Prompt library
Prompt library
Prompts you can paste into a coding agent. Each one asks for a small, checkable change instead of a whole product.
Answer
Copy a prompt, fill in the blanks, and send it. These are instructions for the agent, not claims about what any model will do. If the result is wrong, the next prompt should point at the file and the error.
Start a project
Use these once, at the beginning, so the agent picks a stack and stops adding tools you did not ask for.
Build a one-page app in this empty folder.
What it does: [one sentence].
Who uses it: [one person, on their phone and laptop].
Stack: Next.js, no database, no login.
Done means: I can run it locally and see the page with real copy, not placeholder lorem.
Do not add a backend, analytics, or extra pages.
Before you write code, list the files you will create.I want to build [one sentence]. I am a beginner.
Propose one stack: framework, where it is hosted, and whether it needs a database.
Give the reason in two sentences, and name the official docs you would follow.
Do not write code yet. Wait for me to say yes.Change one thing
Keep the agent inside one file or one behavior. Ask it to show the diff.
Change only [path/to/file].
The change: [what should be different].
Leave every other file alone.
When you finish, show the diff and say how I can see the change in the browser.Add [feature] using the same patterns as [path/to/similar-file].
Do not introduce a new library.
If you believe a new library is required, stop and tell me the name and why, and wait.Explain an error
Paste the real message. Ask for the cause before a rewrite.
This command failed:
[paste the command]
This is the full output:
[paste the error]
Explain the first error in plain English, name the file and line if the output has them, and propose the smallest fix.
Do not refactor unrelated files.This used to work: [what I could do].
Now it does this: [what I see].
The last change was: [commit message or a one-line description].
Compare that change to the failure. Tell me whether to revert or to patch, and why.Git
Ask the agent to use Git, and to stop before a force push or a commit of secrets.
Look at git status and git diff.
Stage only the files that belong to this change: [name the files].
Write a one-line commit message that says what changed.
Do not push.
Do not commit .env or any file that contains a key.Create a branch named [short-name] from the current main branch.
Do the work for [one sentence] on that branch.
Do not merge and do not push until I ask.Database
Make the agent say what is stored, and where the secret lives, before it adds a client.
Add storage for [what you are saving].
Use the database already in this project. If there is none, stop and say so.
List the columns, who is allowed to read them, and which environment variable holds the connection secret.
Do not put that secret in a file that ships to the browser.Look at the existing schema or migration files.
Tell me the table and column I should use for [job].
If it does not exist, propose one migration and wait for me to approve it before you apply it.Deploy
Ask for the host's steps and the list of environment variables, not a new host.
The app runs locally with [command].
I want it on [Vercel or the host already in the repo].
List the environment variables the host needs, which of them are secrets, and the exact dashboard or CLI step from that host's docs.
Do not switch hosts.
Do not print the secret values.The deploy log says:
[paste the log]
The host is [name].
Explain the failing line and the smallest change that addresses it.
Link the sentence to a file in this repo or to a line in the log.Review before you trust it
Use this after the agent says it is done. The point is a list of risks, not more features.
Review the diff for this change.
Flag any API key, password, or token in the diff.
Flag any import or API field you are not sure exists, and say what I should check in the docs.
Do not change files. Reply with a list.Read the routes and the data writes in this change.
List what a person who is not logged in can read or change.
If you are not sure, say so and name the file you would need to read next.
Do not add features.Constrain the agent
Paste this at the start of a session when the agent keeps widening the task.
For this session:
- Touch only files I name.
- Ask before you add a dependency.
- Ask before you run a command that deletes files or talks to a paid API.
- If a step needs a secret, tell me the variable name and where to set it. Do not invent a value.
- When you finish, list the files you changed and how I can check the result.
The task is: [one sentence].Stop. Do not edit anything else.
Summarize what you already changed, what is still broken, and the single next step you recommend.
Wait for me to choose that step.