Research Decision Room Agent Skill

Research Decision Room Skill

Create a single-page HTML decision artifact that helps a product or design team

turn messy evidence into a clear next move. The output is not a decorative

research deck. It is a working room for debate: evidence, themes, confidence,

tradeoffs, and recommended experiments stay visible together.

Resource map

research-decision-room/
├── SKILL.md
├── example.html
└── references/
    ├── checklist.md
    └── evidence-model.md

Read references/evidence-model.md before synthesis and run

references/checklist.md before emitting the artifact.

When to use this skill

Use this skill when the user has any mix of:

• Interview notes, usability-test observations, support tickets, sales call notes,

app-store reviews, NPS comments, survey open text, analytics snippets, or

product-decision context.

• A decision that needs evidence: "Should we build X?", "Which onboarding path

should we try?", "Why are users dropping off?", "What do customers actually

mean by slow?"

• A need to share findings with stakeholders who will not read a long research

report.

Do not use it for pure visual inspiration, campaign ideation, or brand moodboards.

Workflow

Step 1 - Establish the decision frame

Identify the decision scope from the user's prompt. If the user did not give a

decision, derive one from the evidence and label it as inferred.

Write a short frame with:

• Decision question.
• Audience or segment.
• Time horizon.
• Known constraints.
• What this artifact will not decide.

If key context is missing and the task is not blocked, proceed with labelled

assumptions instead of asking a broad question.

Step 2 - Build the evidence ledger

Normalize every useful signal into ledger rows using the model in

references/evidence-model.md.

Each ledger row must include:

• id: short stable id, such as I-03, T-14, M-02.
• source_type: interview, usability, support, survey, analytics, sales, field

note, or stakeholder.

• segment: user type or "unknown".
• signal: one-sentence observation.
• quote_or_metric: direct quote, metric, or "not provided".
• strength: strong, medium, or weak.
• limitations: why this evidence may be biased or incomplete.

Never invent quotes, participant counts, dates, revenue impact, or metrics. If

the user did not provide a number, use "not provided" and explain what evidence

would increase confidence.

Step 3 - Synthesize themes and tensions

Cluster evidence into 4 to 6 themes. For each theme:

• Name the theme in plain human language.
• List the evidence ids that support it.
• Explain the behavior behind it, not just the UI complaint.
• Mark confidence as high, medium, or low.
• Note contradictions or segment differences.

Prefer verbs over nouns: "Teams abandon setup when the first blank state asks

for too much" is better than "Onboarding problem".

Step 4 - Score opportunities

Create an opportunity matrix with 3 to 5 options. Score each option on a 1 to 5

scale:

• Evidence strength.
• User pain.
• Business leverage.
• Implementation risk, where 5 means low risk and 1 means high risk.

Show the total score, but do not let the score replace judgment. Add one sentence

on why the top recommendation wins.

Step 5 - Draft the decision memo

Write a decision memo with:

• Recommended move.
• Why now.
• What evidence supports it.
• What could be wrong.
• What to measure next.
• Reversible next step.

Keep the memo short enough to read in under one minute.

Step 6 - Create the HTML artifact

Produce a self-contained index.html. Use the active DESIGN.md for typography,

spacing, color roles, and component tone, but keep the information architecture

stable:

• Header with decision question, confidence, and last-updated label.
• Executive readout with recommendation, risk, and next experiment.
• Evidence ledger with filter chips.
• Theme map with evidence ids and confidence.
• Opportunity matrix.
• Decision memo.
• Experiment queue with owner, metric, and success threshold.
• Assumptions and limitations.

The artifact should be interactive but durable. Simple vanilla JavaScript is

allowed for filtering evidence, switching views, or highlighting related ids.

No framework dependency is required.

Step 7 - Self-check and emit

Run the checklist. Then emit one concise orientation sentence and one HTML

artifact:

<artifact identifier="research-decision-room" type="text/html" title="Research Decision Room">
<!doctype html>
<html>...</html>
</artifact>

Nothing after the closing </artifact>.