Hatch Capital Consulting

Build Your Company Brain

Week 2: turn business context and one workflow into agent-ready memory
July 29, 2026
AI Bootcamp | Context + Workflow Readiness
The course arc

Week 1 named the pieces.
Tonight we build the context.

Brain, Body, Context, Prompt still drive the whole bootcamp. Tonight is not an install party. Tonight is where your business becomes legible enough for an agent to help.

Brainthe model thinking through the work
Bodythe runtime or tool layer that acts
Contextthe company brain and workflow truth
Promptthe instruction and quality standard
Tonight's focus
Maximize Context and Prompt first. Week 3 is where we choose the Brain + Body/runtime layer and connect tools.
Four-week map

The bootcamp moves from understanding to execution.

Week 1Brain, Body, Context, Prompt
->
Week 2Company brain + workflow context
->
Week 3Runtime, tools, first build
->
Week 4Run, proof, refine

The order matters. A powerful runtime on top of messy context creates messy automation.

Definition

What is a company brain?

A company brain is the organized place where your agent can find the truth of the business.

Business contextwho you serve, what you sell, how the company works.
Workflow truthhow work starts, moves, gets reviewed, and finishes today.
Examplesgood output, bad output, messy cases, screenshots, source docs.
Receiptsproof the work happened: files, logs, screenshots, sent items, decisions.
Important
Obsidian is not where you do all the work. Obsidian is where your agents store, retrieve, and organize the truth so they can produce the work in whatever format you need.
Why this comes first

Every valuable workflow starts with context.

A workflow cannot become useful just because an agent can click buttons. It becomes useful when the agent understands the people, rules, SOPs, examples, source systems, exceptions, and proof standards around the work.

People contextCRM notes, roles, preferences, history, promises, owner/client/vendor relationships.
Process contextSOPs, checklists, examples, edge cases, review rules, templates, decision criteria.
Project contextgoals, current status, related people, source meetings, open questions, next actions.
Proof contextwhat counts as done: sent item, screenshot, file, calendar event, system log, reviewed output.
Homework frame
Between tonight and next week, start moving the context required for your selected workflow into Obsidian and linking it.
The obvious question

Why not just connect the agent to every platform?

You will eventually connect tools. But a CRM, SOP doc, Drive folder, PMS, Slack channel, or spreadsheet usually shows only one slice of the business.

CRMperson + pipeline slice
+
Drive / SOPprocess + file slice
+
Projectswork + owner slice
->
Obsidianconnected context layer
Core point
Connecting to tools gives the agent access. A linked company brain gives the agent understanding.
Neural context

Wiki links make context come alive.

When one person connects to meetings, projects, SOPs, outputs, and receipts, the agent can follow the same relationship map a human builds in their head.

[[Maya Chen]]person
[[BrightPath Wellness]]business
[[New Agent Onboarding]]workflow
[[Monday.com]]source system
[[School of Hard Knocks]]related project
[[Hatch Capital Consulting]]relationship context
[[Onboarding Checklist]]SOP / output
[[Review Gate]]human decision
Agent benefit
The agent does not see an isolated contact record. It sees the connected business context around that person, workflow, and project.
Fresh vault

Start with a simple structure agents can understand.

AI Bootcamp Company Brain/ 00-Inbox/ 10-Business-Context/ 20-Workflows/ 30-Source-Systems/ 40-Prompts/ 50-Outputs/ 60-Receipts/ README.md

This is not about having a beautiful vault. It is about giving an agent a reliable map of the business.

Step 1 - Prompting

Point Claude Cowork at the vault folder.

First, give Claude Cowork access to the fresh Obsidian vault folder. Then use the prompt pack to create the starter structure and detailed folder READMEs.

You are pointed at my fresh Obsidian vault folder. Create the starter company-brain folder structure in this vault. Create README.md files inside every folder. The READMEs are agent instructions. Each README should include: - what belongs there - what does not belong there - note naming rules - required note sections - wiki-link rules - agent instructions - Open Questions rules Use this structure: 00-Inbox/ 10-Meetings/ 20-People/ 30-Projects/ 40-Workflows-and-SOPs/ 50-Source-Systems/ 60-Content-and-Outputs/ 70-Prompts/ 80-Receipts/ 99-Archive/ Do not create fake business notes yet. Only create the structure and README instructions.
Student prompt pack
Open 02-Company-Brain-Prompt-Pack in the shared Week 2 resources folder.
Step 2 - Download the artifact

Give Claude Cowork one fake client meeting.

Open the BrightPath Wellness example transcript from the shared Week 2 resources folder, save it inside the fresh vault, then tell Claude Cowork exactly where the file lives.

Shared resourcedownload training transcript
->
Save in vault00-Inbox or 10-Meetings
->
Run ingest promptpopulate linked notes
Training transcript
Open 03-Fictional-BrightPath-Fathom-Transcript in the shared Week 2 resources folder.
Step 3 - Ingest

Tell Claude Cowork to read the READMEs first.

The folder READMEs are the operating rules. Claude should read them before deciding where anything belongs.

You are pointed at my Obsidian company-brain vault. First, read the main README.md and every folder README.md so you understand the vault rules. Then read this file: [PASTE SAVED FILE PATH HERE] This is fake training data. Preserve the native transcript as a meeting note in 10-Meetings/. Then create linked notes in the correct folders for people, projects, workflows/SOPs, source systems, outputs, and receipts. Use Obsidian wiki links everywhere useful. Do not invent facts. Put unclear items under Open Questions. After updating the vault, summarize the files created and explain the graph connections.
Starter vault

Use folders that work for almost any business.

AI Bootcamp Company Brain/ 00-Inbox/ 10-Meetings/ 20-People/ 30-Projects/ 40-Workflows-and-SOPs/ 50-Source-Systems/ 60-Content-and-Outputs/ 70-Prompts/ 80-Receipts/ 99-Archive/ README.md

Every folder gets a README. The README tells the future agent what belongs there, what does not, and how to update the notes.

Agent instructions

READMEs turn folders into operating rules.

Meetings READMEstore transcripts, summaries, attendees, decisions, and action items.
People READMEone note per person with company, role, meetings, projects, workflows, and action items.
Projects READMEgoals, owners, related people, source meetings, decisions, and next actions.
Workflows READMEtrigger, inputs, steps, systems, human gates, outputs, receipts, missing info.
Why this matters
The README is not for decoration. It is the instruction layer the agent can read before updating that part of the brain.
Graph view

Wiki links make context retrievable.

Names should connect to meetings. Meetings should connect to projects. Projects should connect to workflows. Workflows should connect to source systems and receipts.

[[Maya Chen]]person note
[[BrightPath Wellness]]company context
[[New Patient Inquiry Follow-Up]]workflow
[[HubSpot]]source system
Agent benefit
The graph shows relationships. The agent can follow links from a person to the meeting, from the meeting to the workflow, and from the workflow to the source systems.

Links should explain how the work moves.

Triggerevent or request
→
Meeting / inputsource artifact
→
Workflowsteps + decisions
→
Outputdraft / update
→
Receiptproof / blocker
PeopleProjectsSource systemsRulesExamplesOpen questions
Readback
Ask: “Which source controls this fact, and which workflow step uses it?”
Clean before real use

Archive the example data.
Keep the structure.

After the demo, the fake BrightPath notes should move into an example archive. The READMEs and folder structure stay.

Archive all example information into: 99-Archive/Example-Training-Data/ Keep: - folder READMEs - main README - empty starter structure Next: drop in a real transcript, screenshot, note, or workflow walkthrough.
Homework setup

Move workflow context into the brain.

Once you choose the workflow, identify every piece of context an agent would need to do that work well. Then start moving or summarizing that context into Obsidian.

If it lives in the CRMexport or summarize the people, company, pipeline, promise, and relationship context.
If it lives in Google Drivebring in the SOPs, checklists, templates, examples, screenshots, and process docs.
If it lives in a tooldocument what the tool stores, what the workflow needs from it, and how an agent might access it later.
If it lives in your headrecord a walkthrough. Say the judgment, exceptions, stop signs, and what good output looks like.
Goal before Week 3
Do not connect everything yet. Build enough connected context that the first tool test has something real to operate against.
Workflow context prompt

Have Claude interview the workflow out of you.

Interview me step by step to extract this workflow: [WORKFLOW NAME] Ask one section at a time: - what starts the workflow - who owns it - what tools it touches - what data lives where - what steps happen now - what triggers the next step - where judgment happens - what exceptions can happen - what fallback rule to use - what output good looks like - who reviews it - what receipt proves it happened Then create an agent-ready workflow note with wiki links and Open Questions.
Example exception
If the workflow creates a new email and the preferred name is already taken, the workflow needs the backup naming rule before an agent can run it safely.
Pick one workflow

Choose a wedge,
not the whole business.

QuestionWhat we are testing
Does it repeat?Agents need patterns, not one-off chaos.
Does it have a trigger?Something has to start the work.
Does it touch systems?Email, PMS, Monday, Drive, Slack, spreadsheets, docs.
Does judgment happen?We need to know what humans still decide.
Is there an output?Brief, checklist, draft, update, proposal, owner note.
Can we review it?No review gate means no safe first workflow.
Workflow audit

The questions that make a workflow agent-ready.

StartWhat triggers the workflow? Who owns it today?
InputsWhere does the information come from? What source systems matter?
StepsWhat happens now? What decisions are being made?
Stop signsWhen should the agent ask instead of acting?
OutputWhat should the agent produce? What does good look like?
ProofWho reviews it? What receipt proves the work happened?
Runtime preview

Tonight is context and prompts.
Next week is the body.

Claude / ChatGPT todayFastest way to organize messy notes, draft workflow docs, and create prompts.
Obsidian nowPortable company brain the agent can later read and update.
OpenClaw nextRepeatable execution layer for files, tools, APIs, browser, receipts, and real workflows.
Key line
The runtime does not matter yet. Every runtime gets better when the company brain is clean.
Homework

Bring connected context for Week 3.

Starter company brain
One selected workflow
Workflow walkthrough recording
CRM / people context
SOPs from Drive
Source docs or screenshots
Good output example
Messy edge case
Tool/system list
Wiki links between notes
Review gate
Proof receipt idea
Recording prompt
Start with the trigger. Walk through the real systems. Say what you are thinking. Say where judgment happens. Say what good output looks like. Say what can go wrong.
Homework

Your homework.

Turn one real workflow into connected, reviewable context—then bring the evidence.

Archiveremove fictional facts
→
Interviewextract the real process
→
Gathercollect only what matters
→
Gatename stop signs
→
Receiptprove the result
Prompt 3

Archive the lesson. Keep the structure.

1
Move every fictional note to the training archive.Meeting, people, project, workflow, source-system, output, and receipt notes.
2
Keep all READMEs and the starter folders.The operating system remains; the example facts leave active lanes.
3
Create an archive label.State clearly that the archived notes are fictional training data.
4
Verify by search.BrightPath should appear only in the archive—not in active business lanes.
Archive verification
Do not add real company context until the fictional entities are cleanly separated.
Prompt 4

Interview the real workflow one section at a time.

SectionWhat to extract
Overview + triggerOutcome, why it matters, what starts it, and what “now” means.
Owners + inputsProcess owner, participants, required data, authoritative systems.
Steps + judgmentCurrent sequence, decisions, rules, and where a human still decides.
Exceptions + outputMissing fields, duplicates, fallbacks, examples, anti-examples.
Review + proofApproval gate, receipt, links, and open questions.
Prompt constraint
“Ask me one question at a time. Do not give me a giant form.”
Prompt 5

Gather only the context this workflow needs.

PeopleRoles, relationship history, preferences, commitments, process owner.
Process + SOPChecklists, policies, templates, examples, approved language.
ProjectStatus, goals, next actions, related meetings and decisions.
Source systemsWhat each system stores, what controls, freshness, access questions.
Stop signsMissing, conflicting, stale, sensitive, high-consequence decisions.
Output + receiptDestination, format, review state, proof of completion or blocker.
Workflow readiness

The context map should answer nine questions.

What starts it?Who owns it?Which sources control?What steps happen?Where is judgment?What can go wrong?When must it stop?What does good look like?What proves it?
Honest readiness
If an answer is missing, record the owner and the open question. A diagnosed gap is better than fabricated completeness.
Stop signs + human review

A review-first workflow knows when not to act.

Missing or conflicting factStop, cite the conflict, ask the named owner, update the durable note after review.
External communicationDraft only. Human approves before send, publish, or customer-facing change.
Money, legal, compliance, accessNo purchase, charge, credential change, binding decision, or sensitive release.
Unexpected tool behaviorStop and produce a blocked-run receipt instead of improvising.
Tonight's boundary
No sends, publishes, deletes, purchases, account changes, or irreversible actions.
Receipts before trust

“Done” is a claim. A receipt is evidence.

RequestedOutcome, scope, approved sources, review gate.
AttemptedFiles read, prompts used, tools touched, time.
ChangedCreated, updated, moved, or explicitly unchanged.
ReviewedHuman reviewer, decision, edits, approval status.
BlockedExact failure, last proven step, no false success.
NextOpen questions, owner, smallest safe follow-up.
Current cohort — generic working lanes

Different businesses; the same context discipline.

Knowledge consolidationTurn scattered research and notes into a sourced, structured knowledge base.
Operational handoffMake recurring communications, follow-up, or coordination steps visible and reviewable.
Process-owner interviewWhen expertise lives with another operator, interview that person rather than inventing the process.
Activity
Choose one lane tonight. Describe one trigger-to-receipt path without exposing private client or regulated data.
1/22