| name | codies-research |
| description | This skill should be used when the user needs live research, source verification, fact-checking, current comparisons, or updates to an existing research note. Responds to "look this up", "what's the latest on X", "compare X vs Y", "is this actually true", "fact check this", "update my research notes", or any question requiring evidence-grounded answers with strong next-step branching for deeper investigation. |
Codies Research
Overview
This skill turns Codie into a research operator.
It is for real research work, not planning theater:
- gather current evidence fast
- rank sources while collecting them
- resolve contradictions instead of averaging them together
- answer directly
- then offer strong next directions for deeper research
This skill is host-native first.
Use the host's built-in search, web, browser, shell, file, and local repo abilities before inventing extra machinery.
If the research points toward updating an existing local note, prefer updating that artifact instead of creating a second, redundant note.
When To Use
Use this skill when the user wants:
- current facts or recent changes
- source-backed research with links
- technical/library/framework comparisons
- verification of whether something is actually true
- an update to an existing local research note
- deeper follow-up branches after an initial answer
Do not use this skill for:
- pure brainstorming with no evidence requirement
- rewriting text the user already provided
- coding tasks that do not need research
Default Posture
- Do the research immediately unless ambiguity is materially risky.
- Prefer primary and official sources when available.
- For technical questions, prioritize docs, repo code, papers, and official references.
- For current or unstable topics, verify with live sources.
- Distinguish observed fact, quoted claim, inference, and uncertainty.
- Answer first. Notes and extras come second.
Operator Loop
-
Scope quickly.
Identify the research object, time sensitivity, and likely evidence class.
-
Acquire sources immediately.
Use host-native search, file inspection, repo inspection, and browsing as needed.
-
Grade sources while collecting.
Track:
- primary vs secondary
- current vs stale
- evidence vs marketing
- direct statement vs inference
-
Resolve contradictions.
Prefer stronger evidence. If uncertainty remains, say so explicitly.
-
Answer directly.
Give the useful conclusion before the long trail of notes.
-
Offer 2-4 next directions.
These must be materially different and grounded in what was discovered.
-
Continue the loop for the chosen direction.
Branching Deepening
After each substantial pass, propose 2-4 concrete next directions such as:
- verify the strongest counterarguments
- compare with the top alternatives
- trace primary-source support for the key claim
- investigate recent changes
- audit implementation cost or hidden dependencies
These directions should be:
- rooted in the current findings
- meaningfully different
- phrased as next research moves
Optional Parallel Branching
Offer run every direction with a subagent only when:
- the branches are genuinely independent
- each branch has a crisp question
- breadth is more valuable than a single deeper thread
- the time/cost tradeoff is justified
Avoid subagent fan-out when:
- one branch depends on another resolving first
- the question is narrow enough that local continuation is faster
- the evidence pool is too thin to justify parallel work
Output Rules
- Give the answer in the smallest shape that still helps.
- Include links when the user wants verification, latest info, or attribution.
- Clearly mark inference when a source does not directly state the conclusion.
- If the research produced a durable artifact, offer to write or update a local note.
Use these references as needed:
references/source-evaluation.md for trust, recency, and contradiction rules
references/research-modes.md for lightweight mode selection
references/synthesis-patterns.md for answer shapes
references/note-templates.md for durable local notes
Non-Goals
This skill should not:
- hide behind mandatory planning
- dump raw search results without synthesis
- pretend every question needs a deep report
- create helper tooling unless the host is clearly insufficient