Your agent is fast, confident, and happy to build the wrong thing for three hours. The problem was never the code. It's that nobody agreed what “done” meant, and the decisions that were yours got made without you.
These are the skills I use every day to fix that. Built by a product person and designer who doesn't read code, and refined on real work.
Good work starts with a good brief. Agents skip straight to the work.
Ask for something big and your agent starts immediately. No questions, no brief, no idea how you'll judge the result. Every gap it hits, it fills with a guess. You only find out which guesses were wrong once the work is built on top of them.
A good product person does the opposite. They ask first. They come back with a goal you can correct. They tell you which calls are yours. That's all these skills do: they make the agent work that way.
AgentBefore I start: how will you judge whether this worked?
YouMore visitors starting a trial.
AgentGot it. Which decisions do you want brought back to you — copy and layout, or only the pricing structure?
YouJust pricing. You own the rest.
AgentThen here's the goal I'd work to. Correct anything before I start.
One is ready to install. Three are in the workshop.
Each one takes a job a product person does by hand and teaches it to your agent. They work in Claude Code, and they're written in plain English: no code required to use them, or to read them.
-
1.
Available now
Context is Queen
Stops your agent building the wrong thing. It interviews you one question at a time, researches what it can answer itself, then proposes a goal for you to correct, and holds itself to that goal until the work actually meets it.
Read the docs → -
2.
In the workshop
Turning Research Into Strategy
You have call transcripts, a painpoint tracker and a deck claiming eleven customers asked for the thing. This walks back down to the raw evidence, counts it under a rule you set first, and rebuilds the direction from what's actually there.
-
3.
In the workshop
Judgment Queue
When your agents owe you a pile of decisions, this collects them into one page: each call as a plain question, the evidence sitting right next to it, answerable in a word. No more decisions buried in a wall of chat.
-
4.
In the workshop
AI Output Clinic
For when AI-written copy comes back generic, flat, or subtly off-tone. It works out which part of the instruction caused it, then fixes the prompt behind the words, not just the words.
Context is Queen, in about a minute.
Open Claude Code and paste this into the chat. “Globally” makes it available in every project, not just this one.
› Install the Context is Queen skill globally from github.com/hgambas/think-like-a-pm
Prefer the plugin system? Run
/plugin marketplace add hgambas/think-like-a-pm, then install
context-is-queen from the /plugin menu. That version
also sets up the session-start nudge, so fresh sessions offer the briefing
on their own.
- 1.Ask for something complex, in your own words
“Rework our onboarding so people actually finish setup.” No special syntax needed. The briefing starts on its own, and simple asks skip it.
- 2.Answer the questions, one at a time
Three to eight of them, covering only what your agent can't research for itself.
- 3.Correct the goal it proposes
Then it works to that goal, and checks itself against it before claiming anything is done.
“Work aimed at a goal you've corrected is simply better work — and you stay the decision-maker on your own project.”
I'm Hiba Ganta. I'm the product person and the designer; my agent is my engineer. I don't read code, which means I can't catch a wrong turn by reading the diff. I have to catch it in how the work gets decided.
Every skill here started as a rule I kept repeating to my own agent until it was worth writing down properly.