Research Spike
Time-boxed technical investigation with structured, citation-backed output.
Use for a spike before adopting unfamiliar technology, an architecture decision
record, a comparison of tools or libraries, a feasibility study, a migration or
upgrade assessment, or a survey of best practice in a domain.
Modes
| Mode |
Output emphasizes |
| Technical spike |
Practical findings, code examples, implementation recommendations |
| Competitive research |
Feature matrices, trade-off analysis, selection criteria |
| Feasibility analysis |
Constraints, risks, effort factors, go/no-go recommendation |
Upfront Questions
Use AskUserQuestion to settle three things before researching anything:
- Mode — spike, competitive, or feasibility.
- Depth — quick (high-level overview), standard (actionable depth), or deep
dive (comprehensive, production-ready guidance).
- Scope — technologies to include or exclude, constraints (budget, team
expertise, existing stack), and the decision factors that define success.
Method
- Scope — confirm the question and its boundaries, the key terms, and the
known constraints.
- Plan — choose sources (web, documentation, the codebase), split into
sub-questions, and prioritize by the user's decision factors.
- Gather — search and read; explore the codebase where relevant. Record
every source as you go, not afterwards.
- Synthesize — organize findings by theme, identify patterns and
trade-offs, derive recommendations from the evidence, and name the gaps.
- Write — produce the report, then check it against the quality gates.
Anti-Hallucination
- State only what a source explicitly confirms.
- Hedge inferences: "suggests", "likely", "based on X, we can infer".
- Keep opinion visibly distinct from fact.
- Label anything you cannot verify as "unverified" or "needs confirmation".
- When comparing options, include only documented features.
Output
Write to research/[topic-slug]/spike-report.md, progressively as you work
rather than in one pass at the end. Read
references/report-template.md for the section
template and the quality gates to verify before finishing. Close by telling the
user where the file is and what it concluded.
1---2name: research-spike3description: Runs a time-boxed spike, tool comparison, or feasibility study and writes a cited report. Use for architecture decisions.4---56# Research Spike78Time-boxed technical investigation with structured, citation-backed output.910Use for a spike before adopting unfamiliar technology, an architecture decision11record, a comparison of tools or libraries, a feasibility study, a migration or12upgrade assessment, or a survey of best practice in a domain.1314## Modes1516| Mode | Output emphasizes |17| --- | --- |18| Technical spike | Practical findings, code examples, implementation recommendations |19| Competitive research | Feature matrices, trade-off analysis, selection criteria |20| Feasibility analysis | Constraints, risks, effort factors, go/no-go recommendation |2122## Upfront Questions2324Use AskUserQuestion to settle three things before researching anything:25261. **Mode** — spike, competitive, or feasibility.272. **Depth** — quick (high-level overview), standard (actionable depth), or deep28 dive (comprehensive, production-ready guidance).293. **Scope** — technologies to include or exclude, constraints (budget, team30 expertise, existing stack), and the decision factors that define success.3132## Method33341. **Scope** — confirm the question and its boundaries, the key terms, and the35 known constraints.362. **Plan** — choose sources (web, documentation, the codebase), split into37 sub-questions, and prioritize by the user's decision factors.383. **Gather** — search and read; explore the codebase where relevant. Record39 every source as you go, not afterwards.404. **Synthesize** — organize findings by theme, identify patterns and41 trade-offs, derive recommendations from the evidence, and name the gaps.425. **Write** — produce the report, then check it against the quality gates.4344## Anti-Hallucination4546- State only what a source explicitly confirms.47- Hedge inferences: "suggests", "likely", "based on X, we can infer".48- Keep opinion visibly distinct from fact.49- Label anything you cannot verify as "unverified" or "needs confirmation".50- When comparing options, include only documented features.5152## Output5354Write to `research/[topic-slug]/spike-report.md`, progressively as you work55rather than in one pass at the end. Read56[references/report-template.md](references/report-template.md) for the section57template and the quality gates to verify before finishing. Close by telling the58user where the file is and what it concluded.
gpu-cli