminology.governance/— An artifact index so the agent (and you) always knows what exists and what's current..skills/— Project-specific agent capabilities (more on this below).
The agent remembers your project's terminology, people, decisions, and patterns. When you say "the Phase 1 deployment issue", it knows you mean the specific deployment problem and the root cause — no explanation needed.
Every doc, design decision, and meeting note becomes searchable. Instead of digging through emails or SharePoint, you ask: "What did we decide about DR failover?" and get an answer with the source file cited.
The search is hybrid — keyword matching for exact terms plus semantic/vector search for concept-level queries. So "what's the auth flow for SSO users" finds docs even if they never use the word "flow".
The agent learns how to produce outputs your way. Examples from a real workspace:
| Skill | What It Does |
|---|---|
| branded-docx | Generates client-ready Word docs with corporate branding, headers, footers, table styles. No manual formatting. |
| web-search | Researches vendor docs, KB articles, cloud docs in context of your project — not generic results. |
| excel-reports | Builds tracking spreadsheets, migration matrices, status reports. |
| browser automation | Logs into admin consoles, extracts config, takes screenshots for documentation. |
| custom CLI tools | The agent can help you build project-specific tooling — CLI wrappers for APIs, bulk operations, etc. |
Each track (auth/, cicd/, infrastructure/) has its own docs, diagrams, and archive. Context stays contained. You can even have different agent instructions per workstream.
Docs are versioned (v0.1.0, v0.2.0). Old versions move to archive/. You always know what's current, and you can trace how a design evolved.
This is the key part. It's not a big setup exercise — it's a 5-minute daily habit:
"Today I'm working on X. Here's what happened in the call..."
"We decided to go with approach B because..."
"I need to prep a design doc for the DR review on Thursday"
"What's the status of the auth workstream? What's still open?"
The agent captures it, organizes it, and builds cumulative context. After a few weeks, it's like having a technical EA who was in every meeting and read every doc.
| Without project context | With project context |
|---|---|
| Every conversation starts from zero | Agent already knows the background |
| You re-explain acronyms, people, decisions | Agent remembers all of it |
| Generic search results | Answers grounded in your project docs |
| Manual doc formatting | Agent produces branded deliverables |
| Context scattered across email, chat, wiki | Single searchable source of truth |
| "Let me Google that" | "Let me check our project docs" |
There's more depth to this pattern that's better covered in a walkthrough:
- Scheduled agents — Automated daily/weekly tasks (status rollups, inbox digests, build monitoring)
- Operator runbooks — Step-by-step procedures the agent can execute against live infrastructure
- Cross-project context — Sharing knowledge between related project workspaces
- Agent-driven document generation — Full design docs, architecture specs, proposals from a single prompt
- Local document search engine — BM25 + vector search over hundreds of project docs with sub-second results
Built with Claude Code