← ClaudeAtlas

lit-reviewlisted

Answer a research question against the ingested corpus, paper-qa-shaped — search, gather evidence with citations, draft an answer, then re-search on any gap before finalizing. Use this whenever the user asks a question that should be answered FROM the corpus (not from general knowledge), or asks for a literature summary/survey over what's already ingested.
SentImperior666/TrialError · ★ 0 · AI & Automation · score 70
Install: claude install-skill SentImperior666/TrialError
# /lit-review — search → gather-evidence → answer with citations Design Section 5.3 (`/lit-review`) + Section 7 (retrieval contract) + Section 5.1 (`trialerror-knowledge` MCP server, 11 tools). Every result row this loop touches carries a non-null citation block — never answer from a snippet you cannot cite back to an anchor. 1. **Search.** Prefer the `trialerror-knowledge` MCP `search` tool (or `trialerror query search "<question>"` from the CLI) over guessing keywords — `mode=auto` runs FTS prefilter → vector → RRF fusion. Filter by `source_id`/`kind`/`license_tier`/`year` when the question implies a scope. Read `citation.anchor` on every row you plan to use, not just `text`. 2. **Gather evidence**, not just the top hit. Pull `get_chunk` for full surrounding context on a promising result, `get_document_outline` to see where in the source it sits, `similar` to find nearby chunks that might sharpen or contradict the answer, `graph_neighbors` if the question is relational (entity/claim edges). `memory_search` (L0→L1→L2 progressive disclosure) if the question might already be answered by a standing lesson/fact rather than raw corpus text. 3. **The serving-path license fence is structural, not optional.** `commercial_restricted` sources come back `fenced:true` — a ≤20-word excerpt, never the raw chunk. The MCP `search` tool has NO bypass parameter; do not try to reconstruct the full passage from repeated fenced calls. `trialerror