lit-rescue

name: lit-rescue description: Last-resort skill. Invoke when no obvious or coherent solution is available and hallucination risk is high. Searches peer-reviewed literature and validated sources (Perplexity, bioRxiv, PubMed) before attempting an answer. Generalist — applies to any domain.

Lit-Rescue — Literature-Grounded Problem Solving

When to invoke this skill:

• No confident answer exists from training knowledge
• The problem involves a specific algorithm, parameter, protocol, or library behavior that could be misremembered
• Previous attempts at solving the problem have been inconsistent or failed
• The question is novel, niche, or at the edge of a domain
• The user is debugging something that "shouldn't happen"

Rule: If there is >20% chance the answer could be fabricated or outdated → invoke this skill first.

Step 1 — Classify the Problem

Identify the query type before formulating search queries:

Multiple types can apply — activate each relevant search.

Step 2 — Formulate Queries

Load references/perplexity-prompts.md if Perplexity MCP is available.

Otherwise, construct queries following these principles:

1. Be maximally specific — include library name + version, domain, and exact task
2. Ask for citations — request DOIs, paper titles, or package documentation references
3. Anchor to peer-reviewed content — explicitly exclude opinion/blog content for scientific questions
4. Specify recency — for fast-moving fields, constrain to last 2–3 years

Step 3 — Execute Search Waterfall

Run MCPs in this order (stop when sufficient evidence found):

A — Perplexity (if available)

Check if mcp__perplexity tools are accessible. If yes, use the prompts from references/perplexity-prompts.md.

Perplexity covers: documentation, GitHub issues, StackOverflow, technical blogs, and papers. Best for BUG and METHOD types.

B — bioRxiv / medRxiv MCP (always available)

bioRxiv.search_preprints(
  query="[core concept] [method/tool] [domain]",
  date_range="2022-2026",  # recency matters for methods
  limit=10
)

Use for: recent methods papers, benchmarks, protocols, negative results.

Refine with category filter if domain is clear (e.g., category="bioinformatics" or "biochemistry").

For a promising hit, fetch full abstract + DOI:

bioRxiv.get_preprint(doi="10.1101/XXXX")

C — PubMed MCP (always available)

PubMed.search_articles(
  query="[concept] [method] [organism/domain]",
  max_results=10,
  sort="relevance"   # or "date" for recent work
)
PubMed.get_article_metadata(pmid="XXXXXXXX")

Use for: established methods, clinical/biological protocols, theoretical foundations.

For methods specifically: add "protocol" OR "method" OR "algorithm" to the query.

For benchmarks: add "comparison" OR "benchmark" OR "evaluation" to the query.

Step 4 — Synthesize Results

After collecting hits, apply this synthesis protocol:

4.1 — Extract method/answer

State the answer as found in literature. Do not infer beyond what the sources say.

4.2 — Cite properly

Every claim derived from literature must carry a citation:

[Author et al., YEAR] — DOI: 10.XXXX/XXXXX
or
[bioRxiv YEAR-MM-DD] — DOI: 10.1101/XXXX

Never cite URLs. Always use DOIs when available.

4.3 — Rate confidence

★★★ — Multiple independent peer-reviewed sources agree
★★☆ — One peer-reviewed source or multiple preprints
★☆☆ — Single preprint or indirect evidence
☆☆☆ — No direct source found (see Step 5)

4.4 — Flag limitations

Note explicitly:

• Whether the source is peer-reviewed or a preprint
• Recency (methods older than 5 years may have superseded alternatives)
• Whether the source matches the exact library version / domain in question

Step 5 — Honest Reporting

If evidence is found (★★☆ or better): Present the answer with citations. Note any caveats.

If evidence is partial (★☆☆): Present what was found, state clearly that evidence is thin, and describe what would be needed to confirm.

If no evidence found (☆☆☆):

⚠️ LIT-RESCUE NEGATIVE RESULT

Searched: [list of queries and MCPs used]
Result: No peer-reviewed or preprint source found for [specific question].

Options:
1. Reformulate the question — the concept may use different terminology
2. Consult domain-specific documentation directly (package docs, official manual)
3. Run an empirical test (write a minimal script to probe the behavior)
4. Ask a domain expert or post to the relevant community forum

I will NOT speculate without a source.

This negative result is itself valuable — it tells the user that this is an open question or requires empirical investigation.

Decision Tree

Problem encountered
  │
  ├─ Is the answer confidently known from verified knowledge?
  │    └─ Yes → answer directly (don't invoke this skill)
  │    └─ Uncertain → continue
  │
  ├─ Classify type (Step 1)
  │
  ├─ Is it a BUG/implementation question?
  │    └─ Yes → Perplexity first (broader web coverage)
  │    └─ No → bioRxiv + PubMed first
  │
  ├─ Is the field fast-moving (ML, comp bio, cheminformatics)?
  │    └─ Yes → bioRxiv (preprints = 6–18 months ahead of journals)
  │    └─ No → PubMed (peer-reviewed = more reliable for established fields)
  │
  └─ Synthesize (Step 4) → Report honestly (Step 5)

Related Skills

• chem-brainstorm — for comp chem problems, run this before lit-rescue to check local tools
• scientific-skills:literature-review — for systematic multi-paper literature analysis
• scientific-skills:pubmed-database — for complex PubMed query strategies
• scientific-skills:biorxiv-database — for preprint-focused searches