Task queue
What needs doing, by whom and by when. Lives in the task tool. The knowledge keeps the decision and its context.
Personal second brain for business and sales
Without context, an agent
starts from zero every time.
Agents can already reason and act. What is missing is maintained context: what we know about the client, what we have decided and what has been proven. A second brain is the vault for this context, read by both people and agents.
How agents use itFor readers who know agents. All examples and figures are invented.
Source: note after a call
“We want to start with one branch. The manager makes the decision. Follow up again on Thursday.”
Context prepared by the agent
- Goal
- Test the solution at a small scale
- Decision-maker
- The manager
- Unknown
- Budget
- Next step
- Propose a pilot on Thursday
Sources: 2 notes, 1 previous proposal
What it is, if you already know agents
A maintained, interlinked knowledge vault: facts, decisions, experience and sources. Not tasks, not chat history.
Chat memory can persist between conversations, but in a second brain knowledge is deliberately checked, linked and updated. A vector index lets you search, but it maintains nothing by itself. The difference is maintenance: someone regularly organises, links and updates the knowledge so that the agent works with current, verified context.
The model provides reasoning. Tools provide action. A second brain provides context that accumulates.
Five stages from source to result
Value comes from a repeated cycle. At each stage it is clear who does the work.
- 1
Sources
Conversation notes, transcripts, emails and documents, with date and origin.
Person or connection - 2
Maintained knowledge
Interlinked notes on clients, projects and decisions. Facts, assumptions and decisions kept apart.
Person, agent helps - 3
Context selection and reasoning
The agent finds the notes the task needs, rather than reading everything, and reasons from them.
Agent - 4
Concrete output
Meeting prep brief, reply draft, handover summary, suggested tasks.
Agent - 5
Review and feedback
A person reviews, corrects and approves. The correction flows back into the knowledge.
Person
The capture and organisation principles are based on Tiago Forte's Building a Second Brain method and the PARA structure. The agent use shown here is illustrative.
What a second brain is not
These things live alongside it and are connected, but they are not the same.
CRM
Deal stage, contacts, forecast. The CRM is the primary source for deal status, not for conversation context.
Calendar
When meetings happen. Can be a signal for the agent, but is not part of the knowledge.
Agent instructions
How the agent should work: tone, rules, permitted actions. A separate instructions folder, kept apart from knowledge notes.
Tool execution
Sending an email, writing to the CRM, creating a task. Happens through connections and their permissions, not through notes.
How agents use a second brain
Four ways of working that recur in almost every practical setup.
Reads it as context
Before a task, the agent opens the client's or project's hub note and follows links to decisions and sources. This gives more targeted context than searching the whole vault.
Writes back
Summaries and drafts go to the inbox folder, marked “needs review”. Configure the workflow and tool permissions so that the agent can change existing notes only after human approval.
Follows instructions from files
Rules on tone, the requirement to cite sources and permitted actions live in text files in a separate instructions folder. They are kept under version control and reviewed just like code.
Runs when triggered
In the morning, before a calendar event or after a transcript arrives. This needs a scheduler or a connection to an event source.
What the agent does not do on its own
It does not know what happened in email or the CRM unless those tools are connected. It does not update notes unless the workflow provides for it. Sending to a client is limited by tool permissions, not by the instruction text.
In practice: maintaining it in Obsidian
Obsidian is convenient because the vault is an ordinary folder of Markdown files.
The agent reads and writes in the same folder that a person sees in Obsidian. No separate database is needed. Links [[like this]] and properties (YAML frontmatter) at the top of a note provide structure that both people and agents understand.
A structure that is enough to start
Vault/
├── 00 Inbox/ ← new items, drafts
├── 10 Projects/ ← outcome with a deadline
├── 20 Areas/ ← sales, clients
├── 30 Resources/ ← arguments, templates
├── 40 Archive/ ← finished work
├── Daily notes/ ← quick capture
└── Templates/ ← client, conversationA client note the agent understands
---
type: client
status: active
updated: 2026-10-15
---
# Northern Build (fictional company)
## Context
Wants to reduce manual data re-entry.
The manager makes the decision.
## Decisions
- 2026-10-02 Start with one branch
→ [[2026-10-02 conversation]]
## Open questions
- Budget not yet clarifiedHabits that keep the vault alive
One hub note. Each client and project has one note that links to everything else.
Links, not copies. A fact lives in one place. Elsewhere it is referenced with a link.
Decisions with date and source. Without them, a month later it is unclear whether a decision still stands.
Weekly tidy-up. Empty the inbox folder, update statuses, archive what is finished.
Version control, such as Git, gives you history and lets you undo agent changes; it is set up separately from Obsidian. The mobile app is useful for notes after a meeting, if the vault is synced.
Connections to tools
Three types of connection, and a clear line on what is the primary source for each type of data.
File level
The tool exports to a folder.
Transcripts, documents and notes from your phone land in the inbox folder. A simple start.
Agent tools
The agent reads live data.
Through MCP or an API, the agent reads the CRM stage or the calendar and combines it with context from the vault.
Automation
An event starts the work.
A scheduler or webhook starts an agent task: a meeting tomorrow, a new transcript, a changed deal stage.
| Tool | What goes into the vault | What goes into the tool | What is needed |
|---|---|---|---|
| Meeting transcripts | Summary, decisions and open questions, with a link to the full text | Nothing | Export keeps the source; the summary needs a workflow |
| Selected facts and commitments, not the whole thread | Reply draft that a person sends | Read permission, with saving drafts allowed separately; a person sends. Without a connection the agent cannot see email | |
| CRM | Link to the deal, not a copy | Stage change or note | MCP or API connection for reading; writing only after approval |
| Calendar | Nothing | Nothing | Connection; a meeting becomes the signal for a brief |
| Task tool | Link to the task next to the decision | Next steps suggested by the agent | Connection; the agent creates a draft, a person approves |
Rule against conflicting copies: the vault is the primary source for context and decisions, the CRM for deal status, the task tool for actions. Everywhere else, keep only links.
Three situations.
One vault, one agent.
Fictional examples. Each shows what the agent reads, what it does and what a person reviews.
Preparation triggered by an event
“What have we already agreed with this client?”
The second conversation is tomorrow. The previous one was three weeks ago.
Agent reads
The client's hub note, two conversation notes and a description of a similar pilot from resources. From the CRM, it reads that the deal is at the proposal stage.
Agent does
Prepares a one-page brief: the need, commitments made, open questions. It notes that the budget is unknown rather than inventing it.
Person reviews
Reads it in two minutes, adds one detail from memory and goes into the call with a plan: budget, pilot, success criteria.
Task for the agent
“Prepare me for tomorrow's conversation with this client. Use only the vault and the CRM. Cite a source for every claim and list separately what is unknown.”
Connections needed: calendar as a signal, CRM for reading. Value: fewer repeated questions and a more focused conversation. A person still closes the deal.
Reusing evidence
“Do we already have a good example for this objection?”
The client is comparing price but has not yet compared the scope of work.
Agent reads
The scope of work in the proposal, the approved arguments in the resources folder and one client case study that the client has allowed us to use.
Agent does
Writes a reply draft about the differences in scope, plus one clarifying question. Separates what can be proven from assumptions: the competitor's full scope is unknown.
Person reviews
Decides whether to send the question or suggest a conversation. A discount is not the first answer.
Task for the agent
“Find our approved arguments for this proposal and prepare a short reply. Use only case studies marked as cleared for quoting. Do not invent client results.”
Connections needed: none, the vault is enough. Value: consistent arguments and less need to write from scratch.
Preserving decision context
“Why did we choose this option back then?”
A colleague is taking over a client project. They need the context behind decisions, not just the files.
Agent reads
The project note, the dated decisions with links to conversations, and the open items in the task tool.
Agent does
Prepares a handover brief: goal, current decisions and the reasons behind them, open questions, owners. Each decision links to its source.
Person reviews
Hands the colleague one page instead of twenty messages. The colleague can check any decision through its link.
Task for the agent
“Prepare a project handover brief for a colleague: goal, current decisions with reasons, open questions, owners and next steps. Include only notes cleared for sharing.”
Connections needed: task tool for reading. Value: less dependence on one person's memory. Before the handover, a person checks that the brief contains only information the colleague is allowed to see.
Where the value comes from
Each subsequent agent action uses the context already built up.
- Less repeated work. Preparation and drafts do not start from zero.
- Consistency. Everyone uses the same approved arguments and facts.
- Less dependence on one person. Knowledge is not held in just one head.
- Verifiability. The agent's output has sources, so trust in it can grow step by step.
The calculator is a simplified estimate of the value of work time saved, not a profit forecast. The starting figures are invented.
How to build a system from it
Six layers. You can change them one at a time.
- 1Obsidian + Git
Vault
An Obsidian folder of Markdown files, with version control set up separately. The primary source for context and decisions.
- 2Instructions folder
Instructions
Text files on tone, the requirement to cite sources, permitted folders and forbidden actions. In a separate folder, not in knowledge notes.
- 3Agent CLI or API
Agent
A model with access to files and tools. A command-line agent is enough to start; an API is useful both for internal automation and for a product.
- 4MCP servers, API
Connections
Calendar, CRM, email and transcripts through MCP or an API. Add only those the routine actually needs.
- 5Scheduler, webhook
Workflows
When the agent works: in the morning, before a meeting or after a transcript arrives. A scheduler or an event starts the task.
- 6Inbox folder, chat
Approval
The place where a person reviews and approves: the inbox folder, a chat message or a simple interface. Without this layer the system cannot be trusted.
Start with 1, 2, 3 and 6. Add connections and workflows once the routine already works with a manually started agent.
How to make money with it
One system, four models. The examples are fictional; the pricing models are general.
The client pays for a concrete, verifiable result, not for the number of notes saved. The price can be justified by time savings verified in a pilot or by faster responses. With your first client, start with a one-week pilot, not an annual contract.
Context for the sales team
What you build: a vault with client notes. The agent prepares a brief before a meeting and afterwards proposes an update in the inbox folder. Also works for proposal drafts.
Benefit: less time spent on preparation and fewer repeated questions. Verify this in a pilot rather than assume it.
Risk: nobody maintains the notes. Without a weekly tidy-up the brief goes stale.
Setup with a maintenance subscription
What you build: a vault, instructions and one routine for the client. Then you maintain it: review the inbox folder, adjust the instructions, add connections.
How you earn: a fixed setup price and a monthly maintenance fee. The client knows the content, so they remain responsible for some of the organisation.
Risk: promising “AI that knows everything”. Sell one routine with a clear result.
A product for one industry
What you build: ready-made templates, folder structure, instructions and routines for one industry, such as brokers or accountants. The client adds their own data. A variant: customer support that answers with a link to the source.
How you earn: a monthly fee per user or per company. The value lies in industry knowledge of what to capture and how to use it.
Risk: client data. If you process personal data on a client's behalf, you need a data processing agreement under the GDPR. Also find out where the model processes the data.
Knowledge as a sellable product
What you build: from notes on your experience, the agent helps create guides, checklists, templates and training materials.
How you earn: paid materials, courses, template packs or consulting with ready-made content.
Risk: publishing unverified or confidential content. Publish only verified content that is cleared for publication, whatever folder it sits in.
“You already have valuable knowledge, but it is scattered across conversations, emails and files. We organise it into one vault that the agent uses before a client conversation or a proposal. We start with one process and check whether it really saves time.”
A 30-second explanation for someone who does not know agents.
A pilot in one working week
One vault, one agent, one routine. No full migration.
Pick one routine
For example, preparing for a follow-up conversation with a client. Write down how much time it takes now.
Create a minimal vault
An Obsidian folder with inbox, projects and resources folders. A few conversation notes you are allowed to use, with date and source. One hub note for the client.
Give the agent access and rules
The agent reads the vault and writes only to the inbox folder. An instructions file requires sources and forbids inventing things. Add connections only if the routine needs them.
Use, measure, decide
Prepare a brief before each conversation. Afterwards the agent proposes an update and you approve it. At the end of the week, compare the time spent and the number of errors. Continue only if the benefit is visible.
Four rules for a trustworthy system
Source before confidence. Every claim has a link. Without evidence, the agent writes “unknown”.
A person approves what matters. Proposals, prices and messages to the client are reviewed before sending.
Only the access needed. Keep personal and company information apart. The agent sees only permitted folders; passwords are not stored in the vault.
Permissions, not promises. What the agent may write and send is set by tool permissions and the workflow, not just by the instruction text.
Read further
Verified sources, with one sentence each on why they are useful.
- Building a Second Brain: overviewThe basis of Tiago Forte's method: what to capture, how to organise it and how to use it.
- The PARA methodA folder structure based on what the information is useful for, not on file type.
- Obsidian: links between notesHow to link notes and parts of notes; the hub note relies on these links.
- Obsidian: propertiesFields at the top of a note that let the agent select notes by type and status.
- Obsidian SyncObsidian's service for syncing a vault across devices, including your phone.
- Model Context ProtocolThe standard through which an agent connects to a CRM, a calendar and other tools.
- Building effective agentsWhen a simple workflow is enough and when you really need an agent.
- Effective context engineering for AI agentsWhy an agent's output depends on the context it is given, not only on the model.
Start with one routine and one vault.