voc-research
Skill1 starsChanged 32 days agoMITView on GitHub
This skill should be used for voice-of-customer and qualitative conversion research: mining reviews, support tickets, chat logs or sales calls; designing and analysing on-site polls and post-purchase surveys; coding open-ended responses; planning and synthesising user tests; and building a customer-language bank for copy. Trigger phrases: "what are customers saying", "analyse these reviews", "code these survey responses", "why do people abandon", "customer objections", "survey questions", "user testing plan", "voice of customer", "VoC", "review mining".
In this repository7 other files it gives its agents.
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 44
Read ./cro-engagement.md first and add any research assets you find (review export, survey tool, ticket system) to its Data & access …
- Distinctive: line 51
Keep quotes verbatim and attributable to a source type (review, survey, ticket).
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Voice-of-Customer Research
Find out why visitors don't buy by asking and watching real customers, then turn what they say into evidence-tagged hypotheses and copy. This is the research layer the audit and the tests depend on.
Why this skill exists (evidence)
- AI is good at this part. GPT-4o coded survey open-ends at kappa 0.76 against human coders zero-shot and 0.94 fine-tuned (von der Heyde et al. 2025, https://arxiv.org/pdf/2506.14634). ChatGPT beat crowd-workers on 4 of 5 annotation tasks at about 1/20 the cost (Gilardi et al., PNAS 2023, https://arxiv.org/abs/2303.15056).
- AI is bad at replacing customers. Synthetic users were idealised and sycophantic next to real study data (NN/g 2024, https://www.nngroup.com/articles/synthetic-users/). Synthetic survey data had half the real variance and flipped the sign on a third of the relationships that differed (Bisbee et al., Political Analysis 2024). Never substitute generated personas or "simulated customers" for this research.
- Small qualitative samples go far. Five users find ~85% of usability problems; three rounds of five beat one of fifteen (Nielsen, https://nngroup.com/articles/why-you-only-need-to-test-with-5-users/). CXL/Speero's ResearchXL asks for 200+ survey responses for stable proportions.
Methods (pick by question and budget; details in references/methods.md)
| Question | Method | AI's job | Human's job |
| What stops people buying? | Exit-intent / on-page poll (1 open question) | Code and count answers | Choose the question and page |
| Why did buyers buy, and what nearly stopped them? | Post-purchase survey (5–8 questions) | Code, cluster, pull quotes | Approve survey; read a sample |
| What do customers already say unprompted? | Review, ticket, chat, return-reason mining (own + competitors) | Extract, code, quantify | Validate codebook on a sample |
| Where do people get stuck? | Moderated or unmoderated user test, 5 per round per segment | Draft tasks; summarise sessions | Run or watch sessions; judge severity |
| What words do they use? | Language bank from all of the above | Collect verbatim phrases by theme | Pick the ones copy will use |
Workflow
- Read the brief (
./cro-engagement.md): the surface, the funnel step that leaks, the ICP. - Start with what already exists: reviews, tickets, return reasons, past surveys. Mining costs nothing and needs no traffic.
- Add the cheapest new signal for the leaking step: a one-question poll on that page, or a post-purchase survey. Question bank in
references/methods.md. - Code with the protocol in
references/ai-coding-protocol.md: build a codebook from a first sample, have AI code a 50–100 item sample, have a human code the same sample, runscripts/agreement.py, refine until kappa ≥ 0.7, then code the full set. With under ~100 items in total, have a human review every AI code instead and call the themes exploratory. Thevoc-coderagent runs the AI half. - Quantify and rank themes: share of responses per code, split by segment (new vs returning, device, product line). Keep 2–3 verbatim quotes per theme.
- Turn themes into hypotheses: each one names the theme, its share, the page element it maps to, and the lever (Cost, Trust, Usability, Comprehension, Motivation). Hand them to
ecommerce-cro-auditfor prioritisation andexperiment-designfor testing. - Build the language bank (
references/codebook-template.mdhas the format) forconversion-copywriting.
Engagement context
Read ./cro-engagement.md first and add any research assets you find (review export, survey tool, ticket system) to its Data & access section.
Related skills
ecommerce-cro-audit (consumes themes as evidence), experiment-design (tests the hypotheses), conversion-copywriting (uses the language bank), report-information-design (presents findings).
Guardrails
- Report counts and shares with the sample size, never "customers feel" without a number.
- Keep quotes verbatim and attributable to a source type (review, survey, ticket). Don't paraphrase a quote and present it as verbatim.
- Strip personal data before sending feedback to any external tool.
- One survey answer is an anecdote; a theme needs volume or a consistent pattern across sources.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

