产品经理用 AI 编程工具打造的 daily-ai-builds 小工具集

Small, working AI tools built by a product manager. One problem per build, each with a live app or worked examples you can open without an API key.

I'm Prerna Kakkar, a Senior Product Manager with 10+ years across conversational AI, API-first SaaS and analytics. Each build starts from a problem I've seen at work or in daily life. I write the brief, make the product and design calls, set the guardrails and review every version. The code is written by AI coding tools working from my prompts and review.

Latest

Which Tier? · live app Sort an AI feature under the EU AI Act, with the dates as amended by the 2026 AI Omnibus. You describe the feature and answer chip questions on your role, banned practices, regulated products, Annex III uses, profiling and transparency. Printed rules set the tier, show the rule trail and a confidence level, and list the duties for your role with the article behind each and whether it applies now or in how many days. A model reads the description and flags any answer your own words contradict, quoting them. Each quote is checked against the description, and nothing changes until you apply it.

What It Counts · live app Read the SQL behind a metric before you quote it. Twelve pattern checks run in the browser as you type, with no model, and flag traps such as double counting after a join, a LEFT JOIN undone by a filter, an OR that skips the other filters or a share that rounds down to 0. A model explains the query in plain English line by line, says what one row stands for and whether the query answers your question, and quotes the exact line for every point. Each quoted line is checked against the query. Printed rules set the severity and a verdict that gets stricter when the number is for a decision or a board deck, then write the message to the analyst and a caveat line for the slide.

The builds

Every build has its own folder with a README. Live opens the deployed app, where the worked examples run without a key.

PM practice and career

Including tools for the candidate's side of hiring, drawn from my time in HR tech.

Build	What it does	Try it
Frontline Pulse	Groups field notes into recurring blockers, counts the accounts and money behind each, routes them to an owner with a date and writes what the field should say now	Live
So What?	Turns findings into ranked actions that fit your capacity, each with an owner, a first step this week, a metric, a stop condition and a rough impact range	Live
Who Says Yes?	Maps who has to agree to a proposal, gives each person a play, sets the order to meet them and says whether you are ready to ask	Live
Worth Building?	Turns an AI feature idea into a business case with payback, scenarios, the assumptions to check first and a pitch per approver	Live
How Did That Go?	Debriefs an interview round: what each question was testing, where your answer fell short and what to fix before the next one	Live
Loop Ready	JD and hiring stages in, prep plan per round out, with verified resources only	Live
AI or Me?	Helps a PM decide whether a task needs AI at all, or is better done by hand to keep the skill	Live
Exit Criteria Builder	Gives every stage of a roadmap a measurable finish line the team can check	Live
Ship Check	Describe how something was vibe coded and get a verdict, ranked risks and questions for engineering	Live

PM workflow diagnosis

Tools that help a PM work out what is wrong before pulling in engineering.

Build	What it does	Try it
What It Counts	Explains a SQL query in plain English line by line, marks the traps that could skew its number and writes the questions for the analyst	Live
Signal Translator	Turns a stack trace, failed API response or webhook payload into a plain verdict, blast radius and draft ticket	Live
Repro Builder	Turns a vague user complaint into a reproduction script that isolates one variable at a time	Live
Handback	Audits a draft PRD and lists the questions engineering will send back, ranked by what they block. No model, rules only	Live
Data or Product?	Says whether a metric anomaly looks like broken measurement or a real change in behaviour	Live
Who Owns This?	Ranks the systems most likely to own a bug and drafts the handoff message	Live
Pattern or One-Off?	Judges whether a bug report is a pattern or a one-off, and what would change that verdict	Live
Works on Staging	Finds why something works in one environment and breaks in another, with checks to run before involving engineering	Live

AI reliability and evaluation

Build	What it does	Try it
Wrong Turn	Reads an agent trace step by step, checks each quote against it and names the first step that went wrong, the layer to fix and how to confirm it	Live
Pass the Call	Plans when a voice agent hands a call to a person: a lane per call moment, live triggers, out-of-hours rules and the context packet	Live
Who Does What?	Splits a workflow into steps and puts each in a lane, from plain rules to a person deciding, with approval points, tool access and a rollout plan	Live
Found or Fumbled?	Takes one wrong answer from a document-backed assistant and names which of eight stages broke, with checks, the fix owner and a draft ticket	Live
Judge Calibration Lab	Compares an AI judge with human scores on a golden set, shows which way it leans and finds the rubric wording behind each miss	Live
Prompt Drift Watch	Runs the same messages under two versions of a system prompt and flags behaviour changes you didn't ask for	Live
Eval Starter Kit	Turns a description of an AI feature into a starter eval set with pass and fail conditions	Live
Root Cause Detector	Points to the stage of an AI agent's workflow most likely to have caused a failure	Live
Silent Failure Detector	Flags AI answers that sound confident about claims that cannot be checked, with a redline of every claim	Live
Escalation Quality Scorer	Scores AI-to-human handoffs on context passed and questions repeated, and rewrites the handoff note	Live
Cost-of-Error Estimator	Estimates the monthly cost of each AI failure type, ranks them and tests what a fix would save	Live
Prompt Injection Test Harness	Runs 14 injection attempts against a prompt and checks each reply for the secret, with a positive control	Live
A/B Eval Dashboard	Compares two system prompts on one test set with a judge model, and checks the verdict survives swapping the order	Live

AI governance and compliance

Build	What it does	Try it
Which Tier?	Places an AI feature in an EU AI Act tier with printed rules, lists the duties for your role with dates as amended in 2026, and checks your answers against your own description	Live

Customer success diagnostics

Drawn from my time running customer success alongside product.

Build	What it does	Try it
Pulse Check	Account health score from 0 to 100 across five weighted categories, with top risks and one next action. Fixed formula, no model	Live
Expansion Radar	Expansion readiness score and the play to run next. Fixed formula, no model	Live
Churn Autopsy	A post-mortem on a lost account's timeline	Live

Teen safety with AI chatbots

Tools aimed at reducing harm when AI chatbots give teenagers dangerous advice. No sign-in, nothing stored, and worked examples written without any real method or dosage detail.

Build	What it does	Try it
Duty of Care	Reads a pasted chat and scores three softly named signals: physical safety, emotional signals, trust and boundaries	Live
How to Bring It Up	For parents and guardians: an opener, what to avoid, what to listen for and when to get professional help	Live
What Happens Next	Once a concern is confirmed: who to contact first, what to say and named support lines by country. Rules only, no model, for reliability	Live

Everyday life

Several of these came from living in Spain as an expat.

Build	What it does	Try it
Dígame	Writes a phone script for calling a Spanish office, in the order the call happens, with Spanish read aloud	Live
As Printed	Translates Spanish medical documents literally, expands abbreviations and lists questions for the doctor. It does not interpret	Live
Is This Normal?	Describe something you've noticed about your child and see what official guidance says is typical for that age, and when to ask a professional	Live
Says Who?	Paste a parenting tip for ages 0 to 3 and see how well official health bodies back it	Live
Before the Basket	Scores a grocery item on how it was produced and processed, from its ingredients or a label photo	Live
Lab Report Translator	Reads a soil or leaf analysis and returns per-nutrient status, timed actions and the limits of the test	Live

How the builds are made

The same patterns run through most of them.

Free model by default. Recent builds run on Meta Muse Glimmer through NVIDIA's free endpoint, on my key. Visitors can switch to Anthropic, OpenAI or Gemini with their own key, which is used for one request and never stored.
Prompts stay on the server. The system prompt and API key sit in a serverless function, not in the browser bundle. The newest builds keep that function host-neutral, so the same folder deploys to Vercel or Cloudflare Pages.
Tagged lines instead of JSON. Models reply in lines like STAGE|2|Hiring manager|.... A broken line is skipped, so one bad line doesn't sink the result, and changing provider is cheap.
Three worked examples per build. Saved results load instantly, so anyone can see the tool work without a key.
Rules where a model is the wrong tool. Handback, Pulse Check, Expansion Radar and What Happens Next use fixed, printed rules because the answer has to be the same every time. Is This Normal? keeps its safety rules out of the model for the same reason, Judge Calibration Lab works out its agreement figures and verdict without one, and Worth Building? never lets the model produce a number and checks the written case against the calculated sheet. Who Does What? lets the model describe each step but sets every lane with printed rules, Who Says Yes? does the same for the people behind a decision, Wrong Turn lets the model describe each step of an agent run but picks the wrong turn with rules and checks every quote it cites against the trace, What It Counts finds known SQL traps with pattern checks before any model call and lets the model only explain and point at lines, and Which Tier? sets the AI Act tier, duties and dates with printed rules while the model only checks your answers against your description.
Sources shown. Each app says where its data or judgement comes from and what it cannot do.

Running a build locally

Each folder is its own Vite project.

cd <build-folder>
npm install
npx vercel dev   # runs the app and its API function together

Builds with a model call need these environment variables, set in Vercel or a local .env.local:

Variable	Value
`LLM_BASE_URL`	`https://integrate.api.nvidia.com/v1`
`LLM_MODEL`	The Muse Glimmer model id on NVIDIA
`LLM_API_KEY`	Your NVIDIA API key

To deploy one build on Vercel, import this repo and set the Root Directory to that build's folder.

Builds with a functions/ folder (Silent Failure Detector, Escalation Quality Scorer, Cost-of-Error Estimator, Prompt Injection Test Harness, A/B Eval Dashboard) also deploy to Cloudflare Pages for free: npm run deploy:cloudflare from the build folder, after setting LLM_API_KEY and LLM_MODEL with npx wrangler pages secret put.

Contact

Find me on LinkedIn as Prerna Kakkar, or on GitHub at kakkarprerna.