agents-for-scientific-software

KO | EN

gitlite — search

#claude-code #shell #c #editor #cpp #rust #reactjs #devops #security #kubernetes

agents-for-scientific-software

(★ 9)

companion artifacts for RosettaCon 2026 tutorial - using AI agents for scientific software engineering

파일 탐색기

ZIP 다운로드

- - SKILL.md
- - SKILL.md
- - SKILL.md
LICENSE
ORACLES.md
README.md

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

Companion material for *Sampling is Cheap, Scoring is Hard* — Summer RosettaCon, 2026-08-05.

Full slides here: [RosettaCon 2026 - using AI agents for scientific software engineering](https://docs.google.com/presentation/d/1-cSUsBtvBvFEBKFR7rtbEOAJOuN3aM5p/edit?usp=sharing&ouid=114333546980887011542&rtpof=true&sd=true)

Three reusable prompts, one per act. Each act has its own scoring function, and they get progressively
more mechanizable — which is exactly why you delegate more as you descend.

| Act | Skill | What it does | Who scores it |
|---|---|---|---|
| I | `/problem-statement` | Interviews you until your problem statement holds up | You. Not mechanizable. |
| II | `/design-doc` | Turns that statement into a design, then attacks it | Adversarial subagent + you |
| III | `/validate-science` | Hunts for silently wrong numbers in an implementation | Oracles, tests, CI |

## Start here — pick a tier

Ascending cost. Tier 1 works if you have installed nothing at all.

### Tier 1 — Print the checklist

[`ORACLES.md`](ORACLES.md). Zero tooling. Works with Claude Code, Cursor, Codex, a browser chat window,
or a pen. This is a checklist for *you* — the agent cannot name the invariants for you, so the value
is in your head, not in a config file.

### Tier 2 — Six lines in your project's context file

Ten minutes, once. Add to your repo's `CLAUDE.md` or `AGENTS.md`:

```markdown
## Numerical conventions
- Residue numbering is 1-based pose numbering throughout. Convert at the PDB I/O boundary only.
- Energies are REU. Never mix with kcal/mol without an explicit conversion.
- Float comparisons use `pytest.approx(rel=1e-9)`; tighter tolerances are flaky on this hardware.
- Golden outputs live in `tests/data/golden/`. Never regenerate them from the current code.
```

Best effort-to-payoff ratio here. It fires on every task without being invoked, and it is where the
general checklist becomes specific enough to prevent real bugs. Write it yourself — do not delegate it to
an agent, for the same reason you do not delegate the problem statement.

### Tier 3 — Install the skills

Worth it if you repeat these procedures often, or if you are standardizing practice across a lab.

```bash
git clone <this-repo> rosettacon-skills
cp -r rosettacon-skills/skills/* ~/.claude/skills/
```

Or copy them into a single project's `.claude/skills/` to scope them to that repo. Verify with `/help`;
invoke by name: `/problem-statement`.

**These are not Claude Code — specific.** `SKILL.md` is an open cross-tool format — the same
`name`/`description` frontmatter plus markdown body is read by Codex CLI, Cursor, Gemini CLI, Copilot,
and others. Directory locations and loading semantics differ per tool, and support is uneven (Cursor is
project-scoped only, last we checked). Consult your tool's docs for where the folder goes; the file
contents port unchanged.

### Honest note on what is worth your time

`validate-science` carries most of the value in this repo, and its value is the domain content inside it
— the silent-failure taxonomy and the oracle ladder — not the fact that it is packaged as a skill. It
works just as well pasted into a conversation. `design-doc` is the thinnest of the three; a capable
reasoning model produces most of that structure unprompted, and the genuinely additive parts are the
success-criteria-with-named-oracles requirement and the adversarial subagent charge.

Take the checklist. The rest is convenience.

## Use

```
/problem-statement          # Act I — start here, always
/design-doc                 # Act II — feed it the statement from Act I
/validate-science           # Act III — after the implementation exists
```

Model tiers matter more than model names, which churn quarterly:

- **Act II** wants a reasoning-heavy tier. Currently Fable 5 or Opus 5.
- **Act III implementation** only wants execution competence. Opus or Sonnet is fine.
- **Act III validation** wants reasoning again. Do not economize on the scoring step.

## The one rule

**Never let the agent choose the expected value.**

If expected outputs come from running the agent's own implementation, the test proves self-consistency
and enshrines the bug. Expected values come from theory, from an independent implementation, from
published data, or from your own hand calculation.

## The taxonomy

Three places knowledge can live, and they are not interchangeable:

| | Audience | Loaded | Content |
|---|---|---|---|
| `ORACLES.md` | you | never — you read it | domain-general method |
| `CLAUDE.md` / `AGENTS.md` | the agent | every turn | project-specific facts |
| `SKILL.md` | the agent | on demand | repeated procedures |

General knowledge to the human. Specific facts into always-loaded context. Procedures on demand.

Do not paste the oracle ladder into `AGENTS.md` — it is ~1500 words of general advice charged to every
turn, including when you are fixing a typo. Context-file bloat is the standard failure mode: the file
grows, signal-to-noise degrades, and the agent starts effectively ignoring it.

agents-for-scientific-software

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Download README