Guides adding AI-powered features to SaaS products: LLM API integration, RAG, AI assistants, prompt engineering, cost management, and patterns like smart drafts, summarization, categorization.
npx claudepluginhub whawkinsiv/solo-founder-superpowers --plugin solo-founder-superpowersThis skill uses the workspace's default tool permissions.
AI features should make your product 10x better at its core job, not be a marketing checkbox. This skill helps you choose the right AI pattern, manage costs, and ship AI features that users actually value.
Provides Ktor server patterns for routing DSL, plugins (auth, CORS, serialization), Koin DI, WebSockets, services, and testApplication testing.
Conducts multi-source web research with firecrawl and exa MCPs: searches, scrapes pages, synthesizes cited reports. For deep dives, competitive analysis, tech evaluations, or due diligence.
Provides demand forecasting, safety stock optimization, replenishment planning, and promotional lift estimation for multi-location retailers managing 300-800 SKUs.
AI features should make your product 10x better at its core job, not be a marketing checkbox. This skill helps you choose the right AI pattern, manage costs, and ship AI features that users actually value.
What: AI writes a first draft that users edit and refine.
Examples: Email drafts, report summaries, product descriptions, social posts.
Tell AI:
Add an AI draft feature to [describe where in the app].
When the user clicks "Generate draft," call the Claude API with:
- Context from [what data the AI should use]
- A system prompt that produces [describe the output format]
- User can edit the result before saving
Include: loading state, error handling, and a "regenerate" button.
Use the Claude API with the claude-sonnet-4-5-20250929 model.
What: AI condenses or interprets data the user has collected.
Examples: Meeting notes summary, customer feedback themes, dashboard insights.
Tell AI:
Add an AI summary feature that analyzes [data type].
Input: [describe the data — e.g., "all customer feedback from the last 30 days"]
Output: [describe what you want — e.g., "top 5 themes with supporting quotes"]
Display the summary in a card on [page name].
Cache the result so we don't re-call the API on every page load.
What: AI automatically labels or categorizes incoming data.
Examples: Support ticket routing, lead scoring, content tagging.
Tell AI:
Auto-categorize incoming [items] using AI.
Categories: [list your categories]
When a new [item] is created, call the API to assign a category.
Store the result in the database. Allow users to override.
Use the cheapest model that works (start with claude-haiku-4-5-20251001).
What: Users ask questions and get answers based on their own data.
Examples: "Search my documents," knowledge base Q&A, internal wiki search.
How RAG works (simplified):
1. User's documents → Split into chunks → Store as embeddings in vector DB
2. User asks a question → Convert to embedding → Find relevant chunks
3. Send relevant chunks + question to LLM → Get answer
Tell AI:
Add a Q&A feature where users can ask questions about their [data].
Use RAG (Retrieval-Augmented Generation):
- Embed their [documents/data] using [embedding model]
- Store embeddings in [Supabase pgvector / Pinecone]
- On query, retrieve top 5 relevant chunks
- Send to Claude with context for answer generation
Include: source citations, "I don't know" handling, loading state.
What: Suggest next actions or choices based on user behavior and data.
Examples: "Try this feature next," product recommendations, workflow suggestions.
| Model | Cost | Speed | Best For |
|---|---|---|---|
| Claude Haiku 4.5 | Cheapest | Fastest | Categorization, short responses, high-volume tasks |
| Claude Sonnet 4.5 | Medium | Medium | Most features — drafts, summaries, analysis |
| Claude Opus 4.6 | Highest | Slowest | Complex reasoning, multi-step analysis |
| GPT-4o mini | Cheap | Fast | Alternative to Haiku for simple tasks |
| GPT-4o | Medium | Medium | Alternative to Sonnet |
Rule of thumb: Start with the cheapest model. Only upgrade if quality isn't good enough.
Cost per request = (input tokens × input price) + (output tokens × output price)
Example (Claude Sonnet):
- Input: ~1,000 tokens ($0.003)
- Output: ~500 tokens ($0.0075)
- Cost per request: ~$0.01
1,000 requests/day = ~$10/day = ~$300/month
| Strategy | How |
|---|---|
| Use the cheapest model that works | Start with Haiku, upgrade only if needed |
| Cache responses | Same input = same output. Don't re-call |
| Limit output length | Set max_tokens to what you actually need |
| Batch requests | Combine multiple small requests into one |
| Rate limit per user | Prevent abuse with per-user daily limits |
| Use streaming | Better UX (users see progress) and same cost |
Free plan: 10 AI requests/day
Starter plan: 50 AI requests/day
Pro plan: 500 AI requests/day
Enterprise: Unlimited (with fair use policy)
Your prompts are your competitive advantage. Write them like product specs:
You are [role] helping [user type] with [task].
Context about this user:
- [User's plan/tier]
- [Relevant user data]
Rules:
- [Output format requirements]
- [Tone and style]
- [What NOT to include]
- [Length constraints]
Output format:
[Exact format you want]
Before shipping an AI feature:
- [ ] Clear user value defined (what does this save/improve?)
- [ ] Model selected (cheapest that meets quality bar)
- [ ] System prompt written and tested with 10+ examples
- [ ] Loading state shown while AI processes
- [ ] Error handling: API down, rate limited, bad response
- [ ] Fallback if AI is unavailable (manual mode still works)
- [ ] Usage limits per plan tier
- [ ] Cost monitoring set up (track spend per day/week)
- [ ] User can edit/override AI output (AI assists, user decides)
- [ ] Response cached where appropriate
| Mistake | Fix |
|---|---|
| Using the most expensive model for everything | Start with Haiku. Upgrade per-feature only when needed |
| No cost monitoring | Track API spend daily. Set billing alerts |
| No usage limits | Rate limit per user and per plan tier from day one |
| AI output shown as "truth" | Always let users edit/override. AI assists, humans decide |
| No loading state | AI calls take 1-10 seconds. Show a spinner or stream the response |
| Generic chatbot instead of focused feature | Build specific AI features tied to user workflows, not a general chat |
| No fallback when API is down | App should still work. AI features degrade gracefully |
| Hardcoded prompts with no iteration | Version your prompts. A/B test them. Iterate based on user feedback |