badmuriss Research
Match the evidence workflow to the question:
- Lookup: open the relevant primary source, verify the requested fact and cite it with its access date. No report file, credit ledger or graph is required.
- Synthesis or report: compare the relevant sources, distinguish evidence from inference and name unresolved disagreements. For a requested durable report or consequential research, use reporting.md.
- High stakes: use current primary evidence and proportionate independent review. Never convert missing evidence into a confident conclusion.
Reuse repository material when it already answers the question. A number from a local check does not require web research. Verify volatile external values at the source; record when they were accessed. Treat snippets and agent summaries as leads, not as sources already read.
For public web collection, use the installed scrapingdog skill when keyed.
After a bounded failure or missing key, record why and use Firecrawl, then host
search. For a known official document, direct reading is sufficient. For papers,
start with the free Firecrawl Research Index when available; use paper-search to
cross-check metadata when needed. Missing optional providers do not block local
work that already has adequate evidence.
Convert documents only when needed to extract their content. Verify reading order and tables. Do not package a whole repository to inspect a known file.
Delegate collection only when independent source work would materially help. Give collectors narrow questions and return paths or URLs; inspect the sources before accepting claims. Collection alone does not require an Agent Graph.
Report what is established, what is inference and what remains unverified. A small or uncontrolled sample does not establish a general productivity claim.
Adapted from research-stack, MIT.