102 pieces of feedback + 1 prompt = perfect product insights

_{Day 274/100}

Hey—It's Tim. 

Here's a problem I hope you have:
It’s one I certainly do after you all sent your product feedback for the new competitor gap analysis.

You launch a feature.

You get feedback.

Lots of feedback.

102 pieces of feedback, in my case.

Some in email replies. Some in Slack DMs. Some in that Google Form you set up. Some in random Twitter replies you almost missed.

And you need to know:

• What are the actual patterns here?

• What's working? What's broken?

• Are there any "holy shit" testimonials I can use?

• What edge cases am I missing?

You could read all 102 pieces manually.

Take notes. Categorize. Look for themes.

Take 4 hours.

Miss half the patterns because you're human and humans are bad at spotting patterns in 102 unstructured data points.

Or you could do what I'm about to do:

Feed them all into Claude with one prompt.

Get back patterns, testimonials, edge cases, and themes in 90 seconds.

With none of the human bias that makes you see what you want to see instead of what's actually there.

The Prompt I Built

I'm running this tonight on our new competitor analysis feature.

102 pieces of feedback from the first week.

But this isn't just "hey Claude, summarize this."

That's what everyone does.
And it gives you garbage.
Generic summaries. Made-up patterns. Testimonials that sound like ChatGPT wrote them.

I built a proper prompt.
With validation checks. With severity definitions. With minimum thresholds.

So Claude can't bullshit me.

If you want to know how I write prompts, and why I write them the way I do, I have a whole guide on it here.

<SYSTEM_ROLE>
You are a product feedback analyst specializing in qualitative data synthesis.

Priorities: Accuracy > Completeness > Insight generation > Speed

Translation:
- Never invent patterns that aren't clearly supported by the data
- Surface all significant themes even if they don't fit neatly into 5 categories  
- Provide actionable insights, not just summaries
- Take the time needed to analyze thoroughly
</SYSTEM_ROLE>

<OBJECTIVE>
Analyze 102 pieces of customer feedback about [FEATURE NAME] and extract actionable intelligence for product and marketing decisions.

Deliverables:
1. 5-7 dominant themes with frequency counts and implications
2. 5 marketing-ready testimonials that meet specific criteria
3. 3-5 unexpected use cases that reveal new opportunities
4. Prioritized list of problems/friction points
5. Grouped feature requests with demand signals

Success criteria:
- Every theme is supported by at least 8 pieces of feedback (8% threshold)
- Every testimonial includes specific benefit language and measurable impact
- Every problem is linked to specific feedback with clear user impact
- Feature requests show clear patterns (not one-off wishes)
</OBJECTIVE>

<CONTEXT>
Purpose: This analysis will inform Q1 product roadmap and marketing messaging for [FEATURE NAME].

Constraints:
- We can only tackle 2-3 major problems in Q1
- Marketing needs testimonials within 48 hours for campaign launch
- We're specifically looking for use cases outside our core target persona (B2B SaaS marketers)

Background:
- [FEATURE NAME] launched 6 weeks ago
- Current user base: 450 active users
- These 102 pieces represent 23% response rate from all users
- Feedback was solicited via email survey + in-app NPS
</CONTEXT>

<PLANNING>
Analysis approach:

Step 1: First pass - Categorization
- Read all 102 pieces without analysis
- Tag each piece with initial categories (create new categories as needed)
- Don't force feedback into predetermined buckets

Step 2: Theme identification  
- Count frequency of each category
- Merge similar categories
- Keep themes that appear in 8+ pieces (8% threshold)
- Rank by frequency

Step 3: Quote extraction
- Identify all quotes mentioning specific benefits
- Score each quote on specificity (0-10) and emotional intensity (0-10)
- Select top 5 by combined score

Step 4: Pattern analysis
- Look for use cases mentioned by <3 people (signals we missed)
- Identify problems mentioned across multiple themes
- Group feature requests by functional area

Step 5: Prioritization
- Rank problems by: frequency × severity
- Rank feature requests by: frequency + strategic fit
</PLANNING>

<OUTPUT_FORMAT>

## 1. DOMINANT THEMES

For each theme, use this structure:

**Theme: [Clear, specific name]**
- **Frequency:** X out of 102 pieces (X%)
- **What they said:** 2-3 representative quotes in their exact words
- **Product implication:** One specific, actionable insight
- **Confidence level:** High/Medium/Low based on consistency of feedback

Sort themes by frequency (highest first).

---

## 2. MARKETING TESTIMONIALS

For each testimonial:

**Quote:** "[Exact words]"
- **Author context:** [Role/company size if mentioned, otherwise "Anonymous user"]
- **Specific benefit:** [What concrete outcome they achieved]
- **Emotional signal:** [Excited/Relieved/Surprised + evidence from language]
- **Why this works:** [One sentence on marketing value]

Sort by marketing value (strongest emotional + specific first).

---

## 3. UNEXPECTED USE CASES

For each use case:

**Use case:** [Descriptive title]
- **Who:** [User type/persona if identifiable]
- **What they're doing:** [Specific workflow described]
- **Why it's unexpected:** [How this differs from intended use]
- **Opportunity:** [What this reveals about market or product expansion]
- **Frequency:** Mentioned by X people

Sort by opportunity size (biggest market signal first).

---

## 4. PROBLEM AREAS

For each problem:

**Problem:** [Specific user pain point]
- **Frequency:** X out of 102 pieces (X%)
- **User impact:** [How this affects their workflow - quote evidence]
- **Severity:** Critical/High/Medium/Low
- **Effort to fix:** [Your assessment: Quick win/Medium lift/Major project/Unknown]
- **Priority score:** [Frequency × Severity on 1-10 scale]

Sort by priority score (highest first).

Severity definitions:
- Critical = Blocks core workflow, causes users to abandon feature
- High = Significant friction, users complain but work around it
- Medium = Annoying but doesn't prevent use
- Low = Polish issue, mentioned but not impactful

---

## 5. FEATURE REQUESTS

Group similar requests, then for each group:

**Request category:** [Functional area]
- **Specific requests:** [List unique variations]
- **Demand signal:** X out of 102 pieces (X%)
- **Use case:** [Why users want this - what job it solves]
- **Strategic fit:** [Does this align with product roadmap? Yes/No/Maybe]

Sort by demand signal (highest first).

---

## SUMMARY METRICS

- Total pieces analyzed: 102
- Themes identified: X
- Average sentiment: [Positive/Mixed/Negative based on tone]
- Most requested feature: [Feature with highest %]
- Most critical problem: [Problem with highest priority score]

</OUTPUT_FORMAT>

<VALIDATION>

Before submitting analysis, verify:

**Check 1: Theme frequency accuracy**
- Count supporting quotes for each theme manually
- Minimum 8 pieces required per theme
- If a "theme" has <8 pieces, move it to "minor patterns" section or merge with related theme

**Check 2: Testimonial quality**
- Each testimonial must include at least one concrete, measurable benefit
- Each testimonial must be under 25 words
- Each testimonial must show clear emotional language (exclamation marks, intensifiers, surprise words)
- If a quote doesn't meet all three criteria, replace it

**Check 3: Problem severity justification**
- "Critical" severity requires evidence that users abandoned or avoided the feature
- "High" severity requires evidence of significant workflow disruption
- If severity level isn't supported by evidence, downgrade it

**Check 4: No invented insights**
- Every claim must be traceable to specific feedback
- No synthesis that goes beyond what users actually said
- If you catch yourself inferring beyond the data, flag it as an assumption

**Check 5: Completeness**
- Did any significant patterns get missed? (Scan for outliers)
- Are there quotes that mention multiple themes? (They should be tagged to all relevant themes)
- Did you analyze all 102 pieces? (If you skipped any, note why)

If any check fails: Revise before submitting.
</VALIDATION>

<ASSUMPTIONS_LOG>

At the end of your analysis, include:

**Assumptions made during analysis:**

For each assumption, state:
- What was ambiguous in the feedback
- What assumption you made to proceed
- How this assumption might affect accuracy
- What additional data would remove this assumption

Examples of valid assumptions to log:
- "Assumed 'it's too slow' refers to processing time, not learning curve"
- "Grouped 'export to PDF' and 'download as PDF' as same request"
- "Interpreted excited emojis as positive sentiment even when no explicit praise"
- "Assumed complaints about 'the UI' refer to [FEATURE NAME] specifically, not entire product"

</ASSUMPTIONS_LOG>

---

Now, here's the feedback to analyze:

[PASTE ALL 102 PIECES]

What Makes This Prompt Actually Work

Most people treat AI like a magic box:

"Here's my data. Tell me what's important."

And AI goes: "Sure! Here are 5 patterns I just invented."

This prompt has guardrails:

1. Minimum frequency thresholds

Every theme needs at least 8 pieces of supporting feedback (8% of total).

If something only shows up 3 times? That's not a pattern. That's 3 people.

2. Validation checks

Before Claude can submit the analysis, it has to verify:

• Did you count the frequency correctly?

• Does every testimonial have a measurable benefit?

• Is every "critical" problem actually critical based on the evidence?

• Did you invent any insights that aren't in the data?

3. Severity definitions

"This is a problem" could mean anything.

So the prompt defines:

• Critical = Users abandon the feature

• High = Significant friction, users complain

• Medium = Annoying but doesn't prevent use

• Low = Polish issue

Claude has to justify the severity level with evidence.

4. An assumptions log

At the end, Claude has to list every assumption it made.

"Assumed 'too slow' means processing time, not learning curve"

"Grouped 'export to PDF' and 'download as PDF' as same request"

So I can see where the analysis might be shaky.

5. Testimonial criteria

Can't just be positive.

Must be:

• Under 25 words

• Include concrete, measurable benefit

• Show emotional language (exclamation marks, intensifiers, surprise words)

No "it's good" testimonials.

Only "holy shit this saved me 3 hours" testimonials.

I’m running this tonight.

So, if you wanna get your own competitor gap analysis, reply to the email with your website and the topic you’re focusing on. Then when you get the results lemme know your feedback and I’ll add it to the pile

✌️ Tim "Pattern Recognition > Content Generation" Hanson
CMO @Penfriend.ai

Same brain, different platforms: X, Threads, LinkedIn.

P.S. The validation checks are the key

Without them, your AI will just make shit up.

"Here are 5 patterns!"

(Based on 2 pieces of feedback each and a lot of vibes)

With them, your LLM has to prove every pattern with:

• Minimum 8 pieces of supporting evidence

• Frequency count

• Actual quotes

• Justified severity levels

That's the difference between useful analysis and bullshit.

P.P.S. One more day of comp gap analysis.
Shoot me over your website, and your current writing focus and I’ll tell you what to write.

Penfriend.ai
Made by content marketers. Used by better ones.