Organize knowledge
Build a usable note vault and trusted wiki.
LIVE BUILD WORKSHOPSecond Brain
Second Brain
+ Agent Team
Obsidian keeps your knowledge. The wiki keeps it useful. Codex helps turn it into action.
Your knowledge firstAgents with clear rolesClean team handoffs
YOU
notesCodexagentsFinish line: describe one helpful agent and pass its work to the next team member.
Leave with a system you can run
Every idea becomes a structure, template, example, or team activity.
Describe an agent
Give it a job, values, communication style, and limits.
Set permissions
State what it may view, change, share, or approve.
Pass the work
Give completed work to the next agent without repeating it.
Start with one useful pattern. Add more tools only when needed.
A pile of notes is storage, not memory
A useful system helps the right knowledge return when you need it.
Everything goes in
Nothing gets cleaned up.
The reason is missing
Old fragments no longer make sense.
Search is difficult
You cannot remember the exact words.
Nothing comes out
Notes never become a decision or useful result.
Capture → organize → keep the best → find → use → check.
Three layers, one source of truth
Each layer has a distinct job.
Obsidian vault
Notes, projects, decisions, people, research, specs, logs.
Karpathy-style LLM wiki
Maintained, linkable pages compiled from vault evidence.
Codex
Reads allowed paths, writes artifacts, runs agent loops.
Obsidian remembers. The wiki clarifies. Codex executes. Notion publishes selected outputs.
AI is one part of a larger workspace
Think of the system like a person working at a desk. The technical names are optional.
The model
Recognizes patterns and creates responses.
Current information
What the AI can see right now.
Vault + wiki
Knowledge saved for later.
Codex + tools
Ways to read, write, and act.
Coordinator
Assigns work and knows when to stop.
The AI brain needs information, tools, and clear instructions to be useful.
Folders that reveal work state
A boring map helps humans and agents predict where truth belongs.
vault/
├── incoming/ # fast, unprocessed capture
├── projects/ # outcomes with an end state
├── areas/ # ongoing responsibilities
├── resources/ # research and reference
├── wiki/ # durable promoted knowledge
├── agents/ # specs and access policies
└── logs/ # runs, decisions, evaluationsRules
- Everything enters through
incoming/. - Projects and areas hold active context.
- Only reviewed knowledge enters
wiki/. - Agent work stays inspectable in
logs/.
Folder names are an access-control surface, not decoration.
Move notes through visible states
Low-friction capture needs a processing path.
Capture
Write in incoming/.
Atomize
One clear idea or decision.
Place
Project, area, or resource.
Promote
Compile durable wiki knowledge.
Drill: restructure 10–15 notes and create a one-page stack gap map.
Turn useful notes into trusted reference pages
The wiki is the small, tidy shelf of knowledge you expect to reuse.
Move into the wiki
- An idea used in several projects
- A decision and the reason behind it
- Your current view on an important topic
- An open question someone must answer
Leave as a regular note
- A temporary or unverified thought
- A quick record of one moment
- A copy of an existing wiki page
- Something nobody can understand yet
An agent may suggest an update. A person approves important changes.
Five questions for durable knowledge
Expose evidence, stance, and uncertainty.
problem: What recurring question does this resolve?
decision: What was decided, if anything?
sources: [[vault links]] · dates · evidence quality
current_stance: What do we believe/do now—and why?
open_questions:
- question · owner · review dateA wiki page is a maintained position with provenance—not polished prose alone.
Give the AI the right notes before it answers
This is often called RAG. Remember the habit, not the technical label.
Ask
Write a clear question.
Find
Search notes by meaning.
Choose
Keep the useful notes.
Answer
Use them as evidence.
Check
Open the source and verify.
Find evidence, answer, then check.
Five technical words—made simple
Use the plain meaning first. The label only helps you recognize it later.
- LLM
- The AI model that reads and writes language.
- Context
- The information the AI can see right now.
- Embedding
- A way for a computer to compare meaning.
- Semantic search
- Finding related ideas even when the words differ.
- RAG
- Finding useful notes before the AI answers.
- Orchestration
- Deciding who does what, and in what order.
You do not need the mathematics to use these workshop patterns.
Treat an agent like a new team member
Before it starts, clearly state what it may see and change.
May read
Named research, projects, and approved wiki pages.
May suggest
Prepare a change for a person to review.
May write
Only its output folder and activity log.
Protected work
Private notes, final decisions, deletion, publishing.
Let it read what it needs and change only what it owns.
Give it a job, values, and communication style
These three choices make behavior easier to predict.
What is it responsible for?
Role, boundaries, tools, and definition of success.
What does it protect?
Quality standards, refusal lines, and what “good” means for your team.
How should it communicate?
Tone, length, structure, and expected format.
Think: job description + team values + communication guide.
Copy this before writing a prompt
Bind identity, access, contracts, and stopping behavior.
name:
persona: role · scope · success
soul: values · refusal lines · taste
voice: tone · brevity · formats
allowed_vault_paths:
write_permissions:
tools: Codex · search · browser · Notion export
input_contract:
output_contract:
stop_conditions:
human_checkpoints:Quality check
- Can someone predict its behavior?
- Can it cite what it read?
- Is every write location named?
- Can it stop without guessing?
- Does a human own consequences?
Store specs in
agents/; version them like policy.Research Agent · “Scout”
Source-first investigator; never fills evidence gaps with confidence.
Find, evaluate, organize evidence. Success = traceable packet answering the brief.
Values primary evidence, recency, counterexamples, uncertainty. Refuses invented citations.
Neutral and compact: claim → evidence → confidence → unknowns.
Read selected projects/, resources/, wiki/. Write research folder + append logs.
Codex, search, browser. Input: question, decision, scope, date. Output: cited packet, contradictions, risks.
Stop when claims have evidence or gaps are explicit. Human approves scope expansion and publishing.
Chief of Staff · “Mira”
Turns approved evidence into priorities without redoing research.
Produce dashboard, ranked plan, owners, decisions. Success = executable one-page brief.
Values clarity, leverage, ownership, realistic sequence. Refuses false urgency and hidden tradeoffs.
Calm and decisive: decision → rationale → actions → risks → asks.
Read frozen research + wiki. Write briefs. Export only approved summary to Notion.
Codex, vault I/O, Notion export. Output: dashboard, ranked plan, risks/unknowns.
Forbidden to redo research. Human approves commitments, owners, due dates, export.
Discover → Plan → Execute → Verify → Iterate
Single-agent work is the practice ground for a team loop.
Discover
Read task, policy, context.
Plan
Steps, artifact, risks, stop.
Execute
Use allowed tools.
Verify
Test definition of done.
Iterate
Fix bounded gaps or escalate.
Stop conditions and checkpoints are part of the loop.
Tools must return inspectable results
Structured operations replace pretend actions.
Execute
Read files, create artifacts, run checks.
Discover
Find external evidence.
Inspect
Interact with live interfaces.
Export
Publish approved outputs only.
Every call needs inputs, expected output, failure behavior, and permission.
Pass completed work, not a vague message
A handoff agreement tells the next agent what is finished and what happens next.
work_being_passed: research-packet.md
finished_when:
- required claims cited
- contradictions and unknowns explicit
- confidence stated
next_agent: chief-of-staff
do_not_repeat:
- repeat completed searches
- silently replace evidence
- expand scope without approval
next_agent_response: accept | return with missing item
owner: accountable personSimple test: can the next agent begin without asking “What am I supposed to do?”
One manager. Clear specialists. Visible work.
The team needs one place where responsibility and progress are obvious.
MANAGERbreaks down work · assigns · checks · stops
WORK BEING PASSEDnamed file + status + evidence + next owner
If one agent can finish the task well, do not add a team yet.
Research → Dashboard + Action Plan
Evidence becomes an executive artifact without being researched twice.
Research
Cited packet, contradictions, confidence, unknowns.
Synthesis
Rank choices against goal and constraints.
Dashboard
One page, actions, risks, unknowns, owners.
Research owns evidence. Synthesis owns prioritization. A human owns commitments.
One page that drives a decision
Separate facts, priorities, uncertainty, and action.
3 primary sources · 2 assumptions
- Validate highest-risk assumption
- Run bounded pilot
- Review and scale/stop
- Ownership after pilot
- Cost at scale
- Missing user evidence
Every action links to evidence, an owner, and a decision point.
Research → Notion summary + next actions
Use Notion for shared visibility, not competing truth.
Research artifact
Evidence, citations, confidence, unknowns.
Approved summary
Concise, audience-safe output.
Coordination state
Summary + owner + due date + vault link.
| Next action | Owner | Due | Vault link |
|---|---|---|---|
| Validate assumption | Project lead | 2026-09-03 | [[projects/pilot/research]] |
| Draft pilot plan | Ops lead | 2026-09-05 | [[projects/pilot/plan]] |
If the decision changes, update the vault first—then re-export.
Librarian → Executor → Reviewer
Maintenance, execution, and quality control stay separate.
Propose wiki diff
Find stale stances, duplicates, missing links.
Apply approved work
Edit allowed paths and log the run.
Verify
Check provenance, scope, links, unintended changes.
No role approves its own consequential change.
Measure the wiki, agent, and handoff
Make failures visible at the layer that caused them.
- Sources traceable
- Stance dated
- Duplicates/stale pages
- Useful retrieval
- Access compliance
- Task completion
- Unsupported claims
- Correct stop
- Acceptance rate
- Redo avoided
- Fields complete
- Owner clear
Log failure, likely layer, fix, and next test in
logs/.Recognize the failure before scaling it
Fix architecture and contracts before rewriting prompts.
Hallucinates.
Require retrieval + citations.Everything promoted.
Criteria, owner, review date.Generic slop.
Values, refusal, taste.No owner.
Manager, baton, turn budget.Truth diverges.
Vault-first, linked export.Each failure needs a detection signal and a guardrail.
Specify one agent and run one handoff
Use a real project and a small existing research note.
- 0–5Choose task, upstream artifact, destination.
- 5–12Fill identity, access, contracts, stops.
- 12–17Write handoff and forbidden redo.
- 17–23Run or role-play next agent.
- 23–25Log one failure and one spec change.
Done = another person can accept or return the artifact predictably.
Build Research → Synthesis → destination
Complete one thin pipeline with real vault evidence.
Prepare
Task, access, specs.
Research
Cited packet.
Handoff
Accept/return, synthesize.
Verify
Evidence, access, ownership.
Improve
Revise and rerun one step.
Stretch: add a reviewer. Do not add agents until the baton works.
Personal memory → accountable team action
Read the vault, retrieve the wiki, coordinate specialists, write back under policy.
Obsidian→Wiki retrieval→Research→Synthesis→Dashboard / Notion→Review + write-back
Demo: task → retrieval → handoff → output → verification → improvement.
The complete pattern
Use this sequence for any agent workflow.
Paths + promotion + retrieval
Persona + soul + voice
Read + propose + write + ask
Discover + plan + execute + verify
Artifact + done + forbidden redo
Wiki + agent + handoff metrics
Obsidian is truth. Wiki is maintained knowledge. Codex executes. Notion receives outputs.
Show the journey, not only the final answer
Each team presents the topic they selected and how their system supports the work.
What did you choose?
Who has the problem, and why does it matter?
What must be remembered?
Show the vault structure and one useful wiki page.
Who starts the work?
Explain its job, values, style, permissions, and result.
How is work divided?
Show the specialists, work passed, and human approval points.
What did the system produce?
Demonstrate the result or walk through a realistic example.
What would you change?
Name one success, one difficulty, and the next improvement.
Format: 5-minute team presentation + 2-minute questions and feedback.