Battle-Tested Research Prompts
An elegant, curated gallery of research prompts with proven yield — the
actual prompts (and full transcripts) that frontier AI labs released alongside
significant, verified research results.
The Gallery
OpenAI
| Entry |
Field |
Yield |
Date |
| Cycle Double Cover Conjecture |
Graph theory |
Complete proof of the CDC conjecture (all finite bridgeless loopless multigraphs), via GPT 5.6 Sol Ultra with 64-agent adversarial search |
2026 |
| First Proof? — solve/verify/refine templates |
Competition mathematics |
Programmatic pipeline producing model-verified solutions to 1stproof.org problems (3/4 correct on the BSS seed idea for Problem #6) |
Feb 2026 |
| Gluon → graviton single-minus amplitudes |
Quantum gravity |
ChatGPT 5.2 Pro generalized the gluon amplitudes paper to gravity from a one-paragraph research instruction, producing the preprint "Single-minus graviton tree amplitudes are nonzero". Full 110-page conversation transcript released |
Feb 28, 2026 |
| Goemans cost conjecture |
Combinatorial optimization / network flows |
Counterexample found by Dmitry Rybin / GPT-5.6 Pro; independently kernel-checked in Lean 4 |
Jul 22, 2026 |
Recent / peer review pending
| Entry |
Field |
Yield |
Date |
| Crouzeix conjecture |
Matrix analysis |
Candidate proof from Shanmu Jin / GPT-5.6 Sol — recent / peer review pending |
Aug 2026 |
FrontierMath / independent researchers
| Entry |
Field |
Yield |
Date |
| Absolute Galois group of Q₂ |
Number theory |
Explicit presentation of the absolute Galois group of Q₂, with two independent Lean formalizations and finite-quotient verification |
Jun 24 / Jul 26, 2026 |
| White + Mason matroid conjectures |
Matroid theory |
Computer-free counterexamples to White's strong conjecture and Mason's log-concavity conjecture, found with GPT-5.5 Pro and preserved in public chat logs |
Jul 2, 2026 |
Anthropic
| Entry |
Field |
Yield |
Date |
| De novo protein binder design campaign |
Computational biology |
Wet-lab-confirmed binders against 14/15 targets; 22–35% hit rates vs. 10–15% human baseline; state-of-the-art affinities. ~30k-token autonomous campaign prompt + kickoff messages |
Aug 18, 2026 |
| Automated W2S researcher |
AI alignment research |
9 parallel Claude Opus 4.6 agents reached PGR 0.97 vs. 0.23 human baseline on weak-to-strong supervision. Full prompt stack published (system prompt + skill + critic prompts, MIT) |
Apr 14, 2026 |
Google DeepMind / Google Research
| Entry |
Field |
Yield |
Date |
| Co-Scientist agent prompts & research goals |
Biomedical hypothesis generation |
Nature-published multi-agent system; lab-validated AML repurposing (KIRA6), anti-fibrotic targets (91% response blockade), and re-discovery of unpublished cf-PICI mechanism |
May 19, 2026 |
Academic / other labs
| Entry |
Field |
Yield |
Date |
| SPARK — System of Pathology Agents |
Cancer pathology |
Nature Medicine: autonomous hypothesis→code→validation pipeline produced survival-validated prognostic biomarkers on real patient cohorts. Full CrewAI agent/task prompts published |
Apr 29, 2026 |
Methodology / autonomous AI-math
| Entry |
Field |
Yield |
Date |
| Autonomous sum-product disproofs |
Additive combinatorics / AI-math methodology |
GPT-5.5 Pro's three-stage plan→construct→review pipeline produced correct proofs in 7 of 8 independent trials; code, transcripts, and generated proofs released |
Jul 9, 2026 |
Yield criteria (firm)
All four requirements must hold for a prompt to enter this gallery:
- The prompt is available — released verbatim by its source (PDF,
dataset, or web page) and collected here as a copy-paste-runnable
.md
file.
- It is related to a research task — mathematics, science, or engineering
discovery, not a demo or toy.
- It is battle-tested — actually used in the run that produced the
result, not a post-hoc reconstruction.
- It produced significant yield in its field — a result independently
validated by publication in a peer-reviewed venue, wet-lab confirmation,
formal verification (e.g. Lean), or expert review. The yield is stated
concretely in every entry README.
Never include a prompt without a proven yield.
Usability requirement: the published prompt must carry the causal weight
of the yield. Entries whose published prompts are one-liners (or otherwise
insubstantial) while the actual instructions lived in undisclosed harness
system prompts, skills, or project instructions are unusable and are
excluded — this repository collects useful prompts, not exercises in
reverse-engineering hidden context. Every entry README still discloses
whatever user-authored context remains unpublished (harness configs, tool
schemas) so yields are never overstated.
Conventions
- Hard requirement: every entry contains one or more pure prompt files
(named
prompt*.md) whose entire content is the prompt — no commentary, no
headers, no provenance notes. Copy the file's full content and run it
verbatim. Supporting material (provenance, analysis, transcripts,
context) lives in separate files (README.md, *-context.md,
transcript.md, etc.).
- Each entry lives in
prompts/<lab>-<entry>/ (flat — no per-company nesting) with:
README.md — metadata table (lab, model, field, yield, date, sources)
and analysis of the prompt techniques that worked.prompt*.md — the verbatim, runnable prompt(s).- Full transcripts where released.
- Extraction provenance is documented in the entry README or context file;
the original source is always linked and authoritative.
Rejected entries
| Candidate |
Rejected because |
| Anthropic — Riemann zeta 67.2% critical-line bound (Aug 10, 2026) |
Has hidden, undiscoverable user prompts: the published human inputs are one-liners ("Take a real stab at the Riemann hypothesis", "Keep going / Believe in yourself"); the yield derives from an unreleased research model's hidden harness system prompt, skills, and CLAUDE.md-style instructions. Unusable — deleted. |
| Anthropic — NMR / LC-MS analytical-chemistry one-liners (Aug 18, 2026) |
Same character: two-sentence prompts whose yield is attributable to the unpublished Claude Science harness and skills. Deleted from the protein-design entry. |
| OpenAI — Graviton amplitudes preprint alone (cdn.openai.com/pdf/graviton.pdf, Mar 4, 2026) |
The preprint itself contains no prompts. The graviton work was ACCEPTED via its companion transcript release (gluon-to-graviton-paper.pdf), which publishes the full ChatGPT conversation — see the Gluon → graviton entry. |
