Validate a SaaS idea in 2 days without writing code by running a simple, repeatable “smoke test”: Confirm the problem is real by finding consistent complaints from your target users (Reddit). Check whether existing solutions are weak or expensive (Google + reviews). Publish a minimal landing page that clearly explains your promise and collects signups. Drive traffic from one primary channel (one Reddit post or one community). Measure results over 7 days and decide: Build if signups 50 Pivot if signups < 10 Improve and re test if 10 to 49 (change one variable only) This workflow is designed to be reproducible : each step has a clear goal, timebox, concrete outputs, and defined ownership between a Human and AI agents.

End to end workflow for building a production FAQ chatbot using Retrieval Augmented Generation (RAG). Covers knowledge base preparation, hybrid retrieval pipeline (BM25 + bi encoder + cross encoder reranking), LLM generation with input/output guardrails, multilingual golden dataset creation, and automated evaluation. When to use: You need an AI powered FAQ system that retrieves answers from an existing knowledge base (help center, docs, wiki) and generates natural language responses with safety guardrails. Preconditions: Existing FAQ or help center content (articles, docs, or structured Q&A) Access to an embedding model and an LLM (open source or API based) A vector database or search index Domain volunteers or annotators for golden dataset labeling (if multilingual) Key tuning points: Chunk size and overlap (step 1) — adjust based on article length distribution Model selection (step 3) — open source multilingual (e.g. Mistral &lt;8B) vs. proprietary tradeoff Retrieval weights (step 5) — BM25 vs. semantic score balance Evaluation thresholds (step 7) — minimum aggregate score for go/no go Target languages (step 4) — which languages to include in golden dataset Evaluation formula: Aggregate score = (answer relevancy + faithfulness + no hallucination + prompt alignment) / 4 Reference: Based on patterns from Amex GBT Egencia's FAQ RAG system (470K+ conversations/year). See the 3 part series on the Amex GBT Technology blog for detailed implementation context.

Generate a professional client report every month — with revenue, project status, traffic insights, trends, and strategic recommendations — in under 10 minutes. Claude pulls your raw data from Stripe, Airtable, and Google Analytics, then generates a complete PDF ready report with executive summary, month over month comparisons, charts, and actionable recommendations. Add a client email, hit send on the 1st. Done. What this replaces: The 3 hour monthly scramble of logging into dashboards, copying numbers into a doc, making charts, and writing "everything is going well" with no real insights. Or the $2,000/month retainer for someone who does exactly this. What you need to start: A Stripe account (for revenue data) An Airtable base (for project/deliverable tracking — or any project tool you use) Google Analytics (for traffic and marketing data) Access to Claude A way to create PDFs (Google Docs → Export as PDF works fine) How it works: One time setup ( 10 min): Define your report structure and let Claude build your reusable template + data export checklists Monthly execution ( 10 min): Pull data from 3 sources (5 min) → Claude generates the full report + client email (5 min) → review and send The report includes: | Section | What it covers | | | | | Executive Summary | 3 bullets: biggest win, key number, top recommendation | | Revenue | Total, MRR, new vs recurring, growth % (from Stripe) | | Projects | Completed, in progress, blocked, milestones (from Airtable) | | Traffic | Sessions, top channels, conversions, top pages (from GA) | | Month over Month | Side by side with trend arrows | | Key Wins | 2 3 highlights the client should feel good about | | Watch List | Things that need attention before they become problems | | Recommendations | 2 3 specific actions for next month | No coding required. Data export is copy paste. Report generation is one prompt. Pairs with: The full lead to client pipeline — AI Lead Qualification → Email Drip → Appointment Booking → this workflow for ongoing client reporting.

Find gaps in your documentation by cross referencing support tickets against your knowledge base — then draft the missing articles ranked by customer impact. Before You Start You need the following to run this workflow: Claude Code (desktop app, CLI, or IDE extension). Claude analyzes tickets and drafts articles during Steps 3, 4, and 5. Exported support tickets — a CSV or spreadsheet export from your helpdesk (Zendesk, Intercom, Freshdesk, HubSpot, or similar). You need at least the ticket subject, description or first message, and any tags or categories. A month's worth of data is ideal. Your knowledge base content — either a folder of markdown/HTML files, a sitemap URL, or a list of article titles and URLs that Claude can visit. 30 45 minutes per run. The first run takes longer because you are calibrating. Subsequent weekly runs are faster. What this workflow does not do: It does not connect directly to your helpdesk API or publish articles automatically. You export tickets, Claude analyzes them, and you publish the drafts through your normal editorial process. This keeps a human in the loop for every published article. Quick Reference | Step | Who | What you do | What you get | | | | | | | 1 | You | Export tickets, index your knowledge base | Clean inputs ready for analysis | | 2 | You | Define your documentation style and structure | Style template Claude follows | | 3 | Claude | Cross reference tickets against docs, find gaps | Prioritized gap report + stale article list | | 4 | Claude | Draft articles for the top missing topics | Review ready article drafts | | 5 | Claude | Draft updates for outdated articles | Updated versions with change summaries | | 6 | You | Review, publish, and measure ticket deflection | Live articles reducing support load | Going Further Once this workflow is running regularly: Automate the export. Most helpdesks support scheduled CSV exports or API access. Set tickets to export automatically so your Monday morning starts with fresh data, not a manual export. Track deflection rates. Build a simple dashboard: articles published per week vs. ticket volume for those topics. This is the ROI proof that justifies continued investment in documentation. Expand to onboarding. The same gap analysis works for onboarding flows — cross reference new user support tickets against your getting started guides to find where new users get stuck. Feed insights back to product. If the same feature generates tickets month after month despite good documentation, the problem is not the docs — it is the UX. Share those patterns with your product team.

Agent Package is a shared system for turning successful AI work into reusable operational assets. It gives users two core building blocks: Workflow Pack (reusable execution) and Context Pack (portable understanding). Designed for personal use first, but built to support team sharing and community distribution. Core idea: useful AI work should not disappear into chat history. It should become portable, inspectable, reusable, and continuously improvable.

A pressure test for an unvalidated idea before any building begins. Six forcing questions, asked one at a time, that surface hidden assumptions, refuse polite agreement, and convert vague conviction into a small, falsifiable next step. Why this pack exists Most ideas die from unexamined assumptions, not from bad execution. When a builder pitches an idea — to a friend, a partner, an AI assistant — the default response is some version of "interesting!" That politeness is the most expensive thing in early stage product work. It lets bad ideas survive long enough to consume months of building. This pack replaces the polite mirror with a structured interrogation. The six questions are sequenced so each one closes off a common escape route: vague users, hypothetical workarounds, missing "why now," unexamined beliefs, optimism bias, and the urge to start coding instead of testing. Core idea Specificity is the evidence of understanding. A builder who can name one real person, one real moment, and one real workaround has done the work. A builder who speaks in plurals and abstractions has not. Every question in this pack pushes relentlessly toward the singular and concrete. Who this is for Anyone holding an idea that has not yet met reality — solo builders, side project starters, founders early in formation, internal product owners pitching a new initiative. Not for products that already have paying customers. Not for technical implementation questions. Not for emotional support sessions. When to use it Load this pack when the user signals an unvalidated idea: "I'm thinking about building...", "is this worth doing?", "what do you think of this?", or any pitch of a half formed concept. Do not load when there are already paying customers, when the question is about implementation, when the user is venting, when the decision is irreversible, or when this same idea has been stress tested in a recent session. Expected duration 20–40 minutes of back and forth. Faster usually means worse. The output is a written diagnostic with a single verdict: GO, VERIFY, RECONSIDER, or KILL — plus a 7 day falsifiable test.

A fat, AI ready playbook for turning messy customer evidence into decision quality insight. Based on the pm skills market research methodology: segmentation, personas, customer journey mapping, competitor analysis, market sizing, sentiment analysis, and feedback synthesis. Each block is a discrete skill — use them together for a full research cycle or independently as targeted methods. The goal is always to produce research that changes a decision, not research that fills a slide deck.

A pack for organizing the inside of a hard decision. Not for judging whether the decision is right — for making the shape of the decision visible so the person can decide for themselves. Why this pack exists Most decisions feel hard not because the answer is genuinely difficult, but because everything is tangled. The choices, the reasons behind each choice, what is known versus assumed, what is reversible versus permanent, what is being avoided by not deciding — all of these sit in the same fog at once. The fog is the problem, not the choice. This pack pulls each strand out of the fog and lays it on the table. By the end, the person sees their decision the way an outside observer would: choices named, beliefs labeled, regrets imagined, costs of waiting acknowledged. Often the answer becomes obvious once the picture is clear. When it doesn't, at least the person knows exactly what they are uncertain about, and what would change their mind. Core stance The assistant is a structuring partner, not a judge. The pack does not produce a verdict. It produces a clear page that the person walks away holding. Whether they decide today, next week, or never is their business — the assistant's job is to make sure the shape of the situation is no longer hidden from them. Who this is for Anyone facing a decision that has been sitting unresolved — a career move, a hire or fire, a pivot, a price change, an investment, a relationship choice, whether to keep or kill a project, whether to take an opportunity. Works for personal and professional decisions, big and small. How this differs from "Idea Stress Test" Idea Stress Test asks "is this worth building?" and presses hard for a verdict. Decision Clarifier asks "what actually is this decision?" and stays neutral. Use Idea Stress Test when the user is pitching an idea and wants pressure. Use Decision Clarifier when the user is stuck and wants clarity. When to use it Load this pack when the user signals stuckness around a choice: "I can't decide whether to...", "I'm torn between...", "I keep going back and forth on...", "should I...", "I've been sitting on this for weeks." Also useful when someone is about to make a quick decision and you want to slow them down enough to see what they're actually choosing. When not to use it Do not load when the decision is already made and the person just needs help executing. Do not load when the person is clearly venting and not seeking structure. Do not load when the question is factual ("which framework should I use?") rather than a real choice with stakes. Expected duration 15–30 minutes of dialogue. Output is a one page summary of the decision's shape. The user keeps it, returns to it, updates it. This pack is designed to be re runnable on the same decision as new information arrives.