Practice
The Prompt Workshop
With an agent, your sentences are the source code. A loose sentence gets a loose program. This workshop shows how a loose idea becomes a task an agent can finish, using real prompts from the night this page was built.
From a voice note to a task
FactualA long spoken idea, turned into research tasks.
On 13 November 2026 the site's maker spoke a long idea into a phone. Claude, leading, wrote it down as research requests for Grok. Here is one. Pick a colored phrase to see where it went.
What was said, into a phone
“…deploy Grok to do some research on and then different for all of these things for example …”
What the lead agent wrote down
and find
The worked example. Said loosely, kept, and made exact.
Half of what the worker needed was never said. The lead filled in the count, the shape of the answer and the stopping rule. That is most of the job: deciding what “done” means before anyone starts.
Every line is there for a reason
FactualThe prompt each Grok worker got, line by line, with why.
The prompt the lead then gave each Grok worker. Open a line to see what went wrong before it was added. All of it happened on this project, and the dates come from its journal.
Read AGENTS.md, collaboration.md, the newest handoff, and request R-018.
A fresh agent knows nothing about this project. It only knows what it is told to read. Every session
Do R-018 only, nothing else.
Five workers ran the same night, one task each. A worker that wanders onto another task duplicates or undoes someone else’s work. 13 Nov 2026
Every quote must be copied exactly from the live page you fetched.
Codex checked 79 quotes from one research packet against their pages. Four did not match. 11 Nov 2026
Commit and push after every commit.
Agents had left finished work on one machine, where no other agent could pick it up. 11 Nov 2026
Do not edit content/ or src/.
Only the lead merges, and only checked quotes reach the site. Workers hand findings over; they do not publish them. 8 Nov 2026
No social posts.
An early research run drafted three posts for public accounts. None were meant to go out. 10 Nov 2026
One thing no line could fix: the first worker could not push, because its fresh copy of the project had no installed packages for the pre-push tests. The prompt was fine. The room was not ready. Set up the workspace before you hand it over.
Seven habits of clear instructions
FactualSay what it cannot know, give the why, show the done state, and four more.
- Tell it what it cannot know. The model does not know your norms or workflow unless you say them (Anthropic).
- Say why. A reason helps the model aim its answer (Anthropic).
- Describe the output. Being specific about what you want back improves results (Anthropic).
- Agree on a plan before code. Kelvin Wang writes a spec and a plan and reviews them before anything runs (Kelvin Wang). Amelia Wattenberger approves the spec before any agent starts (Amelia Wattenberger).
- “Done” is not proof. Kelvin says a tool reporting done is not proof (Kelvin Wang). Boris Cherny ranks giving the agent a way to verify its work first (Boris Cherny on X).
- Rules with no exceptions go in code. Kelvin puts them in a hook or a permission rule (Kelvin Wang). See Plate II.
- Say it once, calmly. Piling on capitals and emphasis makes the model over-apply a rule (Anthropic), and the style of your prompt rubs off on the reply (Anthropic). Give the reason instead. More habits like this in the character you are working with.
Try your own
SimulatedPaste a task and see which parts of a good prompt it is missing.
A checklist, not a model. It looks for words that usually mark each part, so it can be fooled. Write a task you actually want done, and see what is missing.
0 of 6 parts a task prompt usually needs
- Says what to read first: missing
- One task, named: missing
- Says what done looks like: missing
- Sets limits: missing
- Asks it to check its work: missing
- Gives a reason: missing