Problem
The model-invoked research skill can be selected for ordinary lookups that do not need its defining workflow.
Its current description includes:
Use when the user wants a topic researched, docs or API facts gathered, or reading legwork delegated to a background agent.
The phrases “topic researched” and “docs or API facts gathered” are broad enough to match routine fact lookups, product/spec checks, and short API comparisons. Once selected, however, the skill requires a heavier and materially different outcome: delegate work to a background agent and leave a cited Markdown file in the repository.
This creates a mismatch between discovery and execution. The model correctly notices that information gathering is needed, but cannot tell from the description whether the user wants a direct answer or a durable research artifact.
Reproduction
Ask a model with the skill installed questions such as:
- “Check whether this product supports feature X.”
- “Compare these two API options.”
- “Confirm what the official documentation says about this parameter.”
The research description makes these plausible automatic matches even when the user only wants a concise answer in the current thread and has not asked for a repository file or background delegation.
Expected behavior
Keep research model-invoked and composable with wayfinder and ask-matt, but select it when at least one defining property applies:
- substantial reading legwork should be delegated to a background agent;
- the user wants a durable, cited research note saved in the repository;
- another skill explicitly delegates a research ticket that requires that artifact.
Ordinary source lookup should remain in the active thread without invoking this skill.
Suggested description
description: Delegate substantial reading legwork to a background agent and save the findings as a cited Markdown file in the repository. Use when the user wants a durable research artifact, wants a substantial investigation performed in the background, or another skill delegates a research ticket. Do not use for ordinary fact lookups, brief product or API comparisons, current spec checks, or questions that can be answered directly in the active thread without creating a repository file.
The exact wording can be shorter, but making the durable artifact and delegation requirements part of discovery should reduce accidental invocation while preserving the intended wayfinder composition.
Why not make it explicit-only?
The current automatic and cross-skill invocation model is useful. The issue is not that the model can invoke research; it is that the discovery description currently treats most information-gathering tasks as candidates.
Problem
The model-invoked
researchskill can be selected for ordinary lookups that do not need its defining workflow.Its current description includes:
The phrases “topic researched” and “docs or API facts gathered” are broad enough to match routine fact lookups, product/spec checks, and short API comparisons. Once selected, however, the skill requires a heavier and materially different outcome: delegate work to a background agent and leave a cited Markdown file in the repository.
This creates a mismatch between discovery and execution. The model correctly notices that information gathering is needed, but cannot tell from the description whether the user wants a direct answer or a durable research artifact.
Reproduction
Ask a model with the skill installed questions such as:
The
researchdescription makes these plausible automatic matches even when the user only wants a concise answer in the current thread and has not asked for a repository file or background delegation.Expected behavior
Keep
researchmodel-invoked and composable withwayfinderandask-matt, but select it when at least one defining property applies:Ordinary source lookup should remain in the active thread without invoking this skill.
Suggested description
The exact wording can be shorter, but making the durable artifact and delegation requirements part of discovery should reduce accidental invocation while preserving the intended
wayfindercomposition.Why not make it explicit-only?
The current automatic and cross-skill invocation model is useful. The issue is not that the model can invoke
research; it is that the discovery description currently treats most information-gathering tasks as candidates.