Coding-agent skills¶

Science ADK ships seven skills that teach a coding agent how to run the research loop properly: how to phrase a falsifiable hypothesis, how to author an agent node, what to do when a gate fails, and when a result is worth recording as a learning.

Install¶

From a checkout of the repository:

node ./science-adk/bin/install.mjs

The installer detects which coding agents you actually have and writes only to those, so it does not litter the home directory of someone who uses one tool.

Target	Detected by	Skills installed to
Claude Code	`~/.claude`	`~/.claude/skills/`
Cline / Roo	`~/.agents`, `~/Documents/Cline`	`~/.agents/skills/`
Antigravity	`~/.gemini`	`~/.gemini/config/plugins/science-adk/skills/` — see Getting started in Antigravity
OpenCode	`~/.config/opencode`	`~/.config/opencode/skills/`
Cursor	`~/.cursor`	`~/.cursor/rules/` (flattened to `.mdc`)

The scientific integrity rules are installed alongside them as science-adk-rules/.

Flags¶

node bin/install.mjs           # install into every detected agent
node bin/install.mjs --local   # install into ./.agents and ./.claude instead
node bin/install.mjs --list    # show what would go where, change nothing
node bin/install.mjs --help    # usage

--local is the right choice for a shared repository: the skills travel with the project rather than the machine.

cd my-research
node /path/to/science-adk/bin/install.mjs --local
ls .agents/skills

agent-authoring  auditing  debugging  evolving
experiment-design  optimizing  research  science-adk-rules

The installer also checks whether the science-adk command is on your PATH and offers to install the Python package if it is not.

The seven skills¶

Skill	Use it when
`research`	You are starting from a question and need the whole loop.
`experiment-design`	You need a hypothesis that is falsifiable and a DAG that tests it.
`agent-authoring`	You are writing the Python for a node.
`debugging`	A run failed, or a gate did.
`optimizing`	The result is real but not yet good enough.
`evolving`	One experiment answered its question and suggested the next.
`auditing`	A run passed its gates and now needs a verdict.

Each is plain Markdown with YAML frontmatter, readable on this site and editable in the repository under skills/.

Using them¶

Once installed, describe the research goal to your coding agent as you normally would:

Use the science-adk research skill. I want to know whether a GNN embedding predicts aqueous solubility better than Morgan fingerprints. Set up the project and run the first experiment.

The agent scaffolds the project, writes the tools, drafts the hypothesis, builds the DAG, writes the node code, validates, runs, and reports back with the gate results.

The agent cannot audit its own work into existence

A coding agent can propose pillar ratings, but the gates are computed by the runtime from the recorded trace. No prompt makes a fabricated result pass tool_use_verified. That separation is the whole point — see integrity gates.

Next¶

Read the skills · Concepts