🎣 BAIT — Biotic AI Toolkit
Teach your AI coding agent to work with IMR Biotic data — safely, on your own machine.
BAIT is a knowledge pack that turns a general coding agent (Claude Code, Codex, Cursor, Antigravity, Mistral Vibe, …) into a competent assistant for the
Institute of Marine Research (IMR) NMD Biotic database. Install it and your agent can:
- Install, update, and maintain the Biotic database
(BioticExplorerServer).
- Connect to it from R using tidyverse syntax.
- Answer data questions — "What is the largest cod in the database?", "Map all cusk
caught on the EggaN surveys."
- Make plots, tables, maps, life-history analyses, and dashboards with
ggplot2,
ggOceanMaps,
leaflet,
ggFishPlots, and
BioticExplorer/Shiny.
- Extract Norwegian catch data — sales notes (landings) and ERS logbooks with
RstoxUtils, shaped for spict-wise and
gadget-wise.
- Build reproducible Quarto analyses and reports in HTML, Word, and PDF, using an
organisation's Word template when supplied, the bundled IMR web-report template when
explicitly requested, or a generic Manuscript design otherwise.
It is a knowledge pack, not a model trained on Biotic records. Local R/DuckDB work
does not guarantee that prompts, tool inputs, logs, or outputs stay on-device or are excluded
from provider processing, retention, or training. See
Privacy.
🔒 Privacy first
Most Biotic data are open; some subsets are restricted or sensitive. Classify data and outputs under IMR rules. Local execution does not guarantee that agent prompts, tools, logs, or outputs stay on-device:
- BAIT contains instructions only — never database records. Examples and cookbook recipes
use code and synthetic values, not real records. Ignore files block common data extensions
as a safety net.
- The database lives outside this repo (
~/IMR_biotic_BES_database/ by default on
macOS/Linux; %USERPROFILE%\IMR_biotic_BES_database\ on Windows). These are defaults; the
configured bes_db_path takes precedence.
- During setup, review provider training, retention, and tool-transmission controls. An
onboarding record reflects a past review; settings may change, and a training opt-out does
not guarantee no transmission or retention
(
skills/biotic-privacy).
🚀 Quickstart
- Get an agent: Claude Code or
Codex, etc.
- Tell your agent:
install https://github.com/DeepWaterIMR/BAIT
The agent runs bait-install and walks you through:
a privacy and data-classification review → cloning BAIT to a sensible place
→ installing the skills globally (so BAIT works in every project) → setting up the
database → verifying the connection.
- Ask away: "What's the largest cod in the database?", "Map cusk on EggaN."
Installed once per machine, then available in all your projects. Keep it current with
"update BAIT" (bait-update).
If you use Fiskwise's managed pack lifecycle, follow its fiskwise-install, fiskwise-update,
and fiskwise-uninstall skills instead of repeating clone and skill-registration steps. BAIT
remains usable as a standalone pack. Shared scientific writing and generic Quarto/package
loading guidance comes from wise-writing and is installed on demand. The Biotic database is
separate user data: an existing database is adopted, a new one is recorded as BAIT-created only
after an approved successful compile, and refresh/removal are separate opt-in actions.
Use fiskwise-workflow for requests spanning packs. For managed R tasks, always use
r-code-style.
Manual / advanced: you can still git clone https://github.com/DeepWaterIMR/BAIT
yourself and open your agent in the folder — it reads CLAUDE.md /
AGENTS.md automatically. The agent-driven install above just does this for
you and sets up cross-project use.
🗂️ What's inside
| Folder | What it holds |
|---|
CLAUDE.md / AGENTS.md | Agent entry points (identity + privacy guardrails + router) |
skills/ | One skill per capability — the agent reads these on demand |
knowledge/ | Vendor-neutral reference: data model, field glossary, connection, packages |
cookbook/ | Standardized query → code recipes (grows over time) |
examples/ | Runnable .R scripts mirroring the cookbook |
scripts/ | Maintenance scripts (e.g. regenerate the field glossary from the XSD) |
docs/ | The BAIT website (Quarto → GitHub Pages) |
🧰 Siblings
BAIT is the data layer. The packs that build on it:
- wise-writing — scientific writing style, document conventions, and the project skeleton the other packs scaffold from.
- index-wise — sdmTMB spatiotemporal survey indices, built from Biotic data through BAIT.
- spict-wise — SPiCT biomass-dynamic stock assessments, driven by an index from index-wise and a catch series from
catch-data.
- gadget-wise — gadget3 length- and age-structured models; BAIT is its default data-loop for Biotic length, age and maturity data.
- wise-presentations — Quarto RevealJS slides.
Install whichever you need alongside; the skills stay out of each other's way.
🧠 It learns from you
When you ask something new, the agent offers to save the solution as a cookbook recipe.
Review it, commit it, and the next person's agent benefits. The cookbook is IMR's shared,
version-controlled institutional memory. See CONTRIBUTING.md.
📚 Website
Example queries and getting-started guidance: https://deepwaterimr.github.io/BAIT/
(built from docs/).
🙋 Contact
Maintainer: Mikko Vihtakari ([email protected]),
DeepWaterIMR. Contributions welcome — see
CONTRIBUTING.md.
