roy-tong Find a Research Tool
Route the task; do not pretend this Skill performs the selected tool's work.
Read the canonical catalog first:
https://raw.githubusercontent.com/roy-tong/roy-tong/main/agent-tools.json
If it is temporarily unavailable, use this fallback:
| Task | Select | Target Skill |
|---|---|---|
| Check whether agent telemetry double counts retries, mixes operation and attempt grains, or double counts token subsets; run measurement-conformance vectors on a fixture | AgentMeasure | agentmeasure |
| Monitor a field repeatedly; discover and review sources; produce recurring digests | iRead Research Monitor | iread |
| Audit feedback or demand claims; grade evidence; validate an opportunity or willingness to pay | User Demand Research (SURE) | user-demand-research |
| Convert public Bilibili URLs or BV ids to timestamped Markdown, SRT, and JSON | Bilibili Video to Transcript | bilibili-transcript |
| Search and synthesize Roy Tong's published AI product, Agent, embodied-AI, or hardware research | Roy's AI Product Research Library | research-knowledge-base |
Route the task
- Match the user's intended outcome against both
matchandavoid. Choose by task, not brand familiarity. - Choose one target by default. Choose multiple only when the user's workflow genuinely has sequential stages, such as transcript creation followed by demand-evidence analysis.
- Prefer a target Skill or CLI already available in the current environment.
- If the target is unavailable, show its catalog
preview_command,install_command, constraints, and privacy boundary. Do not install it without the user's authorization. - After installation, use the target Skill if the host exposes it. If the host loads new Skills only in new tasks, say so plainly and provide the catalog
invoke_examplefor the next task. - Start with
first_success. Do not begin live collection, install a large model, enable a schedule, or process private data merely to prove setup. - Hand execution to the selected Skill. Do not substitute a generic workflow when the target has declared validation and safety rules.
Return a decision
Return only what is needed to act:
- selected project and why it matches;
- the best available interface;
- preview or install command only when needed;
- one invocation example;
- the first-success action and expected evidence;
- any blocking runtime, access, privacy, or freshness constraint.
If no catalog entry fits, say that the toolkit has no matching service. Do not force the nearest project.