| name | research |
| description | Run staged web investigation and produce one durable paper-style markdown report with BibTeX-compatible references. |
research - Staged Web Investigation Protocol
When to Use
- You need to investigate a topic using external web sources
- You need a durable, auditable report (not only chat output)
- The problem requires evidence synthesis from multiple sources
- You need explicit citation traceability from claims to references
Invocation Scope
- This skill is standalone and user-invoked directly.
- v1 scope is direct invocation only.
- Do not couple this protocol to other routes, phases, or automatic invocation flows.
Inputs
Required:
- Research question (single sentence)
Optional:
- Context and constraints
- Time/depth budget
- Preferred source domains
- Output path
Output Contract
Produce exactly one paper-style markdown artifact (.md) with this mandatory structure:
# <Title>
## Abstract
## Research Question
## Scope and Constraints
## Method (Staged Protocol)
## Findings by Stage
### Stage 1 - Problem Framing
### Stage 2 - Source Discovery
### Stage 3 - Source Triage
### Stage 4 - Evidence Extraction
### Stage 5 - Synthesis and Limits
## Discussion
## Conclusion
## Limitations
## References
Rules:
- The report is a single durable artifact.
- Every non-trivial claim must include an in-text citation key, e.g.
[@smith2024llm].
References must be BibTeX-compatible entries embedded in markdown fenced code blocks.- Citation keys used in the body must exist in
References.
Staged Protocol
Follow stages in order and do not skip forward.
Stage 1: Problem Framing
- Normalize the question.
- Define explicit scope boundaries and out-of-scope points.
- Define success criteria for the investigation.
Exit condition:
- Problem statement and boundaries are unambiguous.
Stage 2: Source Discovery
- Collect candidate sources from diverse origins (specs, official docs, papers, benchmarks, incident reports when relevant).
- Capture URL, title, publisher/author, date, and source type.
Exit condition:
- Candidate set covers at least 3 independent sources or a justified smaller set if the domain is niche.
Stage 3: Source Triage
- Score each source for relevance, credibility, recency, and evidence type.
- Drop low-quality or redundant sources with explicit rationale.
Exit condition:
- Curated source set is sufficient to answer the question.
Stage 4: Evidence Extraction
- Extract claims, methods, metrics, and constraints from curated sources.
- Record direct evidence snippets with attribution.
- Distinguish fact vs interpretation.
Exit condition:
- Each major sub-question has evidence coverage.
Stage 5: Synthesis and Limits
- Reconcile agreements and conflicts across sources.
- State confidence per key conclusion (high/medium/low) with rationale.
- Explicitly document unresolved uncertainties and limitations.
Exit condition:
- Conclusions are traceable and uncertainty is explicit.
Citation Standard (BibTeX-Compatible)
In References, provide entries in this style:
@article{smith2024llm,
title={Example Title},
author={Smith, Alice and Doe, Bob},
journal={Journal Name},
year={2024},
url={https://example.org/paper}
}
Minimum fields by source type:
- Article/paper:
author, title, year, url (+ venue field when available)
- Website/doc page:
title, year (or note={Accessed: YYYY-MM-DD}), url
- Book/report:
author or institution, title, year, url or publisher metadata
Quality Gate Checklist
Before finishing, verify:
- Exactly one markdown report was produced
- All mandatory sections are present
- Every non-trivial claim has an in-text citation key
- Every citation key resolves to one BibTeX-compatible entry
- Broken or unreachable links are flagged in Limitations
- The report reflects staged progression, not an unstructured source dump
Non-Goals (v1)
- No automatic routing or scheduler integration
- No coupling to phase-specific workflows
- No generation of multiple report artifacts for one invocation
- No hidden side artifacts as required outputs
