Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Capability menu for the research engine. Lists the 10 freely-composable research packages, what each does, when to reach for it, and a pointer to its full skill table. Read this after north-star crystallization to decide which packages to use — no fixed order. Also serves as the skill-index (capability map).
.claude/skills/yogsoth-ai-research-catalog/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-03 | ✗→✓ | ▲ Improved | -50% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 19% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 18% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -1% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -34% | 0% |
These 10 packages are freely-composable capability domains. There is no prescribed order and no pipeline. After your research direction is crystallized, decide which packages to invoke, in what sequence, and whether to loop back — based on the current research task. Each package is a self-contained research engine with its own campaigns, strategies, tactics, and SOPs.
To use a package, read its reference table under references/<package>.md for the full skill list (sorted by layer, then name). Pick the package whose purpose matches your current need; open its table; select the skills you need.
The 10 packages are listed alphabetically below — the order carries no sequencing meaning.
Research-to-artifact compilation engine: compiles a completed context/ research record into an ARA (Agent-Native Research Artifact) — a 4-layer machine-executable knowledge package (PAPER.md + logic/ + src/ + trace/ + evidence/), not a LaTeX narrative paper — then runs a Level-2 epistemic rigor review. One campaign — ara-from-context.
Reach for it when: a research arc (typically after experiment-execution) has produced results worth packaging for agents to reproduce/extend → compile context into an ARA + epistemic review. Requires the external compiler / rigor-reviewer skills (npx @ara-commons/ara-skills).
Skills: see references/ara-from-context.md
Universal convergence engine: turns an unstructured candidate set into ranked selections, balanced portfolios, and validated decisions. Six campaigns — multi-criteria-scoring, pairwise-ranking, structured-consensus, feasibility-assessment, portfolio-optimization, steel-manning.
Reach for it when: score/rank candidates against multiple criteria; produce a global ranking via pairwise comparisons; multiple perspectives disagree and need convergence; assess feasibility/readiness; select a balanced portfolio; verify rejected candidates or stress-test winners.
Skills: see references/convergence.md
Creative generation engine: transforms hypotheses and research questions into diverse solution spaces. Ten parallel creativity campaigns spanning structural, analogical, destructive, and combinatorial methods.
Reach for it when: SCAMPER / TRIZ / component surgery / structural transformation / function trimming → structural-deconstruction; cross-domain / analogical transfer / bisociation / random stimulus → cross-domain-discovery; assumption negation / reverse brainstorming / worst method → assumption-destruction; biomimicry / biological analogy / BioTRIZ → biomimicry; analogy / metaphor / excursion method → synectics; morphological analysis / Zwicky box / design space → morphological-exploration; PO / lateral thinking / concept fan → lateral-thinking; concept blending / blending / emergence → combinatorial-creativity; perspective switching / six hats / role-play → perspective-forcing; enumeration / coverage analysis / method matrix → systematic-enumeration.
Skills: see references/creative-ideation.md
Deep insight engine: from surface phenomena to root causes, boundaries, assumptions, and the problem itself. Five campaigns — gap-analysis, insight, boundary-analysis, sensitivity-analysis, problem-reformulation.
Reach for it when: gap identification / white-space classification / evidence map / prioritization → gap-analysis; root-cause analysis / stakeholders / tensions / HMW / 5 Whys → insight; validity boundaries / method failure / robustness / distribution shift → boundary-analysis; assumption ranking / sensitivity / variance decomposition / critical path → sensitivity-analysis; redefining the problem / dominant ideas / multiple perspectives / wicked problems → problem-reformulation.
Skills: see references/deep-insight.md
Experiment execution engine: from validated hypotheses/approaches through experiment design, constraint analysis, scenario planning, implementation planning, to actual execution and result collection. Four campaigns — experiment-design, constraint-analysis, scenario-planning, implementation-planning.
Reach for it when: experiment design / factors / variables / ablation / baseline comparison / statistical methods → experiment-design; bottleneck / constraint / insufficient resources / dependencies / conflicts → constraint-analysis; scenarios / future / robustness / worst case / competitors / timeline → scenario-planning; planning / execution / implementation / running experiments / result analysis / reproducibility → implementation-planning.
Skills: see references/experiment-execution.md
Goal-driven hypothesis & research-question formation engine: turns upstream gaps and insights into testable hypotheses and precise research questions. Three campaigns — gap-prioritization, hypothesis-formulation, research-question.
Reach for it when: gap ranking / prioritization / which is worth doing / multi-dimensional scoring / portfolio → gap-prioritization; hypothesis generation / theoretical derivation / falsifiability / If-then / variables / mechanisms / competing hypothesis → hypothesis-formulation; research question / PICO / SPIDER / FINER / scope / sub-question decomposition / success criteria → research-question.
Skills: see references/hypothesis-formation.md
Systematic research-knowledge acquisition engine: academic literature, patent landscapes, benchmark evaluations, cross-study statistical synthesis, and SOTA baselines. Five campaigns — literature-survey, patent-mining, benchmark-archaeology, meta-analysis, baseline-establishment.
Reach for it when: literature review / survey / paper search / PRISMA / snowball → literature-survey; patent analysis / prior art / white space / claims / IPC → patent-mining; benchmark analysis / evaluation methods / metric flaws / leaderboards / saturation → benchmark-archaeology; cross-study statistical synthesis / effect size / heterogeneity / publication bias / GRADE → meta-analysis; SOTA compilation / performance comparison / baseline reproduction / progress curves → baseline-establishment.
Skills: see references/knowledge-acquisition.md
Knowledge-structuring engine: compiles findings into structured artifacts in the wiki vault. Four campaigns — ontology-building (concept hierarchies), causal-modeling (causal graphs), dimensional-analysis (design-space maps), argument-mapping (argument graphs).
Reach for it when: building a domain ontology — concept extraction, relation typing, taxonomy construction → ontology-building; identifying variables and mapping mechanisms into a causal graph → causal-modeling; discovering the dimensions/axes of a design space and enumerating combinations to find gaps → dimensional-analysis; extracting claims, linking evidence, assessing argument strength → argument-mapping.
Skills: see references/knowledge-structuring.md
Research intent crystallization engine: transforms vague research interests into precise, actionable North Star statements through structured dialogue, producing a North Star plus a structured ResearchBrief. One campaign with three strategies — cold-start, warm-start, hot-start — chosen by the user's existing clarity.
Reach for it when: the research direction needs to be (re)crystallized — no direction at all → cold-start; a general direction but not specific → warm-start; a specific topic/problem needing structure → hot-start. Reach for it whenever the current direction feels fuzzy and you want to re-anchor before composing other packages.
Skills: see references/north-star-crystallization.md
Research-artifact stress-testing engine: adversarially validates any artifact — hypotheses, claims, experiment designs, approaches, research questions, gaps, or ideas — until every claim survives or is annotated. Five campaigns — multiagent-debate, red-teaming, failure-anticipation, counterfactual-probing, adversarial-stress-testing.
Reach for it when: adversarial debate / multi-perspective review → multiagent-debate; systematic attack / assumption challenge → red-teaming; failure-mode prediction / risk assessment → failure-anticipation; critical-dependency probing / causal necessity → counterfactual-probing; logical falsification / boundary testing → adversarial-stress-testing.
Skills: see references/stress-test.md
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 28,835 | 24,215 | -16% | 1 | 1 | 0% | 4,229 | 4,955 | +17% | 0 | 0 | — |
case-02 | fail→fail | 47,237 | 42,942 | -9% | 1 | 1 | 0% | 8,278 | 8,075 | -2% | 0 | 0 | — |
case-03 | fail→pass | 59,222 | 17,980 | -70% | 1 | 1 | 0% | 8,279 | 4,144 | -50% | 0 | 0 | — |
case-04 | fail→pass | 19,338 | 12,202 | -37% | 1 | 1 | 0% | 2,400 | 2,865 | +19% | 0 | 0 | — |
case-05 | fail→pass | 16,864 | 10,520 | -38% | 1 | 1 | 0% | 1,995 | 2,363 | +18% | 0 | 0 | — |
case-06 | pass→pass | 22,373 | 5,192 | -77% | 1 | 1 | 0% | 2,636 | 2,434 | -8% | 0 | 0 | — |
case-07 | fail→pass | 15,937 | 9,577 | -40% | 1 | 1 | 0% | 2,443 | 2,411 | -1% | 0 | 0 | — |
case-08 | fail→pass | 24,230 | 9,561 | -61% | 1 | 1 | 0% | 3,935 | 2,614 | -34% | 0 | 0 | — |
case-09 | pass→pass | 22,244 | 5,317 | -76% | 1 | 1 | 0% | 2,869 | 2,479 | -14% | 0 | 0 | — |
case-10 | pass→pass | 24,494 | 5,589 | -77% | 1 | 1 | 0% | 1,559 | 2,551 | +64% | 0 | 0 | — |
case-11 | pass→pass | 23,359 | 6,646 | -72% | 1 | 1 | 0% | 3,334 | 2,465 | -26% | 0 | 0 | — |
case-12 | fail→pass | 26,019 | 13,336 | -49% | 1 | 1 | 0% | 3,794 | 2,831 | -25% | 0 | 0 | — |
case-13 | pass→pass | 53,196 | 4,881 | -91% | 1 | 1 | 0% | 2,817 | 2,408 | -15% | 0 | 0 | — |
case-14 | fail→pass | 14,719 | 4,186 | -72% | 1 | 1 | 0% | 2,351 | 2,213 | -6% | 0 | 0 | — |
case-15 | pass→pass | 18,262 | 6,753 | -63% | 1 | 1 | 0% | 2,327 | 2,447 | +5% | 0 | 0 | — |
case-16 | pass→pass | 19,847 | 4,006 | -80% | 1 | 1 | 0% | 2,852 | 2,345 | -18% | 0 | 0 | — |
case-17 | fail→pass | 21,852 | 2,576 | -88% | 1 | 1 | 0% | 2,723 | 2,113 | -22% | 0 | 0 | — |
case-18 | fail→pass | 18,789 | 12,287 | -35% | 1 | 1 | 0% | 2,364 | 2,820 | +19% | 0 | 0 | — |
case-19 | pass→pass | 29,952 | 3,360 | -89% | 1 | 1 | 0% | 2,252 | 2,267 | +1% | 0 | 0 | — |
case-20 | fail→pass | 22,760 | 23,197 | +2% | 1 | 1 | 0% | 3,910 | 4,193 | +7% | 0 | 0 | — |
case-21 | fail→fail | 6,734 | 11,054 | +64% | 1 | 1 | 0% | 1,133 | 3,412 | +201% | 0 | 0 | — |
case-22 | fail→pass | 16,780 | 5,628 | -66% | 1 | 1 | 0% | 2,903 | 2,727 | -6% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +50 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.