Feedback: All deep research goes through Perplexity

Date: 2026-05-09

Source: Harnoor (explicit, durable rule)

Confidence: explicit · durable

Rule

Optimize for cost. Two kinds of "research":

1. Synthesis / discovery ("what's the latest on X", "compare A vs B", "survey of papers") — use Perplexity (pplx.py) first. Harnoor has Perplexity Pro = flat rate, returns synthesized cited answers, doesn't burn Claude tokens.

2. Raw scraping / known-URL fetch (pull a specific page's HTML, parse JSON from a public API, scrape a Reddit thread, hit an arXiv endpoint) — use local Python tools: urllib, requests, BeautifulSoup, feedparser, subprocess curl. Run via Bash tool. NOT WebFetch. NOT Claude. Local processing keeps Anthropic token costs at zero for grunt work.

Never use WebFetch for bulk scraping — every fetch dumps 5-50K tokens of HTML through Claude's context. Local Python parses it for free.

WebFetch is only acceptable when: Harnoor specifically pasted the URL and wants Claude to read+summarize it conversationally, or for a single one-off page where setting up a local scraper is overkill.

Why

Perplexity Pro returns higher-quality, cited, recent answers vs WebSearch
Harnoor pays for it; not using it wastes the subscription
WebSearch/WebFetch return raw HTML that needs more processing; Perplexity returns synthesized + cited

How to apply

Use the CLI wrapper: python F:/TITAN/scripts/pplx.py "your query" --model sonar-pro --recency week --max-tokens 1200

- --model: sonar-pro (default), sonar-reasoning, sonar-deep-research for thesis-grade

- --recency: day, week, month, year, or omit for all-time

- --max-tokens: 1200 default; use 3000+ for deep memos

Use first before falling back to WebSearch / WebFetch
WebFetch is still appropriate for: known specific URL Harnoor gave you, GitHub README/source code, a specific blog post linked from a Perplexity citation
WebSearch acceptable as a secondary check or for queries Perplexity doesn't handle well

Agent-by-agent

| Agent | Pattern |

|---|---|

| ORACLE | Already uses Perplexity — no change |

| SCOUT | Use pplx.py first for any research synthesis. WebSearch + WebFetch only as fallback / specific URL fetch |

| forge / general-purpose | When delegated a research-heavy task, pplx.py first |

| /feed | Already uses Perplexity (per skill description) |

Cost note

Perplexity Pro subscription = flat-rate, no per-query cost. Use freely for legitimate research. Logged in nightly token-audit (F:/TITAN/scripts/nightly_report_builder.py already tracks Perplexity spend).

Don't

Don't loop pplx.py 50 times for cheap follow-ups; batch the query
Don't bypass and use WebSearch when the query is open-ended ("what's the latest on memory architectures") — that's a Perplexity query