Trevin Chow
12 KiB
ce-riffrec-feedback-analysis
Turn raw Riffrec recordings into structured product feedback — quick bug reports for short captures, extensive analysis for longer ones, with handoff to
ce-brainstormwhen requirements emerge.
ce-riffrec-feedback-analysis is the product-feedback consumption skill. Riffrec is a separate capture tool that records synchronized screen + voice + event sessions and emits a riffrec-*.zip bundle. This skill is the consumption side: it analyzes those bundles (or any video / audio / meeting-notes file), routes between three paths based on length and intent, and produces the right artifact for what's actually inside the recording.
TL;DR
| Question | Answer |
|---|---|
| What does it do? | Analyzes a Riffrec zip (or video/audio/notes file), routes to setup / quick-bug / extensive-analysis path, produces structured feedback artifacts |
| When to use it | A riffrec-*.zip lands in chat; a video, audio, or meeting-notes file is shared as feedback; the user asks how to capture and share Riffrec sessions |
| What it produces | Quick: one concise bug report. Extensive: requirements-shaped analysis + handoff to /ce-brainstorm |
| Three paths | Setup (no recording yet), Quick bug report (under ~60s, single issue), Extensive analysis (longer, multiple issues) |
The Problem
Raw user-feedback recordings don't reduce to structured input automatically:
- Long recordings get ignored — a 12-minute walkthrough is too dense to act on without preprocessing
- Multi-issue recordings collapse to a single fix — the recording covers 4 distinct problems but only the first one gets attention
- Audio nuance lost in transcription — what the user said matters less than what they were trying to do; raw transcripts miss the intent
- Privacy bleed — raw screen captures and audio on disk get committed accidentally
- No bridge to building — the feedback exists, but nothing turns it into a brainstorm or plan
- Setup friction — the user wants to share feedback but doesn't know how to install the capture tool
The Solution
ce-riffrec-feedback-analysis runs a routed flow with three paths:
- Setup path — when the user has no recording yet and asks how to install / capture / share, this surfaces the Riffrec install guide and capture instructions
- Quick bug report — short recording, single issue, "just transcribe" intent → one concise bug report, no full artifact set
- Extensive analysis — longer recording or multiple issues → structured analysis with screenshots, requirements-shaped output, mandatory handoff to
/ce-brainstorm
The skill defaults raw recordings, audio chunks, zip contents, and extracted screenshots to local-only — no automatic commit. Text artifacts (analysis summaries, problem analyses) can be committed when traceability matters and there's no sensitive data.
What Makes It Novel
1. Three-path routing on length + intent
The path is chosen by what the input actually warrants, not by a flag:
- Setup — user has no recording yet, asks how to install Riffrec, capture a session, or share feedback. The skill loads
references/install-riffrec.mdand walks through it. - Quick bug report — under ~60 seconds, single issue, or the user explicitly asks for "quick", "small", or "just transcribe". Output is one concise bug report; the skill skips the full artifact set and brainstorm handoff.
- Extensive analysis — longer, multiple issues / requirements / workflow walkthroughs, or the user wants requirements material. Output is the full structured analysis with screenshots; the skill always continues into
/ce-brainstorm.
When the input is ambiguous (a zip arrived without context), the skill inspects recording length and event count before choosing. If still unclear, it asks the user.
2. Privacy-default for raw artifacts
Raw recordings, audio chunks, zip contents, session dumps, and extracted screenshot frames stay local-only by default. The raw/ and frames/ directories are not committed unless the user explicitly asks and confirms privacy is acceptable. Text/metadata artifacts (analysis summaries, problem analyses, source manifests) may be committed when needed for traceability and they contain no sensitive data.
Repo-relative screenshot paths in committed docs ensure later agents can open the evidence without absolute local paths.
3. Single analyzer entry point — accepts multiple input shapes
All non-setup paths share one entry point: python scripts/analyze_riffrec_zip.py /path/to/input. Accepted inputs:
- A Riffrec
.zipbundle - An
.mp4/.mov/.webmvideo - An
.m4a/.mp3/.wavaudio file - A meeting-notes
.md
This is what makes the skill useful beyond Riffrec — any recorded feedback fits the same pipeline.
4. Context-aware output location
In repos with docs/brainstorms/, the default output directory is docs/brainstorms/riffrec-feedback/ so the analysis lands where downstream skills (ce-brainstorm, ce-plan) expect to find requirements material. The quick path overrides to a temp location so it doesn't pollute the repo with single-issue bug reports.
5. Compound Engineering output format for extensive analysis
The extensive path produces output in the Compound Engineering feedback format documented in references/compound-engineering-feedback-format.md — structured to feed cleanly into /ce-brainstorm as raw input. Multiple issues / requirements / observations get separated, each with the relevant screenshot frames and timestamps. The brainstorm receives a starting point, not a transcript.
6. Mandatory ce-brainstorm handoff for extensive analysis
The extensive path always continues into /ce-brainstorm after the analysis lands. The recording captured what the user experienced; the brainstorm clarifies what to build in response. Without the handoff, the analysis sits on disk and nobody knows what to do with it.
The quick path skips the handoff — a single bug report is its own deliverable.
7. Lazy reference loading
The skill loads only the reference for the chosen path: install-riffrec.md for setup, quick-bug-report.md for quick, extensive-analysis.md for extensive. The other references stay unread, keeping the skill's runtime context small.
Quick Example
A teammate shares a riffrec-2026-05-04-checkout-flow.zip in your chat. You drag it in.
The skill detects a Riffrec zip, runs the analyzer to inspect length and event count: 8 minutes, 47 events, multiple distinct UI surfaces touched. It routes to extensive analysis.
The analyzer extracts: synchronized voice transcript, screen capture frames at event boundaries, event log with timestamps. It identifies 4 distinct issues: (1) the "Buy now" CTA hides on mobile, (2) form validation doesn't surface the error inline, (3) confirmation email subject line is unclear, (4) a confused "wait, why did it skip step 3?" moment that signals a flow gap.
It produces a structured analysis at docs/brainstorms/riffrec-feedback/2026-05-04-checkout-flow-analysis.md with each issue, the relevant frames, and timestamps. Raw recording stays local-only.
The skill loads /ce-brainstorm with the analysis as the starting point. Brainstorm clarifies which issue to address first, what success looks like, and produces the requirements doc.
When to Reach For It
Reach for ce-riffrec-feedback-analysis when:
- A
riffrec-*.ziparrives and you want to act on it - Someone shares a video, audio, or meeting notes as product feedback
- The user asks how to install Riffrec or capture a session for feedback
- A long walkthrough recording needs to become structured input for
/ce-brainstorm
Skip ce-riffrec-feedback-analysis when:
- The feedback is text-only and short — just paste it directly into
/ce-brainstorm - The recording is for a debug session, not feedback — handle the bug directly
- You just want to transcribe audio with no further structure — use a transcription tool, not this skill
Use as Part of the Workflow
The skill is a front-door entry point to the chain:
recording → /ce-riffrec-feedback-analysis → (extensive) → /ce-brainstorm → /ce-plan → /ce-work
→ (quick) → bug report (standalone)
→ (setup) → instructions for capturing
The extensive path always continues into /ce-brainstorm so the captured feedback becomes a real artifact downstream skills can use. The quick path produces a complete artifact on its own (a bug report) without forcing brainstorm overhead.
Use Standalone
The skill is most often invoked directly with a Riffrec zip or other input file:
- Riffrec zip —
/ce-riffrec-feedback-analysis riffrec-2026-05-04-checkout-flow.zip - Video file —
/ce-riffrec-feedback-analysis demo.mp4 - Audio file —
/ce-riffrec-feedback-analysis voice-memo.m4a - Meeting notes —
/ce-riffrec-feedback-analysis meeting-notes.md - Setup question —
/ce-riffrec-feedback-analysis "how do I install riffrec"(no input file; routes to setup path)
When the input is ambiguous (a zip arrived without context, or the path-routing signals are mixed), the skill inspects length and event count before choosing — and asks if still unclear.
Reference
| Argument | Effect |
|---|---|
<file path> |
Analyzes the file — Riffrec zip, video, audio, or meeting notes |
| Setup intent ("how do I install", "how to capture") | Routes to setup path with install instructions |
| Length + intent inferred from input | Routes to quick or extensive path |
Analyzer: scripts/analyze_riffrec_zip.py. Compound Engineering output format: references/compound-engineering-feedback-format.md.
FAQ
What's Riffrec?
A separate capture tool (github.com/kieranklaassen/riffrec) that records synchronized screen + voice + event sessions and emits a riffrec-*.zip. This skill is the consumption side — it doesn't capture, it analyzes captures.
Do I have to use Riffrec to use this skill?
No. The analyzer accepts videos (.mp4, .mov, .webm), audio (.m4a, .mp3, .wav), and meeting notes (.md) as well. Riffrec just provides a structured zip with synchronized event timestamps that produces richer analysis.
Why does the extensive path always continue into /ce-brainstorm?
Because the recording captures what the user experienced — that's evidence, not a decision. Without continuing into brainstorm, the analysis sits on disk as a transcript-with-screenshots and nobody knows what to build in response. The handoff is what makes the feedback actionable.
Why is the quick path different? A 30-second recording showing a single bug doesn't need the full artifact set or a brainstorm — the bug report itself is the deliverable. The user already knows what's wrong; the skill produces a concise report and stops.
What stays local vs gets committed? Raw recordings, audio, zip contents, frames stay local by default — privacy-first. Text artifacts (analysis summaries, bug reports) can be committed when traceability matters and the content is safe. Repo-relative paths in committed docs let later agents reference local screenshots without absolute paths.
What if the input is ambiguous? The skill inspects length and event count first. If still unclear, it asks the user which path applies. Better to ask one question than to run the wrong path.
See Also
ce-brainstorm— extensive-analysis output feeds directly into brainstormce-plan— downstream of brainstorm; receives the requirements doc that came from the recordingce-debug— when a quick-path bug report has a clear root cause to investigate- Riffrec — the capture-side tool (separate project)