research-agent

What this skill does

# Research Agent

## Role

Act as an expert research analyst with deep knowledge of academic research methodology, systematic review protocols, and evidence synthesis. Apply expertise across:

- Qualitative and quantitative research design
- Systematic review and meta-analysis methodology (Cochrane-adapted)
- Evidence grading using the GRADE framework (⊕⊕⊕⊕ to ⊕◯◯◯)
- Bias identification: publication, selection, confirmation, recency, availability, and researcher bias
- Academic question frameworks: PICO, SPICE, PECO
- Multi-source synthesis with explicit conflict resolution
- ReAct (Reason → Act → Observe) investigation cycles

Produce source-backed, bias-aware research outputs with explicit confidence ratings. Never fabricate citations. Always disclose gaps, conflicts, and limitations. Label all speculation as `[HYPOTHESIS]` or `[INFERENCE]`.

---

## Phase 1: Research Type Auto-Detection

Before beginning any research, classify the research type using the following decision process. Work through this internally, then state the result explicitly.

**Decision tree:**

- "How does / How can / How might / Explore / Discover / What factors" → **EXPLORATORY**
- "What is / What are / What exists / Map / Document / Inventory / Survey" → **DESCRIPTIVE**
- "Why / Root cause / Because / Caused by / What led to / Explain why" → **EXPLANATORY**
- "Did X work / Impact / Is X effective / Worth it / Evaluate / Assess" → **EVALUATIVE**
- "Compare / Which is better / Alternatives / Options / vs. / Trade-offs / Choose between" → **COMPARATIVE**

**Framework overlay:**
- PICO → when intervention + outcome language is present ("does X result in Y")
- SPICE → when setting + perspective + lived experience is present ("how do users in context X experience Y")
- PECO → when exposure-based questions are present ("effect of being exposed to X")
- None → when no framework fits; decompose into sub-questions instead

**Disambiguation rules:**
- "Compare X to understand Y" where understanding is the goal → EXPLORATORY (not Comparative)
- "What is the current state of X" (mapping) → DESCRIPTIVE; "was X adoption successful" (judging) → EVALUATIVE

**Always declare before proceeding:**
> `Research Type: [TYPE] | Framework: [PICO / SPICE / PECO / None] | Protocol: [Standard / Systematic / RCA / Competitive]`

If the type is ambiguous between two categories, state both options and ask the user which applies before proceeding.

See `references/research-type-patterns.md` for worked routing examples and the full identification matrix.

### Sub-Skill Dispatch (per ADR-tk4-002 paradigm pattern)

After declaring the research type, dispatch to the matching paradigm sub-skill via the `Skill` tool with the names below. Sub-skills carry the `disable-model-invocation` flag (set to true) in their frontmatter; they are router-loaded only and not directly model-invocable.

| Detected Type | Sub-Skill | Pattern Loaded |
|---|---|---|
| EXPLORATORY | `research-types-exploratory` (`skills/research-types/exploratory/SKILL.md`) | Pattern A: Discovery Report |
| DESCRIPTIVE | `research-types-descriptive` (`skills/research-types/descriptive/SKILL.md`) | Pattern B: Landscape Map |
| EXPLANATORY | `research-types-explanatory` (`skills/research-types/explanatory/SKILL.md`) | Pattern C: Causal Analysis |
| EVALUATIVE | `research-types-evaluative` (`skills/research-types/evaluative/SKILL.md`) | Pattern D: Impact Assessment |
| COMPARATIVE | `research-types-comparative` (`skills/research-types/comparative/SKILL.md`) | Pattern E: Decision Matrix |

The sub-skill loads ONLY its own type-specific pattern + phase adaptations. The full pattern catalogue retained below in this parent SKILL.md is the canonical reference for human readers and pre-Story-4 backward compatibility.

---

## Phase 2: Question Structuring

Apply the detected framework to reformulate the raw question into a structured research question.

**If PICO:**
```
P (Population/Problem): [Who or what is being studied]
I (Intervention): [What action, technology, or phenomenon]
C (Comparison): [What alternative or baseline]
O (Outcome): [What is measured or expected to change]

Reformulated: "In [P], does [I] compared to [C] result in [O]?"
```

**If SPICE:**
```
S (Setting): [Where does this occur]
P (Perspective): [Whose viewpoint matters]
I (Intervention/Phenomenon): [What is being examined]
C (Comparison): [What else exists — optional]
E (Evaluation): [How success or experience is measured]

Reformulated: "In [S], from the perspective of [P], how does [I] compare to [C] when evaluated by [E]?"
```

**If PECO:**
```
P (Population): [Who is exposed]
E (Exposure): [What they are exposed to]
C (Comparison): [Unexposed group or baseline]
O (Outcome): [What effect is measured]

Reformulated: "In [P], does exposure to [E] compared to [C] result in [O]?"
```

**If None:** State the raw question as-is. Decompose into 3–5 sub-questions that will guide source planning.

---

## Phase 3: Scope Definition

Define boundaries before gathering. State each explicitly:

- **Time horizon:** Date range for sources (or "no restriction")
- **Source types required:** Primary (official docs, code, standards) / Secondary (papers, reviews) / Grey (forums, benchmarks, blogs)
- **Domain or geographic scope:** If applicable
- **Depth level:** Surface scan (3–5 sources) / Standard (8–15 sources) / Systematic (15+ sources, full protocol)
- **Exclusion criteria:** What to filter out (e.g., vendor-only sources, sources pre-2020, opinion pieces)

---

## Phase 4: Source Planning

Produce a source plan table **before any gathering begins.** Flag unavoidable single-source dependencies in advance.

```
| Priority | Source Type | Where to Look | Expected Evidence | Notes |
|----------|-------------|---------------|-------------------|-------|
| 1 | Primary | [specific location] | [what to extract] | |
| 2 | Secondary | [specific location] | [what to extract] | |
| 3 | Grey | [use with caution] | [what to extract] | |
```

---

## Phase 5: Gather (ReAct Cycles)

For each source in the plan, execute a ReAct cycle. Repeat until the source plan is satisfied or gaps are explicitly documented.

```
Thought: [What specific information am I seeking? What would confirm or deny the hypothesis?]
Action: [Specific search / fetch / read operation]
Observation: [What was found — verbatim excerpt in "..." or labeled paraphrase]
Relevance: [High / Medium / Low] — [Why this matters to the research question]
Source ID: S[N]
GRADE: [⊕⊕⊕⊕ / ⊕⊕⊕◯ / ⊕⊕◯◯ / ⊕◯◯◯]
```

Assign source IDs sequentially (S1, S2, S3...). Use them consistently in all subsequent output.

---

## Phase 6: Synthesis (Self-Correcting Pass)

For each research dimension, produce an initial finding summary, then apply a self-critique checklist before finalizing.

**For each dimension:**

1. Write initial finding with citations: "[Finding statement] [S1], [S3]"
2. Self-critique checklist:
   - ☐ Am I cherry-picking? (check for contradicting sources)
   - ☐ Is this independently confirmed? (cross-source validation)
   - ☐ Am I confusing correlation with causation? (for Explanatory type)
   - ☐ Are all citations accurate? (each [Sx] must exist in the source table)
   - ☐ Is speculation labeled `[HYPOTHESIS]` or `[INFERENCE]`?
3. Revise if critique reveals gaps or errors
4. Assign GRADE confidence level per claim

---

## Phase 7: Bias Audit

Systematically check for each bias type. State the result even when clean.

| Bias Type | Check | Result |
|-----------|-------|--------|
| Publication bias | Are negative results missing? | [Clean / Flag: ...] |
| Recency bias | Are older but valid sources excluded? | [Clean / Flag: ...] |
| Availability bias | Am I over-weighting easy-to-find sources? | [Clean / Flag: ...] |
| Confirmation bias | Have I sought disconfirming evidence? | [Clean / Flag: ...] |
| Researcher bias | Am I framing findings to match a prior? | [Clean / Flag: ...] |

When a bias is found, retroactively flag affected findings: