Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Snowball · yogsoth-ai | DecimalAI

yogsoth-ai/snowball

SkillScore54Maintained

Citation-chain-driven literature survey starting from seed papers. Traces research lineage in both forward (who cited this) and backward (what this cited) directions until saturation. High deep-read ratio (67%). Use when the user already has key papers and wants to find everything connected to them — ancestors, descendants, and branch points.

researchby yogsoth-ai

Links this skill to your agents and keeps it current with the author’s updates.

Automatic

Research pack

1 · One-time — install the CLI (docs ↗)

pip install decimalai

2 · Get the skill — writes .claude/skills/yogsoth-ai-snowball/SKILL.md

decimalai skills pull yogsoth-ai-snowball --out .claude/skills/

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/yogsoth-ai-snowball/SKILL.md https://app.decimal.ai/s/yogsoth-ai-snowball/SKILL.md

Raw file: /s/yogsoth-ai-snowball/SKILL.md ↗ · machine-readable summary at /s/yogsoth-ai-snowball.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

No production data for this skill yet — usage numbers appear here once agents start running it.

Anonymized aggregate across every org using this skill — proof from real production use, not a demo.

Does this skill help? · Benchmark: with vs. withoutsign in to run this on your agents

24 test cases from the publisher’s eval.yaml · each run twice on gemini-3.6-flash — with the skill injected, then without · last run 18d ago

Yes — +29% lift+5% tokens0% turns54% of whole cases pass with skill (13/24)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-03	✗→✓	▲ Improved	15%	0%
case-07	✗→✓	▲ Improved	102%	0%
case-08	✗→✓	▲ Improved	15%	0%
case-10	✗→✓	▲ Improved	25%	0%
case-11	✗→✓	▲ Improved	56%	0%

Showing 5 of 24 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v4 — 8/10/2026

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

yogsoth-ai / de-anthropocentric-research-engine

skills/snowball/SKILL.md

View on GitHub ↗

License Apache-2.0Imported 8/10/2026Version v1

Skill contents

SKILL.md

Snowball

Purpose: Seed-first, forward/backward tracing — start from known seed papers and trace the research lineage in both directions.

When to use: User already has key papers and wants to find everything connected to them — what they built on, what built on them.

Budget

| Base SOP | Target | ±10% Range | |----------|--------|------------| | web-search | 20 results | 18–22 | | web-research | 3 pages | 2–4 | | paper-overview | 30 papers | 27–33 | | paper-search | 30 papers | 27–33 | | paper-research | 20 papers | 18–22 |

State Ledger

Print this table before each major iteration decision:

| SOP            | Target | Current | % Complete |
|----------------|--------|---------|------------|
| web-search     | 20     | ???     | ???%       |
| web-research   | 3      | ???     | ???%       |
| paper-overview | 30     | ???     | ???%       |
| paper-search   | 30     | ???     | ???%       |
| paper-research | 20     | ???     | ???%       |

Do not exit the strategy until all rows reach ≥90%.

Available Tactics

citation-chaining — forward/backward citation expansion until saturation

Available SOPs

Import (strict protocol execution):

web-search → web-browsing/skills/web-search/SKILL.md
web-research → web-browsing/skills/web-research/SKILL.md
paper-overview → literature-engine/skills/literature-overview/SKILL.md
paper-search → literature-engine/skills/literature-search/SKILL.md
paper-research → literature-engine/skills/literature-research/SKILL.md

Subagent (CC decides when to invoke):

seed-selection — validate and prioritize starting papers
saturation-detection — determine when to stop (diminishing returns)
gap-identification — find what the literature hasn't addressed
survey-synthesis — produce final structured output

Execution Guidance

seed-selection validates and prioritizes the starting papers
citation-chaining is the primary operation — iterate until saturation
saturation-detection determines when to stop (diminishing returns)
Minimal web-search (only for context that papers don't provide)
High paper-research ratio (20/30 = 67% deep-read rate) — trace papers deserve thorough reading
Build a clear lineage: who influenced whom, how ideas evolved

Output Format

Research Lineage Map containing:

Seed papers → ancestors (backward trace)
Seed papers → descendants (forward trace)
Evolution of ideas across generations
Key branch points where the field diverged
Current frontier (most recent descendants)
Lineage visualization (text-based DAG)

Available Tactics

Optional, no fixed order; the final leaf is always a sop.

| Tactic | When to use | | --- | --- | | citation-chaining | Forward and backward citation tracing tactic — expand paper coverage by tracing citation networks in both directions from seed/key papers. Alternates forward (who cited this) and backward (what this cited) passes until saturation. |

Available SOPs

Optional, no fixed order; the final leaf is always a sop.

| SOP | When to use | | --- | --- | | knowledge-acquisition-gap-identification | Identify what the literature has NOT addressed — missing methods, untested combinations, unexplored applications, contradictions without resolution. Used by all strategies. | | knowledge-acquisition-paper-overview | Abstract-level paper scanning for broad coverage. Import of literature-engine/literature-overview skill. Abstract-level only — no methodology conclusions from abstracts. | | knowledge-acquisition-paper-research | Full-depth paper reading with raw text extraction. Import of literature-engine/literature-research skill. Must read fullText (true) — equations, hyperparameters, specific claims extracted. | | knowledge-acquisition-paper-search | AI-summarized paper reading for intermediate depth. Import of literature-engine/literature-search skill. Must call get_paper_content for every analyzed paper. | | knowledge-acquisition-saturation-detection | Determine when additional searching yields diminishing returns. Analyzes the latest expansion batch against existing corpus to judge continue/near-saturation/saturated. Used by snowball and systematic-survey. | | knowledge-acquisition-web-research | Full-page web reading for non-academic perspectives — blogs, tech reports, product pages, industry analysis. Import of web-browsing/web-research skill. Must fetch full page via apify for every analyzed page. | | knowledge-acquisition-web-search | Quick web scanning for landscape understanding. Import of web-browsing/web-search skill. Snippets only — no conclusions from snippets alone. | | seed-selection | Validate and prioritize starting papers for snowball surveys. Evaluates which seeds will yield the richest citation traces based on citation count, recency, and network position. | | survey-synthesis | Final synthesis step — weave all gathered evidence (reading notes, extracted data, categorizations) into a coherent structured output appropriate to the strategy type. Used by all 5 strategies as the final step. |

Illustrative Want to see what yogsoth-ai/snowball would put in front of your model? Preview the rendered body and check its trigger phrases — no LLM call, no signup.

Latest benchmark · publisher runVerified

9/4/2026, 04:53 UTC · 24 cases

Lift

+29%

54% vs 25%

Agent turns

1.0 vs 1.0

Tokens

+5%

6,682.08 vs 6,337.63

fail → pass · 10pass → pass · 3fail → fail · 8pass → fail · 3

Per-case results (24)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-04 I am conducting a systematic literature review on clinical applications of AI in radiology following PRISMA guidelines. I need to construct comprehensive Boolean search strings for PubMed, Embase, and IEEE Xplore, along with inclusion/exclusion criteria for systematic screening. Please provide the search string protocol and screening checklist.	pass→fail	27,830	45,554	+64%	1	1	0%	5,371	9,148	+70%	0	0	—
case-01 I have two foundational papers on diffusion models: Sohl-Dickstein et al. (2015) and Ho et al. (2020). I want to trace the literature backwards to see what early mathematical physics and generative modeling work influenced them, and trace forwards to see all key downstream architectures and applications. Please conduct a forward and backward citation network analysis starting from these seeds. I need a comprehensive lineage report detailing the ancestral influences, downstream branches, major conceptual divergence points, the current frontier, and a text-based directed acyclic graph illustrating how the core ideas progressed over time.	fail→fail	63,833	51,976	-19%	1	1	0%	8,310	9,559	+15%	0	0	—
case-02 Starting from the classic RAG paper by Lewis et al. (2020) and REALM by Guu et al. (2020) as my anchor papers, I need to chart the full citation tree in both directions. Trace backwards to identify the foundational memory/retrieval concepts they built upon, and trace forwards through citing literature to gather modern state-of-the-art developments. Synthesize your findings into a lineage breakdown that explains conceptual shifts across paper generations, highlights where the domain branched into distinct sub-fields, summarizes the active research edge, and provides an ASCII dependency graph of the papers.	fail→fail	51,935	11,457	-78%	1	1	0%	8,316	2,327	-72%	0	0	—
case-03 I am investigating deep Q-learning and want to understand its complete academic evolution. Using Mnih et al. (2015) 'Human-level control through deep reinforcement learning' as my seed, please run a seed-based bidirectional citation expansion. Trace the prior theoretical foundations that enabled this paper as well as the later works that built directly on top of it. I want a structured report outlining historical ancestors, modern descendants, critical pivot points where novel algorithms branched off, the present research frontier, and a text DAG showing the citation network.	fail→pass	50,070	44,718	-11%	1	1	0%	8,299	9,548	+15%	0	0	—
case-05 I already downloaded 10 PDF papers on contrastive learning to my local folder and do not need any external search or citation tracing. Please synthesize these specific 10 papers into a summary table comparing their loss functions, batch sizes, and downstream benchmark scores.	fail→fail	5,445	46,485	+754%	1	1	0%	899	9,490	+956%	0	0	—
case-06 I have collected effect sizes and sample sizes from 8 randomized controlled trials comparing two surgical techniques in a spreadsheet. Please perform a fixed-effects statistical meta-analysis to calculate the pooled odds ratio and 95% confidence intervals.	fail→fail	13,386	44,380	+232%	1	1	0%	1,370	9,484	+592%	0	0	—
case-07 I want to map the literature starting from the original Transformer paper (Vaswani et al., 2017). Please start the bidirectional citation expansion process. I would like to see how you track your planned search execution and SOP budget targets across web and paper channels before making expansion decisions.	fail→pass	32,664	47,240	+45%	1	1	0%	4,712	9,496	+102%	0	0	—
case-08 We are researching parameter-efficient fine-tuning starting from LoRA (Hu et al., 2021). I want deep full-text extraction of methodology details, equations, and hyperparameter setups from the key papers in this lineage, rather than surface-level abstract scanning. Show your strategy execution plan and SOP targets before diving into the papers.	fail→pass	42,513	46,595	+10%	1	1	0%	8,260	9,509	+15%	0	0	—
case-09 I have a list of potential starting papers for a survey on graph neural networks: Kipf & Welling (2017), Velickovic et al. (2018), and Hamilton et al. (2017). Before beginning citation chaining, how do you evaluate and prioritize these initial candidates to ensure maximum coverage of the lineage?	pass→pass	18,125	24,659	+36%	1	1	0%	2,988	5,321	+78%	0	0	—
case-10 Starting from the BERT paper (Devlin et al., 2019), I want to analyze both its intellectual precedents in masked language modeling and the downstream pre-training architectures it inspired. Explain your expansion tactic and trace both directions until saturation.	fail→pass	47,134	55,111	+17%	1	1	0%	7,568	9,488	+25%	0	0	—
case-11 We are tracing the literature on diffusion models starting from Sohl-Dickstein et al. (2015). After several rounds of citation expansion, how do you determine that searching further will produce diminishing returns and that the survey is ready to exit?	fail→pass	17,036	21,055	+24%	1	1	0%	2,599	4,058	+56%	0	0	—
case-12 I need a comprehensive lineage report on deep reinforcement learning algorithms starting from DQN (Mnih et al., 2015). The user requires a visual representation of how ideas evolved from early Q-learning through double Q-learning and dueling architectures to modern Rainbow DQN without using external binary images.	pass→fail	52,166	12,487	-76%	1	1	0%	8,249	2,458	-70%	0	0	—
case-13 When expanding the citation network around AlphaFold (Jumper et al., 2021), we need to gather both academic papers and relevant background context. How should search effort be balanced between full academic papers and general web searches?	fail→pass	14,740	33,817	+129%	1	1	0%	2,266	6,271	+177%	0	0	—
case-14 I am executing a literature expansion strategy and have reached 80% of my paper-overview target and 75% of my paper-research target. Is it permissible to finalize and output the synthesis report now?	fail→pass	15,243	10,654	-30%	1	1	0%	1,564	2,263	+45%	0	0	—
case-15 Beginning with Bahdanau et al. (2014) and Vaswani et al. (2017) as seeds, perform a bidirectional citation expansion on sequence-to-sequence attention mechanisms. Provide the state ledger table tracking execution budgets and outline ancestral mechanisms and downstream efficient attention variants.	pass→pass	44,236	46,828	+6%	1	1	0%	8,250	9,499	+15%	0	0	—
case-16 Use He et al. (2016) 'Deep Residual Learning for Image Recognition' as a seed to map deep computer vision architecture evolution. Track your strategy state across web and paper SOP budgets, and map the ancestral highway skip-connections down to modern vision transformers.	fail→fail	63,147	12,565	-80%	1	1	0%	8,244	2,496	-70%	0	0	—
case-17 Start with Hu et al. (2021) LoRA and trace parameter-efficient fine-tuning literature in both directions. Ensure you track SOP execution metrics in the mandatory table and detail ancestral low-rank matrix techniques alongside modern quantized low-rank adaptations.	fail→pass	43,918	47,972	+9%	1	1	0%	8,242	9,491	+15%	0	0	—
case-18 Conduct a citation expansion starting from Jumper et al. (2021) AlphaFold2. State your SOP target ledger before making expansion decisions, tracing backward to co-evolutionary contact prediction and forward to de novo protein design models.	fail→fail	50,568	12,100	-76%	1	1	0%	8,238	2,541	-69%	0	0	—
case-19 Using Kingma & Ba (2014) Adam optimizer as the seed paper, execute a forward and backward citation network analysis. Present your SOP state ledger to monitor execution budgets and detail optimization ancestors and modern adaptive learning rate successors.	fail→fail	43,068	11,198	-74%	1	1	0%	8,236	2,273	-72%	0	0	—
case-20 Trace the pre-trained encoder lineage starting from Devlin et al. (2019) BERT. Print your state ledger table tracking SOP target counts, and produce a research lineage report including ancestral ELMo/word2vec models, descendant RoBERTa/DeBERTa models, and a text DAG.	fail→fail	49,819	12,601	-75%	1	1	0%	8,252	2,653	-68%	0	0	—
case-21 Using Schulman et al. (2017) PPO as the seed paper, trace reinforcement learning policy optimization backward to TRPO and forward to RLHF applications. Track SOP budgets in the required state ledger format before iteration decisions.	fail→pass	44,651	13,034	-71%	1	1	0%	8,237	2,728	-67%	0	0	—
case-22 Starting from Kipf & Welling (2017) Graph Convolutional Networks, perform bidirectional citation chaining. Report your SOP budget tracking table, detail spectral to spatial evolution, and include a text DAG.	pass→fail	40,049	40,386	+1%	1	1	0%	8,231	9,480	+15%	0	0	—
case-23 Trace unsupervised representation learning using Chen et al. (2020) SimCLR as the seed paper. Print the SOP state ledger with its 5 budget categories and present both ancestral mutual information estimators and descendant non-contrastive methods.	pass→pass	48,211	46,969	-3%	1	1	0%	8,237	9,486	+15%	0	0	—
case-24 Perform citation snowballing starting from Shazeer et al. (2017) 'Outrageously Large Neural Networks' (Sparse MoE). Display your SOP state ledger before making expansion decisions and map historical multi-expert roots down to modern sparse LLMs.	fail→pass	43,289	63,962	+48%	1	1	0%	7,165	11,303	+58%	0	0	—

These cases were also screened by the skill’s intent / safety pass.See the Trust review ↗Reproduce this benchmark in Colab ↗

How this skill was measured

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. 24 cases were attempted. The headline lift of +29 percentage points is the difference between those two pass rates over the 24 comparable cases. 4 cases got worse with the skill loaded, and they are included in that figure.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 25%
Pass rate with the skill: 54.2%
Cases attempted: 24
Cases counted toward lift: 24
Cases that regressed: 4
Skill version measured: v1
Run completed: 9/4/2026, 04:53 UTC

Measured example — same prompt, with vs. without the skill · from the benchmark run · case case-03

Prompt

I am investigating deep Q-learning and want to understand its complete academic evolution. Using Mnih et al. (2015) 'Human-level control through deep reinforcement learning' as my seed, please run a seed-based bidirectional citation expansion. Trace the prior theoretical foundations that enabled this paper as well as the later works that built directly on top of it. I want a structured report outlining historical ancestors, modern descendants, critical pivot points where novel algorithms branched off, the present research frontier, and a text DAG showing the citation network.

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.

Show both answers

SkillScore blends 4 signals1 of 4 complete

BenchmarkPass rate WITH skill on a verified run54

Live evalNeeds ≥10 evaluated traces in 30 days

AI ratingNeeds ≥10 rated traces

AdoptionNeeds ≥10 serves across ≥2 orgs in 30 days

Efficiency · separate axis (at equal correctness)+0% turns+25% tokens

Shown for transparency; not credited unless a verified run shows no correctness regression.

Backed by 1 of 4 signals — each added signal makes the score more robust.

Verified on gemini-3.6-flash · re-tested 17 days ago

Other measured skills in the registry, with their headline benchmark lift.

Ready to supercharge your AI agents?

yogsoth-ai/snowball

SkillSafety

Source

Snowball

Budget

State Ledger

Available Tactics

Available SOPs

Execution Guidance

Output Format

Available Tactics

Available SOPs

How this skill was measured

Measured example — same prompt, with vs. without the skill · from the benchmark run · case case-03

Related skills