Why this matters
Scientific progress depends on methods that others can actually run.
Too much know-how remains private
Labs routinely rely on undocumented prompts, ad hoc command history, and one-off scripts that work once but do not travel well across people, projects, or time.
Scientific tools still carry hidden operational cost
Valuable systems like NCBI E-utilities and standalone RPS-BLAST remain powerful, but much of their real usability still depends on knowing the right choreography, formats, defaults, and failure modes.
Skills create a public operational layer
A skill can preserve exact commands, references, and outputs while making a workflow reusable through natural language. That makes method transfer easier without turning science into a black box.
Manifesto
We should publish operational knowledge, not just results.
A modern scientific result is rarely just a paper and a dataset. It is also a retrieval routine, a screening heuristic, a parsing step, a briefing pattern, a quality check, and a chain of decisions that are often lost between people. Skill-based scientific tooling treats those operational layers as first-class research assets.
How It Works
Natural-language intent becomes a reproducible scientific workflow.
1
Intent
Ask for a scientific outcome in natural language.
2
Skill
Translate that intent into an explicit, guarded workflow.
3
Artifacts
Keep raw outputs, structured files, and inspectable summaries.
4
Reuse
Repeat the process across projects, teams, and time windows.
The key claim is not that language replaces method. The claim is that language can trigger published, inspectable, reusable methods that leave behind stable outputs.
Try The Model
Ask for scientific work the way scientists actually think.
Weekly literature update
Use $ncbi-eutilities-assistant to show me the last 7 days of CRISPR progress on PubMed and write a concise update brief.
Structured review prep
Use $ncbi-eutilities-assistant to run a PubMed review workflow for base editing over the last 30 days and extract structured abstracts.
Conserved domain annotation
Use $rpsblast-assistant to download the minimal CDD assets, run RPS-BLAST on my FASTA, and explain the processed report.
Representative Skills
Two concrete examples of what this model looks like in practice.
NCBI / Literature
ncbi-eutilities-assistant
Wraps official NCBI E-utilities into reusable PubMed workflows for search, review extraction, and recent-progress briefing. The aim is not just easier querying, but a repeatable way to turn literature monitoring into a shareable workflow artifact.
- `esearch`, `esummary`, `efetch`, `elink`, `epost`
- `pubmed-review` for structured JSON and JSONL
- `pubmed-update-brief` for recurring topic surveillance
brief.md
records.jsonl
manifest.json
CDD / Annotation
rpsblast-assistant
Turns the standalone RPS-BLAST plus `rpsbproc` workflow into a local, explainable skill for conserved-domain annotation. The method stays explicit, but the operational burden becomes easier to teach, delegate, and reproduce.
- Asset download, setup checks, and deterministic execution
- Correct archive-mode handling for `rpsbproc`
- Natural-language interpretation of processed outputs
Cdd database
ASN.1 archive
rpsbproc output
Documentation For Scientists
Each flagship skill is paired with manuscript-style exposition.
What the manuscripts do
They frame each skill as an interface-layer contribution: why the underlying scientific tool matters, where usability breaks down, and how a local AI skill can preserve rigor while improving adoption in real research settings.
What they do not claim
These skills do not replace databases, command-line tools, or expert judgment. They standardize retrieval and execution so interpretation, criticism, and extension can happen on top of inspectable artifacts.
Use It Locally
Built for scientific workflows that need to stay inspectable.
Structured Intelligence is meant for local AI-assisted research environments such as Codex and Claude Code. Each workflow is packaged as a plain-text skill bundle so the reasoning, references, scripts, and outputs remain easy to inspect, adapt, and improve.
This project is an independent, unofficial interaction layer around existing scientific tools and services. It does not claim ownership of upstream software, databases, or documentation, and it does not imply endorsement by the original authors or institutions.