feedback-synthesis

Synthesise feedback from $ARGUMENTS using the mandatory process below. Every step is required.

Step 1 — Ingest all feedback

Read every piece of feedback in the input. If pointed at files or directories, use Glob and Read to pull them. Count the total number of data points before proceeding.

For each piece of feedback, extract:

• Source — where it came from (support ticket, NPS survey, app review, social media, sales call, etc.)
• Date — when it was submitted (or "undated" if unknown)
• User segment — plan tier, role, tenure, or any available segmentation
• Raw quote — the user's exact words (preserve verbatim, do not paraphrase yet)
• Sentiment — positive, negative, neutral, or mixed

Step 2 — Categorise each data point

Assign exactly one primary category:

Category	Definition	Examples
Bug report	Something is broken or behaving incorrectly	"The export button doesn't work," "I get an error when..."
Feature request	User wants something that doesn't exist	"It would be great if...," "Can you add...," "I wish..."
Usability issue	Feature exists but is confusing, slow, or hard to find	"I couldn't figure out how to...," "It took me 20 minutes to..."
Praise	Positive feedback about something specific	"Love the new dashboard," "Support was incredibly fast"
Complaint	General dissatisfaction, often emotional, without a specific technical issue	"This is frustrating," "I'm considering switching"
Question	User doesn't understand something (signals a docs/onboarding gap)	"How do I...?," "What does X mean?"

If a single piece of feedback contains multiple categories (e.g., praise + feature request), split it into separate data points.

Step 3 — Build themes

Group related feedback into themes. Mandatory rules for theming:

1. Use the user's language — theme names must come from what users actually said, not internal jargon. "Can't find the export button" not "Export discoverability deficit." "Wish I could bulk edit" not "Batch operations enhancement."
2. Merge only when the root cause is the same — "slow dashboard" and "slow reports" are different themes unless they share a root cause. When uncertain, keep them separate.
3. Every theme needs 2+ data points — a single data point is not a theme, it's an outlier. List outliers separately.
4. Sub-themes are allowed — if a theme has 10+ data points, break it into sub-themes for clarity.
5. Preserve representative quotes — select 2-3 direct quotes per theme that best capture the sentiment and specifics. Choose quotes that are vivid and concrete, not generic.

Step 4 — Quantify and trend

For each theme, calculate:

• Count — total data points in this theme
• Percentage — of total feedback
• Category breakdown — how many are bugs vs. feature requests vs. complaints, etc.
• Sentiment balance — ratio of positive to negative within the theme
• Trend direction — if dates are available: increasing, stable, or decreasing over time. If dates are unavailable, state "trend unknown."
• Segment concentration — is this theme concentrated in a specific user segment? (e.g., "80% from enterprise users," "mostly new users in first 30 days")

Step 5 — Pattern detection

Scan for these specific patterns and flag any that apply:

Pattern	Detection rule	Action
Escalating issue	Theme count increasing over 3+ time periods	Flag as urgent, likely getting worse
Silent churn signal	Complaints + no feature requests from same users	These users have stopped asking and may leave
Onboarding gap	Questions or usability issues concentrated in new users	Recommend onboarding improvements
Power user friction	Feature requests from high-usage or enterprise users	Prioritise — these users are invested
Praise cluster	Praise concentrated on a specific feature	Protect this feature from regression, use in marketing
Bug-complaint bridge	Bug reports and complaints referencing the same area	Bug is causing broader dissatisfaction

Step 6 — Prioritise

Score each theme using:

Impact = Severity × Frequency × Segment weight

Where:

• Severity: Critical (4), High (3), Medium (2), Low (1)
• Frequency: use the count directly
• Segment weight: Enterprise/paid (1.5×), Free (1.0×), Unknown (1.0×)

Sort themes by impact score descending.

Step 7 — Output

Feedback summary

• Total data points analysed: [N]
• Date range: [earliest] to [latest]
• Sources: [list with counts]
• Overall sentiment: [positive/negative/mixed with percentages]

Theme table

Rank	Theme	Category	Count	%	Trend	Segment	Impact score	Representative quotes
1	...	...	...	...	...	...	...	"..."

Patterns detected

List any patterns from Step 5 that were found, with evidence.

Top 3 recommendations

Each recommendation MUST:

1. Name the theme it addresses — with the count and trend
2. State the specific action — not "improve X" but "add bulk edit to the contacts page" or "rewrite the export documentation with screenshots"
3. Provide the evidence — cite the quotes and counts that support this recommendation
4. Estimate the reach — how many users or what percentage of feedback this would address

Format:

### Recommendation 1: [action]
Theme: [theme name] ([N] data points, [trend])
Evidence: [2-3 quotes]
Reach: [N users / N% of feedback]
Rationale: [why this matters and why now]

Outliers

List any single data points that didn't fit into themes but are worth noting (novel ideas, early signals, unusual requests).

Rules

• Never discard feedback because it's inconvenient or contradicts a product direction. Report what users said, not what the team wants to hear.
• Quantify everything. "Users are unhappy" is not a finding. "23 of 45 feedback items (51%) express frustration with search" is.
• Distinguish between what users say they want (stated needs) and what their behaviour suggests they need (latent needs). Note the distinction when relevant.
• If the sample size is small (<30 data points), caveat conclusions explicitly: "Based on a small sample — validate before acting."

Related Skills

• /support:triage-tickets — triaged tickets are the primary input for feedback synthesis. Triage first, then synthesise.
• /support:write-kb-article — when synthesis reveals a recurring user problem, write a KB article to address it.