parallel-research

Parallel Research

Run external research with Parallel.ai and write back a cited report. Use this when a coding or product question needs public-source research: prior art, standards, architecture choices, implementation techniques, thresholds, or trade-offs.

The bundled runner is scripts/run.mjs. It submits an already-framed query to Parallel, polls to completion, and writes a Markdown report. The skill owns the workflow and safety checks; the runner owns the API call.

Before You Call Out

External research is a paid call to a third party, and the submitted query is committed verbatim into the report. Never send client data, secrets, credentials, private customer material, nonpublic business plans, or unpublished case facts to Parallel. Research public methods and public facts in the abstract; apply them to private data locally.

When another workflow or indirect request would trigger research, confirm before making the call unless the user explicitly asked for research this turn.

Frame The Question

State the question precisely and ask for specific, cited findings. Name the figures, thresholds, standards, tools, papers, or implementation approaches you want back. Frame it neutrally: ask for evidence across the field, not confirmation of a conclusion you already hold. The recommendation is yours to draw from the findings afterward, not the engine's to hand you.

Good research prompts include:

• The decision or question being investigated.
• The exact output shape needed, such as a comparison table, risk list, implementation guide, or cited recommendation.
• The public-source boundaries, especially what not to include.
• Any freshness requirement.

Choose Depth

Depth is a property of the question. A quick prior-art scan can use lite, base, or core; a broad architecture or standards question often needs pro; a novel or high-stakes methodology question may justify an ultra* tier.

The runner defaults to pro. Top tiers (ultra2x, ultra4x, ultra8x, and -fast variants) can cost materially more per task, so confirm before using them unless the user explicitly asked for a top-tier run.