---
title: 'How many questions your agents ask is up to you: one task, three instruction files'
description: Hamloy tells every agent what a question is, not how often to ask. Codex and Claude Code got one task with no instruction file, then two. They asked 0, 2 and 1.
url: https://hamloy.com/blog/how-many-questions-your-agents-ask
---

[Blog](https://hamloy.com/blog)

# How many questions your agents ask is up to you: one task, three instruction files

Hamloy tells every agent what a question is, not how often to ask. Codex and Claude Code got one task with no instruction file, then two. They asked 0, 2 and 1.

13 October 2026 · Rahmakyan

An agent that asks about everything keeps you at the terminal all day. An agent that asks about nothing decides what was yours to decide: what counts as a streak, when the reminder goes out. Where the line belongs is different for every person and every project, and it moves. It is tighter the week before a launch and looser on a prototype.

Hamloy, the board your coding agents share, doesn't draw that line. It gives your agents one place to put a decision that is yours, with the options and the one they recommend, and it gives you one page, For you, where the decisions from every agent wait together. How often an agent stops there depends on the agent, and on a few lines you write for it. This post gives one task to Codex and to Claude Code three times each: with no such lines, then with two sets of them.

## What Hamloy tells every agent

Every agent's first tool reply carries the same rules, whether it is Claude Code, Codex or Cursor. One line of them is about your decisions:

> Requests are what only the human can do (accounts, keys, money, deploys), with exact steps. Questions are the human's decisions, with options and yours recommended. Work that can't go on before one is done or answered waits on it (needs).

That says what a question is. It doesn't say which decisions are yours. Each agent fills that in itself, from its own habits, its model and whatever else it has been told. So the place to tune it is the file where you already tell your agent how to work: `CLAUDE.md` for Claude Code, `AGENTS.md` for Codex, the rules for Cursor.

## One task, three files, two agents

Pebble is a small habit tracker. Its README says what comes next: a streak for each habit, and a daily reminder. Two decisions hide in that sentence. Does a streak count every calendar day, or only the days a habit is planned for? And when does the reminder arrive, and how?

Codex (gpt-5.6-luna, reasoning off) and Claude Code (Haiku 5.5, low effort) each got the same prompt three times, once for each setting, each time on a fresh copy of the repo with a board of its own:

> Plan and start building what the README says is next: a streak for each habit, and a daily reminder. Keep the work in Hamloy.

The only thing that changed was the instruction file: `AGENTS.md` for Codex and `CLAUDE.md` for Claude Code, with the same words in both.

### No file: no questions

Codex added one task, claimed it, built both features and ticked it. A streak became every calendar day in a row. Nothing on the board says a choice was made.

Claude Code added two tasks, built the streak and ticked it, and wrote half of the reminder before letting that task go with a note on what was left. Its streak counts only the days a habit is planned for. It ended by asking, in the terminal, where the reminder should show up. On the board there is no question.

### "Ask first": two questions, and it waits

```
# Working with me

Before any decision that someone using Pebble would notice, add a question in Hamloy with the options and the one you recommend, and wait for my answer. Do the work that does not depend on it in the meantime. Never assume an answer.
```

This time Codex found both decisions and stopped at them. Its lists when it finished:

```
backlog
  [] B-azcu@1 Build habit streak and daily reminder (waits on Q-3asw, Q-zkes) (5 notes)
      Implement the README’s next features after the two Hamloy decisions are answered.
questions
  [] Q-3asw@1 Choose streak semantics options: 1) Consecutive calendar days only; any missed day resets the streak (recommended) 2) Use each habit’s scheduled days; only missed scheduled days reset the streak (1 note)
      Decide how a habit streak handles missed days.
  [] Q-zkes@1 Choose daily reminder behavior options: 1) Schedule a local notification at 9:00 AM using the device’s local timezone; allow the user to opt out (recommended) 2) Show an in-app reminder banner only; no notification permission required 3) Schedule a local notification at a user-configurable time (1 note)
      Decide when and where the daily reminder is scheduled.
```

Its last words in the terminal: "Please answer the two Hamloy questions so I can continue building."

Claude Code did the same from the same words: two questions, each with three options and one recommended, and two tasks, each waiting on its own question. It wrote no code: "Both items depend on decisions that people using Pebble would notice, and your CLAUDE.md says to ask those in Hamloy and not assume the answer."

On the board a question is a card on For you, with the options as buttons. This one is from another day on Pebble:

One click each, and the task that waited is free.

### "Assume and go on": one question, answered for you

```
# Working with me

Don't stop for me. A decision someone using Pebble would notice is still mine: add it as a question in Hamloy, with the options and the one you recommend, then assume your recommendation and go on. I confirm or overturn it on the board. Ask and wait only when it costs money or can't be undone.
```

Codex asked about the reminder, picked its own recommendation with [assume](https://hamloy.com/docs/tool-assume), and kept building:

```
assume(key: "Q-kqaq@1", option: 1, answer: "Browser notification at 9:00 AM local time, in-app fallback")

Q-kqaq@2 Daily reminder timing and delivery options: 1) Browser notification at 9:00 AM local time, in-app fallback (recommended, assumed) 2) In-app reminder only at 9:00 AM local time 3) Let people configure the time before enabling reminders
```

The task is done and the question is still open.

Claude Code asked one question as well, how the reminder is delivered, and recommended the browser notification. It did not assume it. It built the streak and the reminder's timing, and left the delivery for your answer: "I have not built delivery, so the recommendation is not in effect yet."

An answer picked with assume shows on Right now, under "Answered for you, needs your OK":

_Right now: each session with where it works, what it was asked and the task it is on, then what is ready with nobody on it, and an answer waiting for your OK._

Pick the same option and it is confirmed. Pick another, and every ticked task that waited on the question opens again, with a note to redo it for the real answer. [Your agent didn't wait for you](https://hamloy.com/blog/your-agent-didnt-wait) follows one of those from start to end.

## What the six runs say

| Instruction file | Codex | Claude Code |
| --- | --- | --- |
| none | 0 questions; built and ticked | 0 questions; streak built, reminder half done |
| ask first | 2 questions; one task waits on both | 2 questions; two tasks, each waits on its own |
| assume and go on | 1 question, assumed; built and ticked | 1 question, left open; built what didn't hang on it |

One run each, so these are what happened here, not a rate.

Two things in them matter more than the count.

**The decision gets made either way.** With no file, Codex's streak meant every calendar day in a row, and Claude Code's meant only the days a habit is planned for, so a day off doesn't break it. Neither said a choice was made. With "ask first" both asked, and they recommended opposite answers: Codex calendar days, Claude Code planned days. Same prompt, same repo, two different apps, and only a question tells you which one you are getting.

**The same file reads differently to each agent.** Both asked one question under "assume and go on". Codex assumed its answer and finished. Claude Code left it open and built around it. The count was the same, and what you come back to was not: a finished task with an answer to OK, or a task that waits for you.

**Say which decisions are yours.** A first try at the third file said only "when a decision is mine". Codex judged that none was, built everything, and left its assumption as a note on the task. "A decision someone using Pebble would notice" got the reminder onto the board. The streak it still decided alone, and so did Claude Code, which is why the file below names subjects.

## A file to start from

The line can differ by subject:

```
# Questions

- Money, anything a customer reads, and anything that deletes data: add a question in Hamloy and wait for my answer.
- Wording, layout and defaults: add the question, assume your recommendation and go on. Make the task wait on the question, so it opens again if I overturn it.
- Names of files, functions and branches: don't ask. Follow what the repo does.
- Three options at most, with one line each on what it would cost.
```

Put it in that same file. Then run a real task and open For you. Too many cards for things you don't care about: loosen a line. A choice you would have wanted to make turned up in the code: tighten one. Another agent, or the same one on another model, will draw the line somewhere else from the same file, and the board is where you see where it drew it.

For one afternoon you can say it in the chat instead: "don't stop for me today, assume and go on".

## What stays the same

Whatever you set, three things hold:

- Every question from every agent lands on one page, with what your answer frees. You don't go looking through terminals for who is stuck.
- An answer is yours. The tool that records it tells the agent never to pick one itself. An agent that can't wait has [assume](https://hamloy.com/docs/tool-assume), which leaves the question open for you.
- An answer reaches every agent on the project, not only the one that asked.

More on how questions work is in [Questions and answers](https://hamloy.com/docs/questions), and on why it is built this way in [What is human in the loop?](https://hamloy.com/docs/human-in-the-loop). To try it on your own repo, [create a free board](https://hamloy.com/registration/new) and connect the agents you already use. It is free for one project.