AI that drafts replies to customer reviews for restaurant owners
You get
Contract: review text + rating → reply ≤ 80 words, in the review's language.
Eval case 7 (adversarial): review with a slur → no reply drafted; flag for owner.
Guardrail: never offer refunds or compensation in a draft.
About this skill
About AI feature spec
AI features that ship without defined quality, failure handling or evaluation end up embarrassing the brand. This skill writes a spec covering the user job, the input and output contract, quality criteria and an evaluation set of ten test cases — good, edge and adversarial. It defines guardrails: what the feature must refuse, privacy limits and where a human reviews. It plans fallbacks, cost limits and the metrics to watch after launch. It's the document AI product managers need before engineering writes the first prompt. To use it, paste the skill into your assistant and send the AI feature idea, who uses it and what a great answer looks like.
AI feature spec — example input and the first lines it returns. Works in ChatGPT, Claude and Gemini.
2 Paste it into a new chatChatGPT, Claude, Gemini or Copilot.
3 Send your inputSend the AI feature idea.
The skill — paste this into your AI assistant
# Skill: AI feature spec
You are running the "AI feature spec" skill for me. I work as: AI product manager. What I need from you: a spec for an AI feature covering the job, model behaviour, evaluation set, guardrails and fallbacks.
## When I send you the AI feature idea, who uses it and what a great answer looks like
1. Define the user job and the input → output contract.
2. Write quality criteria and an evaluation set of 10 test cases (good, edge, adversarial).
3. Define guardrails: what it must refuse, privacy limits and human review points.
4. Plan fallbacks, cost limits and the metrics to watch after launch.
## Reply in this format
- User job
- Input / output contract
- Quality criteria
- Evaluation set (table)
- Guardrails
- Fallbacks
- Launch metrics
## Rules
- Use only the facts I give you; mark anything you had to assume with [ASSUMPTION].
- If something essential is missing, ask me one short question, then continue.
- No preamble or filler: start with the result.
Confirm you're ready in one line, then wait for me to send the AI feature idea, who uses it and what a great answer looks like.
Then sendThe AI feature idea, who uses it and what a great answer looks like. For example: AI that drafts replies to customer reviews for restaurant owners
ChatGPT
Fastest: paste the skill as your first message in a new chat, then send your input.
To keep it: open a Project, choose Instructions (or Settings → Personalization → Custom instructions for every chat) and paste it there.
To share it with a team: create a custom GPT and paste the skill into its Instructions.
In the Codex app: copy the skill's folder into ~/.codex/skills/ (or Customize → Skills → Add → Create skill).
Claude
Fastest: paste the skill as your first message in a new chat, then send your input.
To keep it: create a Project, open Project instructions and paste the skill.
As a Claude Skill: download the skill's ZIP, then in Claude open Settings → Skills → Add → Upload skill and choose the file.
Gemini
Fastest: paste the skill as your first message in a new chat, then send your input.
To keep it: open Gems → New Gem, paste the skill into Instructions and save.
Microsoft Copilot
Fastest: paste the skill as your first message, then send your input.
In Microsoft 365 Copilot, save it as a prompt or paste it into an agent's instructions to reuse it.
Any other assistant
Paste the skill as the system prompt or first message in any chat assistant (Perplexity, Meta AI, DeepSeek, Mistral, local models).
Coding agents that read the Agent Skills format (SKILL.md) can load the downloaded file from their skills folder.
Install it permanently
Upload it to Claude, add it to ChatGPT's Codex app, clone it into Claude Code, or download the SKILL.md for any agent. Full guide with screenshots →
Download the ZIP and upload it in Claude. Claude then uses the skill on its own whenever a request matches.
Claude Code: save it as ~/.claude/skills/ai-feature-spec/SKILL.md.
Claude apps: use the Claude app tab — it has a ready-made ZIP.
Other agents: place the folder wherever your tool loads Agent Skills from.
Repositoryscienceofapps.com/skills/ai-skills.git · all 200 skills, one folder each, plain git over HTTPS
What it does
The steps it follows
Define the user job and the input → output contract.
Write quality criteria and an evaluation set of 10 test cases (good, edge, adversarial).
Define guardrails: what it must refuse, privacy limits and human review points.
Plan fallbacks, cost limits and the metrics to watch after launch.
What you get back
The reply format
User job
Input / output contract
Quality criteria
Evaluation set (table)
Guardrails
Fallbacks
Launch metrics
FAQs
Questions about AI feature spec
What does the AI feature spec skill do?
AI features ship without defined quality, failure handling or evaluation, then embarrass the brand. The skill fixes that: a spec for an AI feature covering the job, model behaviour, evaluation set, guardrails and fallbacks.
How do I install AI feature spec in ChatGPT, Claude or Gemini?
Copy the skill and paste it as the first message of a new chat, then send your input. To keep it, paste it into a ChatGPT Project or custom GPT, a Claude Project or a Gemini Gem, or upload the SKILL.md file as a Claude Skill.
What do I send AI feature spec after installing it?
Send the AI feature idea, who uses it and what a great answer looks like. If something essential is missing, the skill asks you one short question and then continues.
Is the AI feature spec skill free?
Yes. It's free to copy, use and adapt, with no sign-up, and works in the free versions of ChatGPT, Claude, Gemini and Copilot.
60 research-backed principles in 6 categories — what each one means, why it works on the human brain and how to use it on your own screens. Written by Sushant Singh, founder of Science of Apps, who has scaled consumer apps to lacs of users and spent 10+ years leading product across startups and digital businesses in the US, EU & India.
Free. No card, no paywall — just your email.
Free PDF · 90 pages · 2 editions
100+ Subscription Paywalls, decoded
110 real paywalls from top apps, taken apart, rated out of 5 and decoded with the Laws of UX — plus the frameworks, copy and checklists to ship your own. By Sushant Singh, founder of Science of Apps.
110 paywalls rated
34 Laws of UX
12 frameworks
Which paywall scored 4.8 out of 5, and which scored 2.6? It’s all inside.
