mirror of
https://github.com/merlinhu1/Battle-tested-Research-Prompts.git
synced 2026-08-26 12:20:55 +02:00
MerlinH
4b069723f7
- OpenAI: Cycle Double Cover conjecture prompt; First Proof solve/verify/refine templates - Anthropic: Riemann zeta 67.2% critical-line prompts + full transcript; protein binder design campaign prompts - Google DeepMind: Co-Scientist agent prompt templates and research goals - Root README index + CONTRIBUTING with firm inclusion criteria (proven yield required)
De Novo Protein Binder Design Campaign
| Lab | Anthropic |
| Model | Claude Mythos Preview and Opus 4.8 (Claude Science agent harness) |
| Field | Computational biology — de novo miniprotein binder design |
| Result (yield) | Wet-lab-validated binders against 14 of 15 targets (354 binders from 1,320 designs). Hit rates 22.6–35.1% vs. the 10–15% typical of human campaigns; high-affinity binders (KD < 10 nM) against ≥6 targets; best designs exceeded published state-of-the-art affinities (e.g. RBX1: 40% hit rate vs. 3.7% for competition winners). Validated independently by Adaptyv Bio and Twist Bioscience. |
| Date | August 18, 2026 |
| Announcement | https://www.anthropic.com/research/Claude-accelerates-protein-design |
| Prompt release (CC-BY-4.0) | https://huggingface.co/datasets/Anthropic/claude-protein-binder-design |
Why it's battle-tested
A ~30,000-token campaign prompt ran autonomously for 48 hours (multi-target mode, up to 12,500 H100-hours) or 24 hours per target (single-target mode), with no scientific guidance from humans — producing experimentally confirmed binders that match or beat expert human designs.
Prompt architecture
multi-target-prompt.md— the full 14-target campaign prompt (~114 KB), carried as system context by every agent at every depth. Specifies task, all 14 targets with UniProt IDs and oligomeric states, a curated reference corpus, design/folding tools, budgets, spend ceilings, safety gates, scoring, and deliverables.multi-target-kickoff.md— the user message that starts the campaign: a single, precise instruction block with an ordered first-actions list, clock definition (T0 + 48 h), and "Do not ask me any questions or wait for approval."single-target-kickoff.md— the 24-hour, $10k single-target variant.chemistry-prompts.md— the two-sentence NMR and LC-MS prompts from the companion analytical-chemistry result (Opus 5 matched the contract lab's analysis in 23 and 19 minutes).
Notable techniques: explicit sub-agent scope discipline ("your authoritative SCOPE is your task brief from your parent"), spend governors with hard ceilings, originality/novelty checks, counter-target selectivity requirements (GDF-8 vs. GDF-11), and verified-figure grounding.
Files
multi-target-prompt.md— verbatim from the HuggingFace release (CC-BY-4.0).multi-target-kickoff.md— verbatim kickoff message.single-target-kickoff.md— verbatim kickoff message.chemistry-prompts.md— verbatim NMR / LC-MS prompts.
Further reading
- Technical report: https://www-cdn.anthropic.com/30bf50e22a01388bb29bf077ee3f244531594b7a.pdf
- Chemical analysis technical report: https://www-cdn.anthropic.com/9f08da5189ac269b3242ca760de9823805c3f5f6.pdf
- Single-target prompts (16 targets incl. GDF-8-latent): see
prompts/single_target/in the HF dataset.