research-synthesis-loop.md

research-synthesis-loop.md

Theme Output Format (every theme, every level)

### Theme [N]: [the underlying need, solution-free, 4 to 8 words]

- **Frequency:** [recurring across sources / concentrated in one
  segment / isolated]
- **Verbatim:** "[real quote]" — [source, role, date]
- **Verbatim:** "[real quote]" — [source, role, date]
- **Who feels it:** [role/segment — labeled Fact or Inference]
- **What contradicts it:** [counter-evidence, or "none found"]
- **Confidence:** [Fact / Inference / Assumption — one label]
- **So what:** [one sentence: what this means for the decision]

Level 0 — Unseasoned (works today)

/goal Synthesize these eight interview transcripts into key themes and product implications.

You'll get themes. You won't know which transcript carried which theme, which themes are one loud customer, or what got averaged away.

Level 1 — Loop Lingo (the synthesis contract)

/goal Synthesize this research into evidence-cited need themes I can
defend in a roadmap conversation — every theme in the Theme Output
Format.

Base camp (calculate once): research goal, the decision this
informs, personas in scope, and what we believed BEFORE the research
(so contradictions surface as findings, not formatting problems).

Index before you search (build once; do not rescan transcripts
unless I say "refresh"):
1. Sources: who, role, date.
2. Candidate theme categories, one-line definition each.
3. Which sources touch which category.
4. Open questions the material may not answer.

/loop One category per turn.

For each category:
1. Cheap check: does more than one source touch it? One-source
   themes get processed but flagged "isolated" — never promoted to
   headline findings.
2. Pull forward only this category's mapped evidence.
3. Render the theme in the Theme Output Format — every field, real
   verbatims with attribution, one confidence label.
4. Ask me: "Modify, or continue to [next category]?" Wait.
5. On approval, compact: carry the theme and its strongest verbatim
   forward — not the transcripts.

Loop controls:
- No cross-category conclusions until all categories are settled.
- Evidence contradicting the base-camp beliefs gets flagged loudly:
  that IS the finding.
- After the last category, and only then, produce the cross-cutting
  synthesis: top 3 needs, what changed our minds, top 3 assumptions
  to validate next.

Level 2 — Just Enough Jinja2 (the routine form)

{% set routine_version = "research-synthesis-v1.0" %}

{% for category in category_index %}
## Category {{ loop.index }} of {{ category_index | length }}:
{{ category.name }}

Use only this category's mapped evidence: {{ category.sources }}.

Render in the Theme Output Format:

### Theme {{ loop.index }}: [underlying need, solution-free]

- **Frequency:** [recurring / concentrated / isolated]
- **Verbatim:** "[real quote]" — [source, role, date]
- **Verbatim:** "[real quote]" — [source, role, date]
- **Who feels it:** [role/segment — Fact or Inference]
- **What contradicts it:** [counter-evidence, or "none found"]
- **Confidence:** [Fact / Inference / Assumption]
- **So what:** [one sentence]

Ask "Modify, or continue?" and wait before the next category.
{% else %}
STOP: no category index exists. Build the index before synthesis.
{% endfor %}

## Cross-cutting synthesis (only after all categories approved)
- Top 3 needs, ranked, with their strongest verbatims
- What changed our minds vs. the base-camp beliefs
- Top 3 assumptions to validate next

Generated by {{ routine_version }} from
{{ category_index | length }} categories,
{{ source_count }} sources.

Final Step

Offer exactly 4 next options:

1. Produce the cross-cutting synthesis and top opportunity hypotheses (Recommended)
2. Convert the strongest theme into a JTBD canvas
3. Draft the "what changed our minds" note for stakeholders
4. Generate follow-up interview questions for the weakest evidence

Ask the user to reply with 1, 2, 3, 4, 1 and 2, or a custom path.