Why this reading
Working code is not the same as professional procedure.
A scientific coding agent can generate executable analysis code while still applying the wrong conventions, assumptions, or review logic.
The new Scientific Agent Skills paper treats recurring expert procedure as a versioned artifact that can be loaded selectively instead of rediscovered from model memory for every task.
Reading order
Your 30-minute plan.
Explain why code can run and still be scientifically wrong.
Read skill structure, progressive disclosure, context measurements and limitations.
Compare “Thin MCP, Thick Skills,” deterministic rules and LLM-dependent specification work.
Connect reusable skills to structured standards, traceability and protocol-to-CSR pipelines.
Design one reusable SP skill with scope, rules, tools and escalation.
Open-access sources
Fresh agent architecture plus clinical-programming and CDISC context.
Brief background
Put reusable procedure in a reviewable layer.
Scientific Agent Skills packages procedural knowledge into versioned skill directories built around human-readable instruction files, with optional references, scripts and assets.
Its 163 short resident descriptions require 14,246 tokens - about 7.1% of a 200,000-token reference context window - while the full instruction and reference corpus is far larger.
Progressive disclosure keeps 83.7% of the documentation unread unless an activated skill points to it. The median documented workflow uses 23.9% of the reference context, while 29 of 46 workflows would overflow if every reference file were loaded.
The paper also states its limitations clearly: it does not provide task-level evaluation of the whole library, and a selected skill can still be wrong or misapplied.
ClinAgent provides the clinical-programming analogue: stateless data tools stay thin, while prompts, CDISC/SAS rules, decision logic and deterministic validators live in skills.
Key vocabulary
Fifteen terms for reusable agent expertise.
| Term | 中文 | Meaning / use |
|---|---|---|
| procedural knowledge | 程序性知识 / 流程知识 | Knowledge about how to perform a task correctly, including accepted methods, conventions, checks, and caveats. |
| agent skill | 智能体技能 | A versioned instruction package that gives an agent task-specific procedures, references, scripts, and constraints. |
| progressive disclosure | 渐进式披露 | Keeping only short skill descriptions in context and loading detailed instructions or references only when needed. |
| resident context | 常驻上下文 | Information that remains continuously available to the model rather than being loaded on demand. |
| instruction file | 指令文件 | A human-readable file that records the procedure, constraints, and usage guidance for a skill. |
| domain logic | 领域逻辑 | Rules and decisions specific to a professional domain, such as CDISC standards or SAS review conventions. |
| stateless tool | 无状态工具 | A tool that performs a narrow input-output operation without carrying domain-specific workflow reasoning. |
| rule engine | 规匙引擎 | Executable logic that applies explicit rules consistently, often used for validation or classification. |
| skill routing | 技能路由 | Selecting which skill or procedural module should be loaded for the current task. |
| context budget | 上下文预算 | The limited amount of model context available for instructions, evidence, references, and working information. |
| lexical separability | 词汇可区分性 | How distinct skill descriptions are in wording, which affects the risk of confusing similar skills. |
| scope boundary | 适用范围边界 | A clear statement of what a skill covers and what it intentionally does not cover. |
| versioned capability | 版本化能力 | A controlled capability whose instructions, rules, and dependencies are tied to an identifiable version. |
| portable architecture | 可移植架构 | A design that can work across different agent hosts or models rather than depending on one vendor. |
| scientifically defensible | 科学上可辩护的 / 可论证的 | Supported by appropriate procedures, assumptions, standards, evidence, and limitations - not merely executable code. |
Useful phrases
Language for agent architecture and SP automation.
- working code is not the same as a defensible analysis - Working code is not the same as a defensible analysis.
- encode recurring expert judgment once - A skill can encode recurring expert judgment once instead of rediscovering it for every task.
- load detailed guidance only when the task calls for it - The agent should load detailed guidance only when the task calls for it.
- keep data access thin and domain logic thick - ClinAgent keeps data access thin and domain logic thick.
- make deterministic rules executable - High-confidence compliance and validation rules should be made executable.
- separate procedural knowledge from the underlying model - The architecture separates procedural knowledge from the underlying model.
- version the instructions together with the rules - Teams should version the instructions together with the rules.
- declare what the skill does not cover - A trustworthy skill should declare what it does not cover.
- route the task to the narrowest relevant skill - The system should route the task to the narrowest relevant skill.
- treat standards metadata as an authoritative source - CDISC standards metadata should be treated as an authoritative source rather than improvised from model memory.
Comprehension
Five questions.
- Why does the paper distinguish executable code from a defensible analysis?
- How does progressive disclosure reduce context burden?
- Why is a reusable skill different from a stateless tool?
- What does ClinAgent's 72.1% overall derivation accuracy imply?
- Why should standards, deterministic validation, study context and human review remain separate layers?
Retelling
Say it three times.
- 30 seconds · Define an agent skill and progressive disclosure.
- 45 seconds · Description → routing → skill → references/scripts → tool → result.
- 60 seconds · Divide responsibility among the LLM, skill, deterministic validator, CDISC metadata and human reviewer.
5-minute output task
Design one reusable SP agent skill.
- Minute 1: Choose ADAE, ADLB, log QC, TFL denominators, PK windows or SDTM terminology.
- Minutes 2-3: Define description, scope, declines, authoritative references, procedure, rules and tools.
- Minute 4: Add version, CDISC applicability, study configuration, regression tests and edge cases.
- Minute 5: Explain why this knowledge should be a reusable skill instead of one giant prompt.
One sentence to keep
A strong SP agent should not be expected to rediscover professional procedure from model memory on every task; recurring expertise should be versioned, reviewable, selectively loaded, and paired with deterministic checks where correctness must be exact.