wimpysworld/nix-config/research-task

Research Task

Deeply research an existing task, everything it links to, and related work, then recommend cited solutions. The caller supplies the task target and optional context. This skill supplies the method.

Use research-task when a tracked task anchors the work. Defer an open question with no tracked task to deep-research.

Keep every source read-only. For GitHub issue, pull request, code, comment, review, and status reads or searches, follow the global GitHub rule: prefer dedicated gh reads and searches, then gh-api-safe; otherwise use only documented, clearly read-only GitHub MCP operations. Never use a GitHub MCP mutation or a tool whose effect is unclear.

Input resolution

Resolve the task target before researching:

Treat text after the target as extra context. If the target is missing, ask which task to research before starting.

Process

1. Apply communication-rules before writing anything. Read it first unless its complete, current instructions are already in this context.
2. Read the target and extract its problem statement, acceptance criteria, and every embedded or linked source.
3. As the coordinator, fan out to workers in parallel where possible. As a worker, research the assigned scope directly. Cover both linked sources and unlinked related work within that scope.
4. Merge the findings into the output format. Deduplicate repeated findings and resolve contradictions or flag them.

A task written by create-task carries its main context under Outcome, Problem, Context, Scope, Requirements, Acceptance criteria, Validation, Non-goals, Dependencies, and Evidence.

Fan-out

Split work by source or source cluster. Match the available tool to the source:

Reach for BigQuery when the task makes a quantitative claim, when a recommendation would be stronger with a number, or when the size of a problem is unknown. Prefer a measured number over an estimate. Record the query and run date beside the number. If no client is reachable, state that limitation and continue. Absence is not a finding.

Derive search terms from the problem statement, feature names, symbols, and error strings in the target. For each related item, capture what it decided or changed and whether it is open, merged, or abandoned. Use this context to respect and reuse prior work.

Each worker returns findings with source references. Never mutate external state: no comments, approvals, merges, or posts.

Only the coordinator plans and dispatches work. Loading this skill does not change a worker's role. Workers never launch agents. Preserve source coverage, topic searches, measurements, and citations within the assigned scope. If additional specialist work is necessary, return a bounded request to the parent after completing independent assigned work. The coordinator handles the request and continues the original task.

Output

# [Task title] - research
 
## Summary
[Problem and headline recommendation - one paragraph]
 
## Task context
[What the target asks for: problem, acceptance criteria, scope]
 
## Linked sources researched
[Each source with a one-line note on what it contributed]
 
## Related work
[Prior contributions found by topic search on Linear, GitHub, and Slack, not linked from the target. For each: what it decided or changed, its status, and what to respect or reuse rather than undo. Omit if none]
 
## Findings
### [Theme]
[Synthesised findings with source references]
 
## Recommended solutions
| Option | Approach | Trade-offs | → |
| --- | --- | --- | --- |
[Ranked options; mark the recommendation with →]
 
## Open questions
[Unresolved items - omit if none]
 
## Next steps
[Concrete actions]

This skill does not update the task or tracker automatically. create-task or update-task captures the research when authorised. Private report storage follows the inherited delivery rules. Research consumed directly by a later workflow needs no intermediate report file.

Markers: 📌 KEY (critical finding), ⚠️ CAVEAT (limitation or uncertainty), → (recommendation).

Constraints

• Every claim traces to a source. Never present training data as a research finding.
• Search Linear, GitHub, and Slack by topic, not only through links from the task.
• Note prior contributions to respect. Do not recommend undoing valuable work without saying why.
• Measure quantitative claims through a reachable data source and cite the query and run date.
• Deduplicate findings raised by more than one worker.
• Never mutate external state.
• Omit empty sections.
• State each fact once and lead with conclusions.