Use this skill when you need to survey a research domain, discover and formulate open problems worth solving, prioritize them by impact and feasibility, and commission Principal Scientist agents to execute the research. Activate when the user wants to define a research agenda from scratch, identify the most valuable unsolved problems in a field, or systematically advance knowledge across a domain over multiple research cycles — rather than executing a single already-defined project.
Survey a research domain, discover and formulate the problems most worth solving, and commission Principal Scientist agents to execute the research — maintaining an evolving map of what is known, what is open, and what to pursue next.
The Computer Scientist is the highest-level agent in the research hierarchy. It operates above the Principal Scientist and focuses on what to research, not how to research it. Its primary output is a prioritized problem registry that drives Principal Scientist commissions.
Full hierarchy:
Computer Scientist ← you are here
└── Principal Scientist (x M)
├── Lead Researcher (x N per PS)
│ └── hypothesis-generation → literature-synthesis →
│ experiment-design → code-replication →
│ research-writing → ieee-paper-generator
└── Auto-Benchmark
What the Computer Scientist does that no lower layer does:
Collect the domain context before scanning the field. Ask explicitly for any missing inputs.
| # | Question | Why it matters |
|---|---|---|
| 1 | What is the research domain or sub-field? | Scopes the field survey |
| 2 | What is the strategic objective? (advance SOTA / defend competitive position / explore white space / solve a specific bottleneck) | Determines problem selection criteria in Phase 3 |
| 3 | What is the organization's current capability baseline? (existing systems, datasets, compute, team expertise) | Gates feasibility scoring in Phase 3 |
| 4 | Are there known constraints? (time horizon, compute budget, target venues, ethical boundaries) | Shapes prioritization in Phase 3 |
| 5 | Are there Principal Scientist portfolios already active that should be considered? | Prevents redundant commissioning |
| 6 | What should the output of this cycle be? (problem registry only / commissioned research / full pipeline to papers) | Determines how far this activation runs |
Produce a Domain Context Brief (markdown, ~1 page):
Get explicit user confirmation before beginning the field survey.
Conduct a broad scan of the domain to build a current, accurate picture of the state of knowledge.
Scan across all four dimensions before identifying problems:
| Dimension | Questions to answer |
|---|---|
| State of the art | What are the best-performing methods on major benchmarks? Who holds each leaderboard? What are reported ceilings and why? |
| Open problems | What does the field explicitly acknowledge as unsolved? What do papers list as future work? What do practitioners complain about? |
| Recent breakthroughs | What has changed in the last 6–18 months? What newly proposed methods have not yet been fully exploited? |
| Strategic white space | Where has the field under-invested relative to its importance? What do competitors avoid and why? |
For each source type, extract structured findings:
Produce a Field Map structured as:
## Field Map — [Domain] — [Date]
### State of the Art
- [Benchmark]: best score [X] by [System], achieved via [method]. Rate of improvement: [fast/slow/plateauing].
- [Benchmark 2]: ...
### Confirmed Open Problems
- [Problem A]: acknowledged in [N] papers; no solution proposed
- [Problem B]: partial solutions exist but significant gap remains
### Recent Breakthroughs (not yet fully exploited)
- [Technique X] from [Paper, 2025]: reported [Y]% gain on [task]; adoption is low outside the original lab
### Strategic White Space
- [Area Z]: high potential impact, sparse publication activity, no clear leader
From the Field Map, identify and catalog candidate research problems.
Classify every candidate problem into one of four types:
| Type | Definition | Example |
|---|---|---|
| Known-unsolved | The field knows the problem exists and has tried; no satisfying solution yet | Long-context faithfulness in LLMs |
| Newly tractable | Recent breakthroughs make a previously infeasible problem now attackable | Sparse attention enabling 1M-token context |
| Strategic gap | High-impact area where competitors have not invested, matching our capabilities | Domain-specific retrieval for regulated industries |
| Fundamental bottleneck | Solving this would unblock multiple downstream problems | Better uncertainty quantification for active learning |
For each candidate problem, fill in the full formulation before scoring: