research

Name: research
Author: tonyazhuuki

29stars7forks

How to install research?

tonyazhuuki/deep-research-skill/research

$npx -y skills add tonyazhuuki/deep-research-skill --skill research

Installs into the current project.

Summary

Multi-agent iterative research pipeline using Eric Jang methodology. 10-18 specialized agents across 3 cycles produce consensus references, interaction maps, personalized syntheses, and action plans with domain-specific stress-testing, iterative deepening, and anti-pattern guards

Files · 1

View on GitHub

SKILL.md

1	# Deep Research Skill
2
3	A structured multi-agent research pipeline that turns any topic into a comprehensive, fact-checked, bilingual (EN+RU) synthesis with actionable recommendations. Built on Eric Jang's iterative methodology from ["As Rocks May Think"](https://arxiv.org/abs/2602.00000).
4
5	## How It Works
6
7	The skill orchestrates 10-19 specialized AI agents across 3 mandatory cycles:
8
9	```
10	Cycle 1: Broad Search → 4-5 parallel SCOUTs explore the landscape
11	Quality Gates → CRITIC + METHODOLOGIST cross-check findings
12	Reflection 1 → Identify gaps, generate competing hypotheses
13
14	Cycle 2: Deep Dives → 2-3 targeted agents test hypotheses + stress-test questions
15	Iterative Deep. → Auto-resolve CONTESTED claims (WEAK → follow-up DD)
16	Reflection 2 → Convergence analysis, hypothesis verdicts
17
18	Cycle 3: Execute → Python scripts (analysis, models, visualizations)
19	Synthesize → SYNTHESIZER creates integrated document
20	Verify → FACT-CHECKER + CITATION_VERIFIER + DOMAIN_REVIEWER
21	Apply → ACTION MAPPER updates user's protocols/goals
22	```
23
24	## Research Modes
25
26	\| Mode \| Output \| When to use \|
27	\|------\|--------\|-------------\|
28	\| personalized (default) \| `synthesis.md` \| Specific question for your context \|
29	\| consensus \| `consensus_reference.md` \| Building knowledge base (population-level truth) \|
30	\| consensus+interactions \| consensus + `interaction_map.md` \| Cross-effects matter \|
31	\| full \| All three documents \| Deep investigation \|
32
33	## Agent Roles
34
35	\| Role \| Count \| Purpose \|
36	\|------\|-------\|---------\|
37	\| SCOUT \| 4-5 \| Broad literature search, each with unique reasoning style \|
38	\| CRITIC \| 1 \| Cross-stream contradictions, bias audit, weak evidence \|
39	\| METHODOLOGIST \| 1 \| Methodological quality — domain-specific evidence hierarchy (GRADE for health, forecast audit for macro, reproducibility for science) \|
40	\| DEEP DIVER \| 2-3+ \| Hypothesis testing + domain stress-test questions. Auto-spawns on CONTESTED claims (iterative deepening) \|
41	\| SYNTHESIZER \| 1 \| Integration across all sources into coherent document \|
42	\| INTERACTION MAPPER \| 1 \| Cross-domain interactions that change recommendations \|
43	\| DOMAIN_REVIEWER \| 1 \| Domain-specific review: MEDICAL (health), MACRO (markets), MARKET (company), METHODOLOGY (science) \|
44	\| FACT-CHECKER \| 1 \| Top-15 numerical claims verification \|
45	\| CITATION_VERIFIER \| script \| Python API check against Semantic Scholar/PubMed/CrossRef \|
46	\| TEMPORAL DIFF \| 0-1 \| Compares new consensus with previous version (UPDATE mode only) \|
47	\| ACTION MAPPER \| 1 \| Converts findings into TODO blocks in user's files \|
48
49	## v4.1 Features
50
51	### Domain Stress-Test Questions
52	Each Deep Diver must answer ≥2 mandatory adversarial questions from their domain adapter. Examples:
53	- Health: "Under what conditions does this intervention become harmful for MY profile?"
54	- Macro: "What single event makes this consensus irrelevant within 18 months?"
55	- Company: "What does every successful player understand that customers never say out loud?"
56	- Science: "What replication failure would collapse this finding?"
57
58	### Iterative Deepening on CONTESTED Claims
59	When Cycle 2 convergence finds CONTESTED claims (confidence <0.5), the pipeline auto-spawns targeted Deep Divers to find tiebreaker evidence. Max 2 rounds, 15% budget cap.
60
61	### Anti-Pattern Guards
62	Each domain adapter includes a "Common Anti-Patterns" table — pre-flight checklist of domain-specific mistakes (e.g., "citing mouse studies as human evidence" for health, "extrapolating 3-year trend as permanent" for macro).
63
64	### Company Domain Enhancements
65	- Raw language preservation: SCOUTs preserve verbatim customer/user quotes with tags
66	- Opportunity taxonomy: Action Mapper classifies opportunities as CONTRARIAN / TIMING_PLAY / SAFE_BET
67
68	## v4.2 Features
69
70	### Temporal Diff (UPDATE mode)
71	When re-researching a topic that already has a consensus reference, the pipeline auto-detects the previous research and offers UPDATE mode. In UPDATE mode:
72	- Full research runs identically to fresh mode (SCOUTs do NOT see previous consensus — no anchoring bias)
73	- TEMPORAL DIFF agent runs after SYNTHESIZER, comparing old and new consensus claim-by-claim
74	- Produces `_temporal_diff.md` with 5 categories: CONFIRMED / REVISED / CONTRADICTED / OBSOLETE / NEW
75	- Adds a `## Temporal Diff Summary` to the new consensus_reference.md
76	- Archives previous version as `_previous_consensus_[YYYY_MM].md`
77	- Stability score: CONFIRMED / total — tells you how much the field moved
78
79	CLI flags:
80	- `--update

People who liked this also viewed

6 picks

Type

research

29stars7forks

View on GitHub

How to install research?

tonyazhuuki/deep-research-skill/research

$npx -y skills add tonyazhuuki/deep-research-skill --skill research

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Deep Research Skill
2
3	A structured multi-agent research pipeline that turns any topic into a comprehensive, fact-checked, bilingual (EN+RU) synthesis with actionable recommendations. Built on Eric Jang's iterative methodology from ["As Rocks May Think"](https://arxiv.org/abs/2602.00000).
4
5	## How It Works
6
7	The skill orchestrates 10-19 specialized AI agents across 3 mandatory cycles:
8
9	```
10	Cycle 1: Broad Search → 4-5 parallel SCOUTs explore the landscape
11	Quality Gates → CRITIC + METHODOLOGIST cross-check findings
12	Reflection 1 → Identify gaps, generate competing hypotheses
13
14	Cycle 2: Deep Dives → 2-3 targeted agents test hypotheses + stress-test questions
15	Iterative Deep. → Auto-resolve CONTESTED claims (WEAK → follow-up DD)
16	Reflection 2 → Convergence analysis, hypothesis verdicts
17
18	Cycle 3: Execute → Python scripts (analysis, models, visualizations)
19	Synthesize → SYNTHESIZER creates integrated document
20	Verify → FACT-CHECKER + CITATION_VERIFIER + DOMAIN_REVIEWER
21	Apply → ACTION MAPPER updates user's protocols/goals
22	```
23
24	## Research Modes
25
26	\| Mode \| Output \| When to use \|
27	\|------\|--------\|-------------\|
28	\| personalized (default) \| `synthesis.md` \| Specific question for your context \|
29	\| consensus \| `consensus_reference.md` \| Building knowledge base (population-level truth) \|
30	\| consensus+interactions \| consensus + `interaction_map.md` \| Cross-effects matter \|
31	\| full \| All three documents \| Deep investigation \|
32
33	## Agent Roles
34
35	\| Role \| Count \| Purpose \|
36	\|------\|-------\|---------\|
37	\| SCOUT \| 4-5 \| Broad literature search, each with unique reasoning style \|
38	\| CRITIC \| 1 \| Cross-stream contradictions, bias audit, weak evidence \|
39	\| METHODOLOGIST \| 1 \| Methodological quality — domain-specific evidence hierarchy (GRADE for health, forecast audit for macro, reproducibility for science) \|
40	\| DEEP DIVER \| 2-3+ \| Hypothesis testing + domain stress-test questions. Auto-spawns on CONTESTED claims (iterative deepening) \|
41	\| SYNTHESIZER \| 1 \| Integration across all sources into coherent document \|
42	\| INTERACTION MAPPER \| 1 \| Cross-domain interactions that change recommendations \|
43	\| DOMAIN_REVIEWER \| 1 \| Domain-specific review: MEDICAL (health), MACRO (markets), MARKET (company), METHODOLOGY (science) \|
44	\| FACT-CHECKER \| 1 \| Top-15 numerical claims verification \|
45	\| CITATION_VERIFIER \| script \| Python API check against Semantic Scholar/PubMed/CrossRef \|
46	\| TEMPORAL DIFF \| 0-1 \| Compares new consensus with previous version (UPDATE mode only) \|
47	\| ACTION MAPPER \| 1 \| Converts findings into TODO blocks in user's files \|
48
49	## v4.1 Features
50
51	### Domain Stress-Test Questions
52	Each Deep Diver must answer ≥2 mandatory adversarial questions from their domain adapter. Examples:
53	- Health: "Under what conditions does this intervention become harmful for MY profile?"
54	- Macro: "What single event makes this consensus irrelevant within 18 months?"
55	- Company: "What does every successful player understand that customers never say out loud?"
56	- Science: "What replication failure would collapse this finding?"
57
58	### Iterative Deepening on CONTESTED Claims
59	When Cycle 2 convergence finds CONTESTED claims (confidence <0.5), the pipeline auto-spawns targeted Deep Divers to find tiebreaker evidence. Max 2 rounds, 15% budget cap.
60
61	### Anti-Pattern Guards
62	Each domain adapter includes a "Common Anti-Patterns" table — pre-flight checklist of domain-specific mistakes (e.g., "citing mouse studies as human evidence" for health, "extrapolating 3-year trend as permanent" for macro).
63
64	### Company Domain Enhancements
65	- Raw language preservation: SCOUTs preserve verbatim customer/user quotes with tags
66	- Opportunity taxonomy: Action Mapper classifies opportunities as CONTRARIAN / TIMING_PLAY / SAFE_BET
67
68	## v4.2 Features
69
70	### Temporal Diff (UPDATE mode)
71	When re-researching a topic that already has a consensus reference, the pipeline auto-detects the previous research and offers UPDATE mode. In UPDATE mode:
72	- Full research runs identically to fresh mode (SCOUTs do NOT see previous consensus — no anchoring bias)
73	- TEMPORAL DIFF agent runs after SYNTHESIZER, comparing old and new consensus claim-by-claim
74	- Produces `_temporal_diff.md` with 5 categories: CONFIRMED / REVISED / CONTRADICTED / OBSOLETE / NEW
75	- Adds a `## Temporal Diff Summary` to the new consensus_reference.md
76	- Archives previous version as `_previous_consensus_[YYYY_MM].md`
77	- Stability score: CONFIRMED / total — tells you how much the field moved
78
79	CLI flags:
80	- `--update

People who liked this also viewed

6 picks

Type