Last30daysCAUTION

Vigolium - High-fidelity vulnerability scanner fusing agentic AI with native speed, modularity, and precision

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: last30days
description: "Researches a topic from the last 30 days across the web, Reddit, and X, surfacing real discussions with engagement metrics and synthesizing them into actionable insights. In an audit, use it for recent domain attack research — newly disclosed CVEs, fresh exploit techniques, and attack patterns for the target's stack/framework that predate the model's knowledge cutoff. Use whenever current, time-sensitive intelligence beats static knowledge."
argument-hint: 'recent CVEs in <framework>, new SSRF bypass techniques, latest <language> deserialization exploits'
allowed-tools:
  - Bash
  - Read
  - Write
  - AskUserQuestion
  - WebSearch
  - mcp__exa__web_search_exa
---

# last30days: Research Any Topic from the Last 30 Days

Research ANY topic across Reddit, X, and the web. Surface what people are actually discussing, recommending, and debating right now.

## When to Use

- User wants to know what people are saying about a topic in the last 30 days
- User asks for recent community sentiment, recommendations, or trending discussions
- User wants to research a topic before making a decision (tools, products, techniques)
- User asks "what's new with X" or "best Y" or "what are people saying about Z"

## When NOT to Use

- User wants historical data older than 30 days
- User needs official documentation or API references (use docs tools instead)
- User wants to search a specific codebase (use code search tools)
- User asks a factual question that doesn't need community sentiment

## CRITICAL: Parse User Intent

Before doing anything, parse the user's input for:

1. **TOPIC**: What they want to learn about (e.g., "web app mockups", "Claude Code skills", "image generation")
2. **TARGET TOOL** (if specified): Where they'll use the prompts (e.g., "Nano Banana Pro", "ChatGPT", "Midjourney")
3. **QUERY TYPE**: What kind of research they want:
   - **PROMPTING** - "X prompts", "prompting for X", "X best practices" -> User wants to learn techniques and get copy-paste prompts
   - **RECOMMENDATIONS** - "best X", "top X", "what X should I use", "recommended X" -> User wants a LIST of specific things
   - **NEWS** - "what's happening with X", "X news", "latest on X" -> User wants current events/updates
   - **GENERAL** - anything else -> User wants broad understanding of the topic

Common patterns:
- `[topic] for [tool]` -> "web mockups for Nano Banana Pro" -> TOOL IS SPECIFIED
- `[topic] prompts for [tool]` -> "UI design prompts for Midjourney" -> TOOL IS SPECIFIED
- Just `[topic]` -> "iOS design mockups" -> TOOL NOT SPECIFIED, that's OK
- "best [topic]" or "top [topic]" -> QUERY_TYPE = RECOMMENDATIONS
- "what are the best [topic]" -> QUERY_TYPE = RECOMMENDATIONS

**IMPORTANT: Do NOT ask about target tool before research.**
- If tool is specified in the query, use it
- If tool is NOT specified, run research first, then ask AFTER showing results

**Store these variables:**
- `TOPIC = [extracted topic]`
- `TARGET_TOOL = [extracted tool, or "unknown" if not specified]`
- `QUERY_TYPE = [RECOMMENDATIONS | NEWS | HOW-TO | GENERAL]`

**DISPLAY your parsing to the user.** Before running any tools, output:

```
I'll research {TOPIC} across Reddit, X, and the web to find what's been discussed in the last 30 days.

Parsed intent:
- TOPIC = {TOPIC}
- TARGET_TOOL = {TARGET_TOOL or "unknown"}
- QUERY_TYPE = {QUERY_TYPE}

Starting now.
```

If TARGET_TOOL is known, mention it in the intro: "...to find {QUERY_TYPE}-style content for use in {TARGET_TOOL}."

This text MUST appear before you call any tools. It confirms to the user that you understood their request.

---

## Research Execution

**Step 1: Run the research script**
```bash
python3 "{baseDir}/scripts/last30days.py" "$ARGUMENTS" --emit=compact 2>&1
```

The script will automatically:
- Detect available API keys
- Run Reddit/X searches if keys exist
- Signal if WebSearch is needed

---

## STEP 2: DO WEBSEARCH WHILE SCRIPT RUNS

The script auto-detects sources (API keys, etc). While waiting for it, do WebSearch.

For **ALL modes**, do WebSearch to supplement (or provide all data in web-only mode).

Choose search queries based on QUERY_TYPE:

**If RECOMMENDATIONS** ("best X", "top X", "what X should I use"):
- Search for: `best {TOPIC} recommendations`
- Search for: `{TOPIC} list examples`
- Search for: `most popular {TOPIC}`
- Goal: Find SPECIFIC NAMES of things, not generic advice

**If NEWS** ("what's happening with X", "X news"):
- Search for: `{TOPIC} news 2026`
- Search for: `{TOPIC} announcement update`
- Goal: Find current events and recent developments

**If PROMPTING** ("X prompts", "prompting for X"):
- Search for: `{TOPIC} prompts examples 2026`
- Search for: `{TOPIC} techniques tips`
- Goal: Find prompting techniques and examples to create copy-paste prompts

**If GENERAL** (default):
- Search for: `{TOPIC} 2026`
- Search for: `{TOPIC} discussion`
- Goal: Find what people are actually saying

For ALL query types:
- **USE THE USER'S EXACT TERMINOLOGY** - don't substitute or add tech names based on your knowledge
- EXCLUDE reddit.com, x.com, twitter.com (covered by script)
- INCLUDE: blogs, tutorials, docs, news, GitHub repos
- **DO NOT output "Sources:" list** - this is noise, we'll show stats at the end

**Options** (passed through from user's command):
- `--days=N` -> Look back N days instead of 30 (e.g., `--days=7` for weekly roundup)
- `--quick` -> Faster, fewer sources (8-12 each)
- (default) -> Balanced (20-30 each)
- `--deep` -> Comprehensive (50-70 Reddit, 40-60 X)

---

## Judge Agent: Synthesize All Sources

**After all searches complete, internally synthesize (don't display stats yet):**

The Judge Agent must:
1. Weight Reddit/X sources HIGHER (they have engagement signals: upvotes, likes)
2. Weight WebSearch sources LOWER (no engagement data)
3. Identify patterns that appear across ALL three sources (strongest signals)
4. Note any contradictions between sources
5. Extract the top 3-5 actionable insights

**D

CAUTIONgrade C · trust 79/100 Install with care. The audit found things worth knowing before you trust its output.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (2)

**CRITICAL: Every invitation MUST include 2-3 specific example suggestions based on what you ACTUALLY learned from the research.** Don't be generic -- show the user you absorbed the content by referen

AGENTS.md

Gates applied: instruction_override, no_behavioural_pass.