Research Powerpack MCP

# 03 — Scrape LLM Wrapper Audit (`scrape_pages`)

## Metadata

| Field | Value |
|---|---|
| Canonical tool | `scrape_pages` |
| Legacy alias | `scrape_links` -> `scrape_pages` |
| Source | `src/services/llm-processor.ts` (`processContentWithLLM`) |
| Trigger | Per URL when `use_llm=true` |
| Role | Wraps extraction instruction + page content before LLM call |

## Current Wrapper Text (Verbatim)

**Branch A (`what_to_extract` provided):**

```text
Extract and clean the following content. Focus on: {config.what_to_extract}

Content:
{truncatedContent}
```

**Branch B (`what_to_extract` missing):**

```text
Clean and extract the main content from the following text, removing navigation, ads, and irrelevant elements:

{truncatedContent}
```

## Context in Prompt Assembly

1. `src/tools/scrape.ts` builds `enhancedInstruction` (prefix + user target + suffix).
2. `processContentWithLLM` wraps that instruction with this frame.
3. Final request is currently user-role only (no system-role constraints).

## Criticism Table (12)

| # | Criticism | Impact |
|---:|---|---|
| 1 | No system-role for fixed rules | Lower instruction priority |
| 2 | Wrapper repeats "extract/clean" semantics already in prefix | Token waste |
| 3 | Branch A vs B has different policy surfaces | Inconsistent outputs |
| 4 | `Focus on:` is vague boundary text | Weaker controllability |
| 5 | No explicit anti-preamble rule here | Fluff drift possible |
| 6 | No explicit output-shape contract | Parsing variability |
| 7 | Generic `Content:` delimiter only | Instruction/source bleed risk |
| 8 | Branch B bypasses enriched extraction targets | Lower relevance |
| 9 | Fixed overhead repeats per URL | Scales cost linearly |
| 10 | Wrapper language is verbose relative to control value | Poor density |
| 11 | No confidence/uncertainty output guidance | Inconsistent certainty handling |
| 12 | Not normalized with `deep_research` system+user split style | Cross-tool inconsistency |

## Recommended Wrapper (copy-paste design)

```text
SYSTEM: Extract only from SOURCE. No hallucination. No preamble. Structured data -> markdown table; otherwise tight bullets.
USER: TARGETS:\n{what_to_extract_or_default}\n\nSOURCE:\n{truncatedContent}
```

## Alternatives (3)

| Alternative | Pros | Cons |
|---|---|---|
| A — **Recommended** system+user split | Best adherence, clear precedence, lower redundancy | Requires message-shape refactor |
| B — Single-user compact frame | Minimal code change | Weaker policy priority |
| C — Keep current, trim wording | Safest rollout | Retains structural weaknesses |

## System vs User Tradeoff

- Put stable policy in **system** (groundedness, format, anti-fluff).
- Put variable intent + source in **user**.
- Current design blends both, reducing consistency.