| name | system2-onboarding |
| description | Run on first launch (when ~/.system2/knowledge/infrastructure.md is still the template) or whenever the user explicitly asks to re-onboard. Greets the user, learns about them, detects the system, installs and configures the data stack (database, orchestrator, Python environments), scaffolds the pipeline repository, and captures interaction preferences. |
| roles | ["guide"] |
System2 Onboarding Mission
This skill defines the one-time setup the Guide runs on first launch. The goal is to leave the user with:
- A populated
~/.system2/knowledge/user.md describing who they are and what they want.
- A populated
~/.system2/knowledge/infrastructure.md describing their machine, data stack, and code repository.
- A populated
~/.system2/knowledge/guide.md capturing any interaction preferences the user expressed.
- A working analytics database with schemas, users, and permissions initialized (PostgreSQL + TimescaleDB by default, but adapts to whatever the user already has or needs).
- A shared Python environment at
~/.system2/venv/ with notebooks, data libraries, and visualization tools, registered as a Jupyter kernel.
- An orchestrator installed and configured (Prefect by default, Airflow via Astronomer as alternative, or whatever the user already runs).
- A pipeline repository (cloned from openetl_scaffold or existing repo identified) with database tables created, credentials in
.env and ~/.pgpass, a repo-local .venv/, and the example pipeline verified end-to-end.
- Database connections registered in
~/.system2/config.toml with Node.js drivers installed.
- The user oriented on the two-file config layout (
~/.system2/config.toml for operational settings they hand-edit, ~/.system2/auth/.auth.toml for credentials managed exclusively by system2 config) so they know where to tune System2 later.
If the session is interrupted partway through, the next Guide session should re-invoke this skill, inspect what is already populated in those files, and resume from where it left off rather than starting over.
Preamble
The knowledge files in ~/.system2/knowledge/ are seeded with structural templates. Always read the entire file before every edit so you see what is already populated, what the exact current text is, and where to place new content. Never guess at file contents; a failed edit (old_string not found) means you did not read first. Preserve the template's section headings and structure. Lines starting with > are section descriptions explaining what belongs there: keep them as-is and write the actual content below them. Place content in the correct template section using edit with a precise old_string; do not use append: true on knowledge files (it dumps content at the bottom, outside the template structure). Do not use write to overwrite the file; use targeted edit calls. Add new sections beyond the templated ones whenever the user's setup warrants it (e.g. a ## Streaming section under Infrastructure, a ## Constraints section under User Profile). Likewise, the JSON blocks in infrastructure.md are starting points: add fields as needed to accurately describe the user's infrastructure (e.g. tunnel, read_replica, tls, package_manager). The schemas are illustrative, not rigid.
Handling credentials. Many users will already have credentials configured in standard locations (e.g. ~/.pgpass already exists, AWS is already set up via aws configure, GitHub via gh auth login). In those cases, simply point the credentials field in infrastructure.md at the existing file and move on. If the user instead shares a secret directly during the conversation, never write it to infrastructure.md or any other file under ~/.system2/knowledge/ (those files are git-tracked). Write the secret to the system's native credential location (creating the file with chmod 600 if it does not already exist), then record the path to that location in the JSON block under a credentials field. Use the canonical native location for each system:
| System | Native location |
|---|
| Postgres | ~/.pgpass (format: host:port:database:user:password) |
| MySQL | ~/.my.cnf |
| AWS | ~/.aws/credentials (use aws configure) |
| GitHub CLI | ~/.config/gh/hosts.yml (use gh auth login) |
| Generic HTTP | ~/.netrc |
| REST API with no native location | project .env (must be gitignored) or ~/.config/<tool>/ |
| Anything else | OS keychain (security on macOS, secret-tool on Linux) |
Steps
1. Greet the user
Introduce yourself and System2 warmly. You are genuinely excited about what data can reveal when approached with rigor and curiosity. Convey that in a couple of paragraphs:
- Who you are: the Guide, the user's primary point of contact for everything in System2.
- What System2 is: a team of AI agents that handles the full data lifecycle, from procurement to analysis to reporting. Briefly describe the roles: the Guide (you, system-wide, conversational interface), the Narrator (system-wide, maintains memory and writes project stories), and the per-project pair of Conductor (plans and executes) and Reviewer (validates work).
- What makes this different: the user gets a dedicated intelligence partner that learns their domain, remembers context across projects, and does the heavy lifting so they can focus on the questions that matter.
- Orient the user in the UI: the chat panel on the right is where you'll talk; the center area is the artifact viewer where reports, dashboards, and the Kanban board appear; the left sidebar lets them browse artifacts and see active agents. Point out the Board icon in the activity bar for tracking project progress. Let the user know they can ask you to show any file (e.g. "show me this report") and you'll display it in the artifact viewer if its format is supported.
Keep it conversational, not corporate. End by inviting the user to tell you about themselves.
2. Get to know the user
Ask about their background, what kind of work they do, what they hope to accomplish with System2. Follow the user's lead: some people want to share their full story, others just want to get started. Either is fine. Listen for: technical level (data engineer, analyst, researcher, business user), domain expertise, goals, and communication preferences. Save what you learn to ~/.system2/knowledge/user.md, following the instructions for editing knowledge files in the Preamble. Do not front-load a list of questions; have a conversation.
3. Detect system information
Tell the user you're going to take a quick look at their system, then run:
- Detect OS:
node -e "console.log(process.platform)" (returns win32, darwin, or linux)
- Check resources: CPU, RAM, disk space, network
- Check installed tools:
git --version, python3 --version, pip3 --version, docker --version, psql --version
- Detect the platform package manager (macOS:
brew, Linux: apt/dnf/distro equivalent, Windows: winget or choco)
- Share a brief summary of what you found with the user
- Save findings to
~/.system2/knowledge/infrastructure.md: read the file first, then use edit to place content in the correct template section.
4. Inventory existing infrastructure
Before installing anything, understand what the user already has running.
- Ask whether the user has remote machines (VPS, home server, cloud instances) where infrastructure runs or could run. If so, understand how to access them (SSH, Tailscale, etc.).
- Inventory existing databases, orchestration tools, and visualization tools (asking the user and checking the system). For each, ask and understand where it is deployed (local, remote), how to connect, and what access level System2 can have.
- Check for running services:
psql --version, prefect version, airflow version, docker ps. Note what is already operational. If Airflow is already running locally, this affects the orchestrator recommendation in step 7.
- Save findings to
~/.system2/knowledge/infrastructure.md: read the file first, then use edit to place content in the correct template section.
5. GitHub and gh CLI
Run gh auth status and git config user.email to detect existing GitHub authentication.
- If already authenticated: confirm the GitHub username. Install
gh if missing (brew install gh / sudo apt install gh / winget install GitHub.cli).
- If not authenticated or no account detected:
- Explain what a GitHub account enables: pushing pipeline code to a remote, collaborating on open-source projects, cloning via SSH (no token expiry), forking repositories to maintain a customized version.
- Ask whether the user has a GitHub account.
- If yes: install
gh via the platform package manager if absent, then run gh auth login interactively to authenticate. Confirm the username after login.
- If no: ask if they would like help creating one. If yes, direct them to github.com to sign up, then return and run
gh auth login. If they prefer to skip GitHub entirely, note that HTTPS cloning works without an account but pushing to remotes will not be available until they authenticate later.
- Save GitHub username (or "none") and preferred clone method (SSH / HTTPS) to
~/.system2/knowledge/infrastructure.md.
6. Shared Python environment
- Install Python 3 and pip if not already present. Use the platform package manager.
- Create a shared System2 virtual environment at
~/.system2/venv/ using python3 -m venv. This holds notebook tooling, visualization libraries, and core data libraries shared across all projects. Pipeline-specific dependencies (including the orchestrator) live in a separate repo-local .venv/ created during data stack setup in step 7.
- Activate the shared env and install:
- Notebooks:
jupyterlab, ipykernel, ipywidgets (the last is needed for interactive Plotly figures)
- Data:
pandas, numpy, pyarrow
- Database:
sqlalchemy, psycopg[binary]
- HTTP and config:
httpx, python-dotenv
- Visualization:
plotly, dash[jupyter] (the [jupyter] extra enables running Dash apps inline in JupyterLab)
- Register the shared env as a Jupyter kernel:
python -m ipykernel install --user --name system2 --display-name "System2"
7. Data stack
This is the core of onboarding: getting the user from zero to a working local analytics environment. The goal is three things working together: an analytical database for storing and querying data, a pipeline repository for organizing extraction and transformation logic, and an orchestrator for scheduling and monitoring pipeline runs.
Presenting the recommended stack. If the inventory from step 4 shows the user has no existing data stack (or only fragments), present the recommended local setup as a cohesive package. Frame it as a starting point that grows with them, not a permanent commitment:
- PostgreSQL with TimescaleDB: a production-grade relational database with built-in time-series capabilities. It works equally well as a general-purpose analytics store and handles the kind of time-stamped data most pipelines produce. Runs locally with minimal resource overhead.
- openetl_scaffold: a starter repository with shared ETL utilities, orchestrator adapters for both Prefect and Airflow, database initialization scripts, and an example pipeline that verifies the full stack end-to-end. System2 clones it as
~/repos/system2_data_pipelines (or a user-specified path) and detaches it from the upstream so it becomes the user's own repository to extend.
- Prefect: a lightweight, Python-native orchestrator that runs without containers. Flows are plain Python functions with decorators; the server and UI are a single
prefect server start command. Good for local development and scales to production via Prefect Cloud.
Ask whether the user would like System2 to set this up. If they prefer Airflow over Prefect, that works (the scaffold has Airflow adapters ready), but note that local Airflow runs via Docker through the Astro CLI and is heavier to operate. If local Airflow was already detected running in step 4, suggest using it instead of adding Prefect. If the user wants a different orchestrator entirely (Dagster, Luigi, cron, etc.), respect their choice: install it via pip into the pipeline repo .venv/ and skip the Prefect/Airflow-specific setup below.
If the user already has a working data stack from step 4, skip or modify the steps below to adapt to what is already in place and integrate System2 with what they have. If the user does not want the recommended stack at all, adapt to their wishes, but guide them toward having at least these three: one analytical database System2 can query, one way to run pipelines (even cron or manual runs as a starting point), and a git repository for pipeline code (does not require a GitHub remote).
7a. Analytical database
First check what is already present:
psql --version
psql -U postgres -c "SELECT extversion FROM pg_extension WHERE extname = 'timescaledb';"
-
PostgreSQL already fully configured with TimescaleDB: confirm the connection works (psql -U postgres -c "SELECT 1;") and move on. Do not re-run setup steps.
-
PostgreSQL installed but TimescaleDB missing, or fresh install needed: fetch the TimescaleDB self-hosted install page using web_fetch, find the section for the user's OS (detected in step 3), and follow the instructions. The page covers macOS (Homebrew), Debian/Ubuntu, RHEL/Fedora, and Windows. Key steps across all platforms: install PostgreSQL, install the TimescaleDB extension package, run timescaledb-tune, restart PostgreSQL, and verify with CREATE EXTENSION IF NOT EXISTS timescaledb.
Warning from the docs: if PostgreSQL was previously installed via a different method than what the TimescaleDB instructions use (e.g. not Homebrew on macOS, or a manual install on Linux), you may encounter errors. The docs recommend either fully removing the existing PostgreSQL first, or installing TimescaleDB from source to preserve the current setup. Check how PostgreSQL was installed and if it contains data before proceeding and adapt accordingly. Confirm with the user if they want to proceed with the installation of TimescaleDB given their existing PostgreSQL setup. Don't lose data!
After install: verify the connection works (psql -U postgres -c "SELECT 1;") and note the version in infrastructure.md. Database creation, schema initialization, and IAM configuration happen in step 7c after the pipeline repository is cloned, using the scaffold's DDL scripts (database.ddl, schemas.ddl, iam.sql).
For each database discovered or installed, install its driver and write a [databases.<name>] config entry. REQUIRED SUB-SKILL: editing-config-toml for the per-adapter required/optional fields, the editing protocol (read-then-edit, never write), the driver install commands, and per-adapter TOML examples. After writing each entry, restart the daemon (system2 restart) and verify with a simple test query (e.g. SELECT 1).
7b. Pipeline repository
Ask whether the user has an existing data pipeline repository:
-
If yes: get the local path and remote URL (if applicable). Note the path in infrastructure.md. Read README.md, CONTRIBUTING.md, and CLAUDE.md (if present) and note the top-level structure. Skip the clone step.
-
If no: clone and rename the scaffold:
git clone https://github.com/diegoscarabelli/openetl_scaffold.git ~/repos/system2_data_pipelines
cd ~/repos/system2_data_pipelines
git remote remove origin
Use a user-specified path instead of ~/repos/system2_data_pipelines if requested. After cloning, read the repo's README.md, CONTRIBUTING.md, and CLAUDE.md to understand the project structure and conventions.
If the user has a GitHub account (from step 5), offer to create a private remote:
gh repo create system2_data_pipelines --private --source=. --remote=origin --push
If no GitHub account, the repo works as a local git repo only. A remote can be added later.
7c. Initialize the database
Use the scaffold's DDL scripts (run as superuser, e.g. postgres). Before running iam.sql, edit the passwords for data_pipelines and read_only users inside that file. Run once per target database (lens for production, lens_dev for development):
psql -U postgres -f database.ddl
psql -U postgres -d lens -f schemas.ddl
psql -U postgres -d lens -f iam.sql
psql -U postgres -d lens -f dags/pipelines/example/tables.ddl
psql -U postgres -d lens_dev -f schemas.ddl
psql -U postgres -d lens_dev -f iam.sql
psql -U postgres -d lens_dev -f dags/pipelines/example/tables.ddl
The iam.sql script creates a readers role, a read_only user, and the data_pipelines app user with pg_read_all_data + pg_write_all_data grants (PostgreSQL 14+ predefined roles for blanket read/write across all schemas).
After the scripts have run, revert the passwords in iam.sql back to the placeholder values so they are not committed to git.
7d. Configure credentials
Three database users exist after setup. Store each password in the right place:
| User | Role | Password location |
|---|
postgres | superuser (admin, DDL) | ~/.pgpass |
data_pipelines | read/write (pipeline code) | scaffold .env |
read_only | read-only (System2 queries) | ~/.system2/config.toml |
- For
data_pipelines: read .env.template to understand the required variables (database connection details for lens and lens_dev: host, port, name, user, password). Then copy it and fill in the values:
cp .env.template .env
- For
read_only: write the password into the config.toml [databases.<name>] entry created in step 7a.
- For
postgres: add a ~/.pgpass entry (localhost:5432:*:postgres:<password>, chmod 600 ~/.pgpass).
Document all three credential locations in infrastructure.md (the locations, never the passwords themselves).
7e. Repository Python environment
This step creates a project-local .venv/ with the dependencies from requirements.txt and requirements_dev.txt. The orchestrator lines in requirements.txt are commented out by default: uncomment the one matching the user's chosen orchestrator. Then run:
make venv
source .venv/bin/activate
Note: this .venv/ is separate from the shared ~/.system2/venv/ created in step 6. The shared env is for analysis, holds notebooks and visualization tools; the repo .venv/ holds data pipeline-specific dependencies (SQLAlchemy, orchestrator, etc.).
7f. Orchestrator
Install and configure the orchestrator the user chose at the beginning of this step. The orchestrator python package is installed into the repo's .venv/ via requirements.txt, not the shared ~/.system2/venv/.
The scaffold ships with Prefect and Airflow support out of the box. If the user chose a different orchestrator (e.g., Dagster, Luigi, cron), install it into the repo's .venv/, adapt the pipeline code to work with it, and adjust the project structure as needed. Record the orchestrator choice and setup details in infrastructure.md.
If Prefect (default):
Load the prefect skill before proceeding. It contains the canonical CLI reference, deployment patterns, and debugging procedures. Use it throughout the Prefect setup steps below.
Start the server (persistent runs, UI, scheduling). See the Prefect server docs for details:
prefect server start
prefect config set PREFECT_API_URL=http://localhost:4200/api
prefect work-pool create default --type process
prefect worker start --pool default
UI at http://localhost:4200.
Deploying and running flows:
cd dags
prefect deploy pipelines/<name>/flow.py:flow --name <name>-dev --work-pool default
prefect deployment run '<flow-name>/<name>-dev'
Monitoring:
prefect flow-run ls
prefect flow-run inspect <flow-run-id>
prefect deployment ls
Prefect Cloud (managed, no local server): prefect cloud login (browser auth).
If Airflow (only if the user chose Airflow over Prefect):
Load the airflow skill before proceeding. It contains the canonical CLI reference, DAG patterns, and debugging procedures. Use it throughout the Airflow setup steps below.
Astronomer is the recommended way to run Airflow 3 locally: it handles all Docker setup. Requirements: Docker Desktop (macOS/Windows) or Docker Engine (Linux) running.
Install the Astro CLI:
brew install astro
curl -sSL install.astronomer.io | sudo bash -s
winget install -e --id Astronomer.Astro
Initialize Astro CLI inside the scaffold repo, then clean up the placeholder DAG:
astro dev init
rm dags/exampledag.py
Configure .astro/config.yaml to avoid port collisions with the local PostgreSQL on 5432:
postgres:
port: "5434"
webserver:
port: "8081"
Create docker-compose.override.yml to pass database credentials and mount the data directory into the scheduler container. Use host.docker.internal (macOS/Windows) or 172.17.0.1 (Linux) for SQL_DB_HOST so containers can reach the host database:
services:
scheduler:
environment:
- SQL_DB_HOST=host.docker.internal
- SQL_DB_PORT=${SQL_DB_PORT}
- SQL_DB_NAME=${SQL_DB_NAME}
- SQL_DB_USER=${SQL_DB_USER}
- SQL_DB_PASSWORD=${SQL_DB_PASSWORD}
- DATA_DIR=/usr/local/airflow/data
- AIRFLOW__LOGGING__ENABLE_TASK_CONTEXT_LOGGER=False
volumes:
- ${DATA_DIR:-./data}:/usr/local/airflow/data
dag-processor:
environment:
- DATA_DIR=/usr/local/airflow/data
- AIRFLOW__LOGGING__ENABLE_TASK_CONTEXT_LOGGER=False
The ${VAR} references are resolved from the .env file in the project root.
Start Astronomer:
astro dev start
UI at http://localhost:8081 (default credentials: admin/admin).
Managing Astro:
astro dev restart
astro dev stop
Triggering and monitoring DAGs:
astro dev run dags unpause <dag_id>
astro dev run dags trigger <dag_id>
astro dev run tasks states-for-dag-run <dag_id> "<run_id>"
7g. Verify the example pipeline
Run the example pipeline to confirm the full stack works end-to-end. Inform the user beforehand. For Prefect: ensure the server is running (step 7f), then deploy and run the example flow. For Airflow: unpause and trigger the example DAG (astro dev run dags unpause example && astro dev run dags trigger example). Check that data lands in the wid schema of the lens database.
7h. Save to infrastructure.md
Read infrastructure.md once more and verify all data stack details are accurately recorded:
- Databases: For each database, fill in the JSON block with fields appropriate to the type (engine, version, deployment; host/port for on-premise databases; account/project for cloud services like Snowflake or BigQuery). Add a Roles section listing every database role created during setup, its purpose, auth mechanism, and where the credential lives. At minimum there should be a read-only role (used by System2 for queries and dashboards, credentials in config.toml) and a pipeline write role (used by ETL code, credentials in the pipeline repo's
.env or equivalent). Never paste actual passwords.
- Orchestrator: name, deployment mode, UI URL, flows/DAGs location.
- Code repositories: path, remote URL, purpose.
- Data pipelines environment: venv path.
8. Agent instruction files
Check for AI agent instruction files that may contain coding conventions or preferences:
- Per-repo:
CLAUDE.md, AGENTS.md, .cursorrules, .github/copilot-instructions.md
- Global:
~/.claude/CLAUDE.md, ~/.cursor/rules/
- Save findings to
~/.system2/knowledge/infrastructure.md: read the file first, then use edit to place content in the correct template section.
9. Capture interaction preferences
Throughout the conversation the user may have expressed preferences about how they like to interact: verbosity, level of detail, how much autonomy to take, when to ask vs. act, etc. Save any such preferences to ~/.system2/knowledge/guide.md so they carry over to future sessions.
10. Walk through the configuration files
System2 splits its settings across two files. Orient the user on both, in this order, adapting depth to their technical level:
~/.system2/auth/.auth.toml (written by system2 config (credentials), written exclusively by system2 config). Do NOT show this file as an artifact and tell the user not to hand-edit it: every system2 config write rewrites it via parse-mutate-stringify, so any manual edits would be lost. Instead, walk them through what system2 config manages on their behalf:
- OAuth and API key tiers (
[llm.oauth], [llm.api_keys], with per-provider sub-sections). The two credential tiers. OAuth subscriptions (Anthropic Pro/Max, OpenAI Codex, GitHub Copilot) are tried first; API keys (Anthropic, OpenAI, Google, etc.) are the failover tier. Each tier has its own primary and ordered fallback list. Use system2 config to add/remove providers, add multiple labeled keys per API-key provider for rotation, set primary, and reorder fallbacks. OAuth credential JSON files live alongside .auth.toml at ~/.system2/auth/<provider>.json (also 0600). Mention that System2 automatically fails over between keys and providers when errors occur.
- Per-role model pins (
[llm.api_keys.<provider>.models]): a TOML table with role keys inside (e.g. narrator = "claude-haiku-4-5-20251001"). API-keys tier only; the OAuth resolver picks one model per provider. Managed via system2 config.
- Services and the web search toggle (
[services.brave_search], [tools.web_search].enabled): Brave Search credentials and the web-search on/off flag. Setting a Brave key auto-enables web_search. Managed via system2 config.
If the daemon is running when credentials change, restart it with system2 stop && system2 start to pick up the change.
~/.system2/config.toml (user-edited operational settings). Show this file (show_artifact) and walk them through the sections in order. This file holds operational tunables only; no credentials. Hand-editing is expected and safe: the daemon reads it but never writes it.
[agents.<role>]: Per-role behavior overrides for any agent (guide, conductor, narrator, reviewer, worker): thinking_level (off/minimal/low/medium/high/xhigh) and compaction_depth (how many context compactions before pruning). Only specified fields override the library defaults. This is how they can, for example, raise the Guide's thinking level. (Per-role model pins live in .auth.toml under [llm.api_keys.<provider>.models] and are managed via system2 config.)[databases.*]: The database connections just configured. Each entry uses the section name (e.g. analytics in [databases.analytics]) as the identifier throughout System2, including in HTML dashboard queries via the postMessage bridge. Remind them they can add more databases later by adding a [databases.<name>] section and installing the corresponding Node.js driver (npm install --prefix ~/.system2 <package>), or they can just ask you and you'll handle both steps.web_search_max_results (top-level scalar): operational knob for the web_search tool. The on/off enabled flag itself lives in .auth.toml and is managed by system2 config; the result-count cap is a hand-editable operational setting and lives here.- Operational sections (
[backup], [session], [logs], [scheduler], [chat], [knowledge], [delivery]): These exist with sensible defaults and rarely need adjustment. Point out scheduler.daily_summary_interval_minutes as the one users might want to tune (controls how often the Narrator writes summaries).
Keep this conversational, not a lecture. A couple of sentences per section is enough. The goal is awareness of the split (credentials via system2 config, operational settings hand-edited in config.toml) and where each knob lives, not memorization of every field.
11. Wrap up
Summarize what was configured: the user's profile, detected infrastructure, data stack, and code repository. Then ask what they'd like to work on first. Be excited to start working with data!
