Evidence synthesis
Skill cdeust/zetetic-team-subagents/skills/evidence-synthesis
Turn a pile of sources into a defensible claim. Use for "what does the literature actually say?", conflicting studies or benchmarks, "is this result trustworthy?", arguments that need claim-evidence-warrant structure, qualitative data needing coding, or a theory with no catalog of its own contradicting evidence.From its SKILL.md
npx -y skills add cdeust/zetetic-team-subagents --skill evidence-synthesisAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 7 stars7 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
- runs commandsInstructs the agent to run 3 commands, including `tools/genius-invoker.sh route "<problem>"` and 2 more.
SKILL.md
2.9 KB, 614 tokens by cl100k_base, as published. Nobody here has run it
Evidence Synthesis
Problem shape: multiple sources (papers, benchmarks, interviews, texts) must be combined into one graded conclusion — and the current draft cherry-picks, skips heterogeneity, or has no explicit warrant connecting evidence to claim.
Relevant geniuses
| Agent | Use when |
|---|---|
| cochrane | studies must be systematically found, graded, and pooled; publication bias suspected; effect sizes needed, not vote-counting |
| toulmin | an argument needs explicit claim-evidence-warrant structure with qualifiers and rebuttals; standards differ by field |
| feynman | a result is suspiciously clean; a citation is used without the ability to rederive it; integrity audit before publishing |
| darwin | evidence accumulates over a long horizon; the theory needs a difficulty book of its own contradictions; hardest case first |
| laplace | evidence must update a belief quantitatively — priors, likelihoods, calibration |
| gadamer | sources must be interpreted, not just counted — pre-understandings audited, the charitable reading found first |
| strauss | qualitative data (interviews, tickets, reviews) needs open/axial coding until saturation, not anecdote-picking |
| geertz | the evidence is behavior in context — thick description and emic vs etic separation before conclusions |
Invocation
- Pick the best-fit agent above. If two or more fit, run
tools/genius-invoker.sh route "<problem>"and take the top ranked match. - Load it:
tools/genius-invoker.sh invoke <agent> "<problem>", then readagents/genius/<agent>.mdin full. - Apply the agent's
<workflow>step by step and answer in its<output-format>. Every synthesized claim carries its evidence grade and its strongest known counter-evidence (coding-standards §8: no source, no claim). - Typical chain: cochrane grades the corpus → toulmin structures the argument
→ feynman audits the integrity. Run via
tools/genius-invoker.sh compose cochrane toulmin -- "<problem>". - If no shape above matches, use a standard team agent instead.
Refuse when
- The conclusion is fixed in advance and only supporting sources are wanted.
- Primary sources are unavailable and only summaries-of-summaries exist — say "I don't know" rather than synthesize hearsay.
What ships with it
Read from the repository
Just SKILL.md. No reference files, no scripts.