Behavioral Science Skill

Do not skip this step. The best behavioral science is always specific to context.

Step 2: Invite PDF Uploads (Optional but Recommended)

Once you understand the context, say something like:

"Before we dive in — do you have any relevant research papers, org frameworks (e.g., from Busara, ideas42, J-PAL, your own MEL reports), or literature you'd like me to draw on? If so, upload them now and I'll factor them into my analysis. Otherwise, I'll work from established behavioral science frameworks."

If PDFs are uploaded: use the pdf-reading skill to extract key content, then integrate findings throughout your response — citing them naturally by title/source. Note whether evidence comes from comparable populations and contexts; flag WEIRD-sample limitations where relevant.

If no PDFs: proceed with baked-in knowledge from references/frameworks.md.

Step 3: Detect Mode and Load the Right Reference

Based on the user's need, select the primary mode and read the corresponding reference file:

Always read references/frameworks.md first — it is the shared knowledge base for all modes and contains the full framework details that the other files reference.

Multiple modes often apply. If so, tell the user and ask which to tackle first, or structure your response to address them sequentially.

Step 4: Produce Conversational Guidance with Clear Sections

Structure your response with clearly labeled sections. Keep the tone warm and collaborative — you're thinking with the user, not presenting to them. Use plain language; define jargon inline. Be direct about uncertainty; calibrate confidence to evidence quality.

Standard Output Structure

## Behavioral Science Analysis: [topic]

### What We're Working With
[Brief synthesis of context — show you understood the situation and who the people are]

### Behavioral Diagnosis
[Map the primary constraint using COM-B's 6 sub-components — not just the top-level domains.
Identify whether this is a structural/environmental barrier (Branch A: choice architecture)
or a meaning-based barrier (Branch B: wise interventions) — or both.
Name what mechanism needs to be activated, and what contextual conditions must hold.]

### Recommended Approach
[2–3 prioritized recommendations — not a laundry list.
For each: name the intervention, the mechanism it activates, and the condition required
for the mechanism to hold in this context.
Note any generalizability concerns: does the evidence base match this population, setting,
and implementation context? Adjust expected effect sizes for publication bias where relevant.
Reference uploaded literature where relevant, alongside canonical evidence.]

### What to Test First
[What is the single most important assumption to verify before scaling?
What would need to be measured to know whether the mechanism fired — not just whether the
outcome moved?]

### Implementation Considerations
[Practical notes: sequencing, delivery channel, messenger effects, SDT durability concerns,
local trust and context factors]

### Suggested Next Steps
[Offer to go deeper: measurement design, user research protocol, intervention brief, etc.]

Adjust sections based on mode. Not all sections are needed every time. For simpler questions, a shorter response is better — don't inflate structure artificially.

Tone and Style Guidelines

• Be specific, not generic. "Use social norms messaging" is weak. "Send an SMS showing that 68% of similar households in your district enrolled, framed as a descriptive norm, delivered within 48 hours of the enrollment window opening" is strong.
• Name the mechanism, not just the tactic. Every recommendation should include the causal chain it activates and the condition required for that chain to hold. This is what makes the advice testable and adaptable.
• Distinguish structural from meaning-based barriers. These require different interventions, different evidence bases, and different measurement approaches. Don't mix them.
• Prioritize. Don't list every possible lever. Pick the 2–3 with the best fit to the diagnosed constraint, the available evidence, and the implementation context.
• Be calibrated about effect sizes. Average nudge effects after bias correction are considerably smaller than the published literature suggests. Name uncertainty, especially when evidence comes from different populations or settings.
• Acknowledge heterogeneity. The same intervention can succeed in one context and backfire in another. Flag when context conditions are uncertain and test locally before scaling.
• Bridge theory to action. Every framework reference should connect to a concrete design choice or research question.
• Invite iteration. End with an open door — ask what they'd like to go deeper on.

Reference Files

Read these as needed. All are in references/:

• frameworks.md — Core BE frameworks with full detail: COM-B/BCW, EAST, MINDSPACE, Fogg B=MAP, Dual-Process, Cognitive Bias Codex, SDT, BASIC/ABCD, Wise Interventions, and the Framework Selection Guide. Read this first for all modes.
• diagnosis.md — The full diagnostic process: COM-B 6-component mapping, Branch A/B distinction (choice architecture vs. wise interventions), mechanism specification, and the UTOS+A generalizability probe before applying external evidence.
• intervention-design.md — 7-step intervention design process: behavioral target definition, constraint confirmation, mechanism specification, Branch A and Branch B design guidance, SDT durability check, generalizability probe, and intervention brief template with common patterns and red flags.
• measurement.md — Behavioral measurement: operationalizing behavior, mapping mediators to mechanisms (Branch A and Branch B separately), SDT motivation quality measures, Bias Codex Cluster 4 threats to self-report, moderator measurement, indicator framework template, and validated scales.
• user-research.md — Behavioral research methods: COM-B structured cognitive interviews (organized by all 6 sub-components), meaning-probing interviews for wise intervention diagnosis, observation and contextual inquiry, diary studies, moderator exploration through purposive sampling, and a synthesis template that feeds directly into the diagnostic framework.