What you get: the exact 3 Claude Code prompts I used to build my Agentic OS — the system that runs my entire business from one dashboard. Paste them in, one at a time, and you'll have your own working version this weekend. No terminal expertise, no code.
A personal AI is 3 layers, not 1 chatbot
Most people quit because they treat AI like a smarter chatbot. The leverage is in stacking three layers:
- Skills — give your AI hands. It learns to actually DO your repeat tasks.
- Memory (a second brain) — give it memory. Every note, client, and number in one place, so it compounds instead of starting from scratch every time.
- Dashboard — give it a front door. Buttons, not commands — so anyone (even someone who's never written a line of code) can run it.
Below is one paste-and-go prompt for each layer. Do them in order.
Before you start (one-time, ~5 min)
Install Claude Code (Anthropic's coding agent). This is what runs all three prompts.
Open it in a folder you work in — or let it scaffold one for you (the prompts handle that).
Run each prompt in its own fresh Claude Code conversation, in order: Skills → Memory → Dashboard.
Tip: answer each prompt's questions like you're talking to a smart assistant — plain English, no jargon. It does the technical part.
🛠️ Prompt 1 — SKILLS (give your AI hands)
What it does: walks you through a conversational interview, one area of your work at a time, and turns each repeat task into a one-click skill you can fire from any Claude Code conversation. End of session: a folder of working skills.
How to use it: first install Anthropic's skill-creator (one time) by running /plugin install skill-creator@anthropics in Claude Code. Then paste the prompt below and just talk about your work when it asks.
You are helping me design and build my Claude Code skill architecture.
This is Part 1 of a 3-part agentic OS build:
Part 1 — Skills (this conversation)
Part 2 — Memory / second brain (separate prompt)
Part 3 — Dashboard (separate prompt)
Your job ends at skills. Don't try to design memory or dashboard architecture
here — those have their own prompts.
We work in CYCLES, not phases. One workflow area at a time. Each cycle has
the same shape:
1. I describe an area of my daily work (freeform)
2. You reflect back the tasks I mentioned + propose which become skills
3. I confirm / adjust / rename
4. You tag each skill (on-demand / local routine / cloud routine)
5. You offer to build them via /skill-creator
6. You loop with skill-creator per skill (or save specs for later)
7. You ask if I want another cycle or wrap
RULES OF ENGAGEMENT
- ONE question at a time. Wait for my answer.
- Be a thought partner, not a yes-man. If a "skill" I describe is too vague,
too broad, or one-off (not recurring), push back. Ask for the specific
input + specific output + how often it actually runs.
- Plain language. No jargon unless I use it first.
- Remember across cycles. If I describe something in cycle 3 that overlaps
cycle 1, flag the overlap.
- You do NOT hand-write skill files. When it's time to build, you hand off
to /skill-creator (Anthropic's skill that scaffolds skills properly with
trigger phrases, YAML frontmatter, eval cases).
- Don't auto-advance. Every step waits for me.
============================================================
CYCLE START
============================================================
Begin every cycle by asking:
"Tell me about ONE area of your daily work. Doesn't have to be organized
— morning routine, sales, content, finance, admin, research, whatever's
top-of-mind right now. Just talk for as long as you want, however messy.
I'll listen and pull out the patterns."
Wait for my dump. Don't interrupt. When I clearly signal I'm done (or after
an obvious pause), proceed to REFLECT-BACK.
============================================================
REFLECT-BACK
============================================================
After my dump, produce TWO lists:
TASKS I HEARD:
- <every recurring or notable task, even if not skill-worthy>
COULD BE SKILLS (repeatable, clear input + output):
- <verb-phrase name>: takes <input>, produces <output>,
triggered when <natural-language phrase>
For things I mentioned that DON'T make good skills (too vague / one-off /
not actually recurring), list them under TASKS but exclude from COULD BE
SKILLS. Note one-line reason for exclusion.
Ask:
"Did I get the right candidates? Anything to add, drop, rename, merge,
or split? Anything I excluded that you actually want as a skill?"
Iterate until I confirm the list.
============================================================
TRIAGE
============================================================
For each confirmed skill, ask:
"<skill-name>: on-demand (I fire when needed), local routine (runs on a
local schedule while my machine is on), or cloud routine (runs while my
machine is closed)?"
Default to on-demand if I'm unsure. Don't push for routines unless the use
case obviously needs it (e.g. "5am market briefing while I sleep" -> cloud
routine; "draft sponsor reply" -> on-demand).
Show the final tagged list before moving on:
<skill-name> — on-demand
<skill-name> — local routine, daily 8am
<skill-name> — local routine, weekly Monday
============================================================
BUILD GATE
============================================================
Ask:
"Build these N skills now via skill-creator?
(A) Build all now — I'll hand off to skill-creator for each, you
iterate with it per skill, we return here when each is done
(B) Build some, save the rest for later
(C) Save them all as specs, build later"
For each skill to BUILD NOW, output a hand-off message in this exact shape:
----------------------------------------------------
NEXT SKILL: <skill-name>
----------------------------------------------------
Paste this into your next message to hand off to skill-creator:
/skill-creator
Build a skill called <skill-name>. It should: <one-paragraph spec>.
Trigger phrases: <comma-separated list>. Input: <input shape>.
Output: <output shape>. Schedule: <triage tag>.
skill-creator will interview you about edge cases, eval examples, and
sample inputs. Iterate with it until the skill matches what you'd
produce manually. When the skill is built, come back here and say
"built" — I'll set up the next one.
----------------------------------------------------
Wait for me to say "built" (or "next" / "skip this one" / "spec is wrong")
before moving on. If I say the spec is wrong, revise it and re-emit the
hand-off message before moving to the next skill.
For skills SAVED FOR LATER (mode B or C), produce a "save block" I can
copy to a scratch note — each entry has the same /skill-creator hand-off
message ready to paste later.
============================================================
LOOP
============================================================
After all build-now skills are done (or we're in C mode), ask:
"Built <N> skills this cycle. Saved <M> for later. Want to walk through
another area of your work? Or wrap here?"
If another:
- Show me the running list of what's been built so far, grouped by cycle
- Goto CYCLE START
If wrap:
- Produce FINAL SUMMARY:
SKILLS BUILT THIS SESSION:
Cycle 1 (<area>): <name>, <name>, <name>
Cycle 2 (<area>): <name>, <name>
...
SAVED FOR LATER:
<skill-name>: <ready-to-paste /skill-creator hand-off>
...
- Then point me to the next two layers:
"Skill architecture done. You can fire any of these from any Claude Code
conversation by typing the trigger phrase. Next: Prompt 2 (Memory) gives
these skills a shared brain; Prompt 3 (Dashboard) turns them into buttons."
============================================================
GO
============================================================
Begin Cycle 1 now. Ask me for the first dump.🧠 Prompt 2 — MEMORY (give it a second brain)
What it does: builds an Obsidian vault that becomes your AI's long-term memory — a clean folder structure + a locked daily-note format so Claude can read everything you know and write new notes back into it. That's what makes it compound instead of starting from zero each time.
How to use it: paste into Claude Code, in (or next to) the folder where you want your second brain to live.
You are helping me build the MEMORY layer of my personal Agentic OS — Part 2 of 3.
Part 1 — Skills (built in a separate prompt)
Part 2 — Memory / second brain (this conversation)
Part 3 — Dashboard (separate prompt)
Your job ends at the memory layer. The goal: ONE Obsidian vault that becomes my
AI's long-term memory — every note, client, and number in one place, that Claude
Code can READ and WRITE — so it stops starting from scratch and starts to compound.
Work in PHASES. Pause for my confirmation between each. Don't skip the [VERIFY] checks.
RULES OF ENGAGEMENT
- ONE question at a time. Wait for my answer.
- Plain language, no jargon unless I use it first.
- Be a thought partner: if a folder or field I suggest is vague or redundant, push back.
- Show me the structure and get a yes before you create files. Don't dump silently.
============================================================
PHASE 0 — INTERVIEW
============================================================
Ask me, ONE at a time, then summarize and confirm before building:
1. What kind of work should this memory serve? (content creator / founder / clinician / researcher / personal productivity)
2. What are the 3-5 things I'd want my AI to always have on hand? (clients, notes, numbers/metrics, projects, a daily log...)
3. Where should the vault live? (give a path, or say "scaffold one" and create ~/second-brain)
4. Do I already use Obsidian? (yes -> point you at my vault / no -> set it up from scratch)
============================================================
PHASE 1 — VAULT SKELETON (the shape of the memory)
============================================================
Create a four-zone vault (adapt the names to my domain, keep the four zones):
- inbox/ — staging: every new note or idea lands here first
- projects/ — active work (each file has a `status:` in its frontmatter)
- output/ — finished, shipped things
- wiki/ — evergreen, distilled knowledge the AI maintains
...plus: daily-notes/ (daily rhythm), system/ (machine-readable plumbing), _archive/.
Write a CLAUDE.md at the vault root documenting the conventions: what each folder is for,
file naming (YYYY-MM-DD-slug.md), and the status taxonomy (active / in-progress / blocked / done / archived).
[VERIFY] I can open the vault in Obsidian and see the folders + CLAUDE.md.
============================================================
PHASE 2 — THE DAILY-NOTE SCHEMA (the contract everything agrees on)
============================================================
Lock a FROZEN daily-note format so every skill and every read uses the same shape.
- Create system/schemas/daily-note.md defining the exact frontmatter + section order:
Current Focus / Top 3 Priorities / Schedule / Daily Drivers / Activity Log / Notes / EOD Reflection
- Create a template at system/templates/daily.md that stamps a fresh note from that schema.
[VERIFY] A new daily note renders cleanly and its frontmatter parses.
============================================================
PHASE 3 — WIRE THE MEMORY (read + write + connect)
============================================================
- Add an "Indexing rule" to CLAUDE.md: every folder with 3+ files gets an index note,
every file gets tags, and notes cross-link across folders — so the AI can actually
traverse what I know, not just store it.
- Prove READ: open the vault in a Claude Code session and ask it to summarize what's there.
- Prove WRITE: ask it to file a test note into inbox/, tag it, and link it to a project.
[VERIFY] Claude reads a file I point it at AND creates a correctly-placed, tagged, linked note.
============================================================
GO
============================================================
Start Phase 0 now. Ask me the first question.📊 Prompt 3 — DASHBOARD (give it a front door)
What it does: clones a ready-made cockpit and wires it to YOUR skills + data, so every skill becomes a button. Plan my day, prep a call, pull metrics — one click each. This is the part that makes the whole thing usable by anyone.
How to use it: paste into Claude Code. It checks your setup, clones the dashboard, boots it at localhost:8501, then interviews you to wire in your own skills, data, and brand.
Set me up with the Agentic OS Dashboard — a local Streamlit cockpit for my
Claude Code skills with three tabs (Overview / Audience / Research), a token
meter, audience cards, and parallel skill runs through a background queue.
This is a CLONE-AND-WIRE flow. The dashboard code is already production-ready
and lives at github.com/cth9191/agentic-os-dashboard. Your job is to clone it,
boot it in DEMO mode so I can see what we're working with, then walk me through
wiring it to MY environment.
Two phases:
PHASE 1 — Clone + boot the shell (silent, just ship it running)
PHASE 2 — Interview me to customize it
============================================================
PHASE 1 — CLONE + BOOT
============================================================
ORDER OF OPERATIONS (run silently, no questions until phase 2):
1. Check prerequisites:
- python --version (need 3.10+)
- git --version
- claude --version
If any fail, stop and ask me to install before continuing.
2. Clone the repo into the current working directory:
git clone https://github.com/cth9191/agentic-os-dashboard.git
3. cd agentic-os-dashboard
4. pip install -r requirements.txt
If permission errors on Mac/Linux, retry with --user.
5. Copy config:
- Windows: copy config.example.py config.py
- Mac/Linux: cp config.example.py config.py
6. Edit config.py — start in DEMO_MODE so I see the dashboard populated
immediately without needing real data:
- Set DEMO_MODE = True
- Leave VAULT_PATH / VAULT_NAME / CLAUDE_CLI as placeholders for now
- Keep all ENABLED_CARDS = True so I see every component
7. Boot:
streamlit run app.py --server.headless true --server.port 8501 &
(Background it on Mac/Linux. On Windows, spawn it in a new terminal window.)
8. Tell me, exactly:
"COCKPIT UP. Open http://localhost:8501 in your browser. Everything you see
is demo data for now — we'll wire it to your real skills + data in Phase 2.
Open it in another window so you can watch it change live as we work."
============================================================
PHASE 2 — INTERVIEW + WIRE
============================================================
Do NOT stop after Phase 1. Ask me these in order. Wait for each answer before
moving on. Edit config.py in place between questions so I see the dashboard
update live.
QUESTION 1 — WORKING FOLDER
"Where's the folder Claude Code should run skills in + where the dashboard
should read data from? Could be your second-brain vault, your main project
folder, anywhere you already work with Claude Code. If you don't have one, say
'scaffold one' and I'll create a minimal data folder for you."
-> Set VAULT_PATH + VAULT_NAME in config.py.
QUESTION 2 — CLAUDE CLI
"Run `which claude` (Mac/Linux) or `where claude` (Windows) and paste the full
path so the dashboard can launch skills."
-> Set CLAUDE_CLI to the absolute path.
QUESTION 3 — CLAUDE PLAN
"Which plan are you on? pro / max / team / enterprise?" (sets the daily run cap).
-> Set CLAUDE_PLAN.
QUESTION 4 — SKILLS (the leverage moment)
"Paste the list of skills you built in Prompt 1 (names + one-line descriptions).
I'll wire each one to a button, drop the unrelated placeholders, and tell me any
that need a text input (like a topic or URL) when clicked."
-> Rewrite the SKILLS list in config.py to match mine.
QUESTION 5 — AUDIENCE METRICS (optional)
"Do you track any audience numbers (YouTube / Instagram / TikTok)? If yes, I'll
turn on the audience cards; if no, I'll hide them so nothing sits empty."
-> Toggle ENABLED_CARDS accordingly.
QUESTION 6 — BRAND / COLOR
"Keep the default accent color or give me a hex / a vibe (deep cobalt, forest
green, electric purple) and I'll re-skin it."
-> Swap the accent in the CSS if I want.
QUESTION 7 — FLIP TO LIVE DATA
"Ready to leave demo mode? If yes, I'll seed it with your real data and set
DEMO_MODE = False so the cards read from your folder. If not yet, we keep demo
mode for now."
-> If yes: seed data, set DEMO_MODE = False, tell me to refresh.
QUESTION 8 — AUTOSTART (optional)
"Want the dashboard to auto-launch on boot so it's always running at
localhost:8501? yes / no / tell-me-how."
-> If yes, set up the auto-launch and walk me through it.
============================================================
WRAP
============================================================
When Phase 2 is done, tell me:
"YOU'RE WIRED IN. Your cockpit lives at http://localhost:8501 — bookmark it.
Your skills are clickable, your folder feeds the panels, and you can change
anything later by opening this same folder in Claude Code and describing what
you want. To add a skill: build it (Prompt 1), then tell me to add a button."
Never commit config.py, never push to the upstream repo, never ship my paths or
keys to GitHub. All edits stay in my local clone.The 3-second version
- Skills prompt → teach your AI to do your repeat tasks.
- Memory prompt → give it one brain to read + write.
- Dashboard prompt → turn it all into buttons.
Do them in order. You'll have a working Agentic OS by Sunday.
Want this built around YOUR business instead of building it yourself? I set these systems up end-to-end for professionals and clinics — skills, a second brain, and a dashboard your whole team can run. If that's you, DM me BUILD and I'll show you how it works.