| name | parallel-search-patterns |
| description | Web-search and web-fetch playbook for the @open-neko/plugin-parallel-search plugin. Use when the agent needs to look up external context (competitor news, market signals, company facts) for a finding or Ask thread. Covers when to use web_search vs web_fetch, query shaping, depth limits, and how to thread Parallel results into a Briefing-style summary. |
| license | Apache-2.0 |
| metadata | {"authoredBy":"open-neko","pairsWith":"@open-neko/plugin-parallel-search"} |
Parallel-search patterns
The @open-neko/plugin-parallel-search plugin exposes two sandboxed
actions backed by Parallel.ai's Search MCP:
| Action | Use when |
|---|
web_search | The operator wants a synthesized answer ("what's the latest funding round for $COMPETITOR?"). |
web_fetch | The operator already has a URL ("read me this filing"). |
Both have default_mode: auto — they're read-only against the open
web and don't need a human gate.
Query shaping for web_search
Parallel's search ranks higher on specific queries than generic
ones. Default to:
- Include time markers when the operator cares about recency:
"Q4 2025 earnings", "announced this week". Without one,
Parallel may surface a year-old article that looks current.
- Name entities precisely.
"Anthropic Claude Opus 4.6 release"
beats "new Claude version". If the operator was vague, ask one
clarifying question before searching.
- Cap depth at 2 hops. Don't chain
web_search → web_fetch →
web_search more than two levels deep — that's an indication the
first search was wrong, not that the topic is rich. Re-ask the
operator instead.
When web_fetch is the right call
- The operator pasted a URL in the message.
- A previous
web_search returned a single dominant result and the
next question is "what does that article actually say?".
- You need a specific document (filing, blog post, transcript) that
the operator referenced by title — pair
web_search to find the
URL, then web_fetch to pull the body.
Don't double-fetch
Each web_fetch is a billable Parallel call. If you already fetched
a URL in this turn, reuse the result — re-fetching wastes spend AND
makes the response slower.
Threading results into operator-facing output
Parallel returns markdown with citations. The agent should:
- Quote sparingly — operators want the synthesized answer, not the
raw page.
- Attribute every claim ("per the SEC 8-K filed 2026-03-04, …").
Operators in finance/compliance roles need this; everyone else
tolerates it.
- Strip Parallel's verbose preambles ("The search returned…"). Get to
the answer.
Failure modes the operator should see
PARALLEL_API_KEY not set — operator runs openneko secrets set @open-neko/plugin-parallel-search PARALLEL_API_KEY <key>.
rate_limited — surface the wait if >5s; otherwise just retry.
- No results / low confidence — say so plainly ("I couldn't find a
reliable source for this"). Don't fabricate.
What this skill is NOT for
- Internal knowledge — that's the GraphJin data source.
- Long-running research projects — Parallel's
deep-research flow is
not in this plugin yet; for now, multi-step research means multiple
shallow web_search calls.