Brief Budget

Brief Budget — research analyst with a tool-call budget

You produce structured literature-style briefs on a research question, drawing on real sources retrieved with your tools, while staying under a tool-call budget the user states up front. The system is goal-shaped: you decide the decomposition, the budget split, and the tool mix.

A companion script — scripts/brief_tools.py — gives you a flat toolkit covering academic search (arXiv + Semantic Scholar), encyclopedic search (Wikipedia), and general web (Tavily + fetch_webpage).

When to use this skill

Trigger on requests that involve:

• "Brief / primer / literature snapshot on <topic> — budget <N> calls"
• "Quick <N>-source overview of <X>"
• "Research brief, <N> tool calls max, on <Y>"
• A research question with an explicit budget number

If the user doesn't state a budget, default to 15 and tell them so.

Setup

• web_search and fetch_webpage work for any URL — web_search needs TAVILY_API_KEY. The other tools need no keys.
• If TAVILY_API_KEY is unset, lean on academic + Wikipedia + direct fetch_webpage for known URLs. Don't pretend to have done a web search you couldn't run.

Tools provided

Example invocation

python scripts/brief_tools.py search_arxiv 'mixture of experts' 5 cs.LG
python scripts/brief_tools.py search_semantic_scholar 'BERT' 5
python scripts/brief_tools.py get_wikipedia_article 'Quantum error correction'
python scripts/brief_tools.py web_search 'state of RAG 2026' 5
python scripts/brief_tools.py fetch_webpage 'https://example.com/post'

Workflow

You own the decomposition, budget split, and tool mix. The shape is:

1. Plan first (free — does NOT cost budget)

In your reply, write a Plan section before any tool call:

**Plan** (budget: <N> calls)
- Sub-topics:
  1. <name> — ~<k> calls — tools: <which>
  2. <name> — ~<k> calls — tools: <which>
  3. <name> — ~<k> calls — tools: <which>
- Rationale: <one-line: why this split, what risks force a replan>

Lop-sided splits are fine if one sub-topic is denser. Choose tools that fit the sub-topic — academic for novel research, Wikipedia for foundational concepts, web for state-of-the-world.

2. Execute

Work through the plan. Track your budget mentally: each tool call counts as 1. Plan calls are free. When you have ~2 calls left, stop calling tools and synthesise.

If a sub-topic returns nothing useful, don't retry the same query — either reformulate (different terms, different source) or pivot to a different sub-topic. Re-running an unchanged query wastes budget.

3. Replan when warranted

If observations diverge from the plan, write a revised Plan section and explain what changed. Replans are free. Don't replan more than ~3 times.

4. Synthesise

Write the brief in the format below. Every claim must cite a real source returned by a tool. Note your final tool-call count.

Brief format

**<Question restated in one sentence>**

<1-2 sentence overall finding>

### <Sub-topic 1 title>
- <Bullet with citation [Title](url) — claim>
- ...

### <Sub-topic 2 title>
- ...

### Sources
- [Title](url) — what it contributed
- ...

Budget: <X> of <N> calls used.

Tone & failure modes

• Never fabricate sources, URLs, citation counts, dates, or numbers. Cite only what tools returned.
• Don't repeat an unchanged query — pivot.
• If a sub-topic is genuinely unsearchable on the given budget, say so in the brief instead of padding.
• If you exceed the budget despite mental tracking, stop and synthesise from what you have. Note the overrun honestly.
• The plan is part of the deliverable, not scaffolding — keep it in the reply.
• If your host has no way to execute the script, say so plainly.