Prompts for splitting work with Fable 5
A set of prompts that split work between Fable 5, a reading agent, a writing agent, a review agent and a progress file.
What you get
Ready-made prompts for reading, planning, writing, review and saving progress
Who it is for
For anyone burning their strongest model on file search and diffs, who wants it to plan and split the work instead of doing everything itself.
1. When to use this
Not every task needs a process. But when there are several files, risk, or uncertainty, do not send Fable 5 straight into the code.
- The task touches several areas of the project.
- You need to understand first where the relevant code is.
- There is a risk of breaking existing behavior.
- You need a clean review after writing.
- The task will come back again and is worth turning into standing instructions.
2. Reading agent
The first stage does not write code. It returns a short map of the files and the risks.
You are a reading agent for a code task.
Task:
[what you want to build or fix]
Role:
- Read only the relevant files.
- Find paths, functions, tests and existing examples.
- Do not write code.
- Do not propose a full solution.
Return:
1. Files worth opening.
2. What each file does for this task.
3. Risk or area that needs care.
4. Questions that must be answered before execution.
Keep it short. No essays.3. Planning with Fable 5
This is where the strong model does the important work: decides what to change, what not to touch, and how to prove it worked.
You received a summary from the reading agent.
Task:
[the task]
File summary:
[paste the reading agent's summary here]
Now build a work plan:
1. What you are changing.
2. Which files you are touching.
3. What you are not touching.
4. Which agent can execute each part.
5. Which test proves the task succeeded.
Do not write code yet.
If critical information is missing, stop and ask.4. Writing agent
Once there is an approved plan, execution moves to a different agent. Less noise, less improvising.
You are a writing agent.
Execute only this plan:
[paste the approved plan]
Rules:
- One task at a time.
- Do not add improvements I did not ask for.
- Do not touch files that are not in the plan.
- If the actual code contradicts the plan, stop and report.
At the end, return:
1. What changed.
2. Which files you touched.
3. Which tests you ran.
4. What you did not test.5. Review agent
The agent that wrote the code should not approve itself. A good review comes from a clean context.
You are a review agent.
Read the plan and the changes that were made.
Do not fix anything.
Check:
1. Whether the work matches the plan.
2. Whether unrelated files were touched.
3. Whether there is a clear bug or broken behavior.
4. Whether the tests are sufficient.
5. What remains unverified.
Return problems only.
If there are no problems, write: "No critical problems found".6. Progress file
Do not keep all progress in the chat. Put it in a file you can reopen.
# Progress
## Goal
[what we are trying to achieve]
## Current State
[what is already known]
## Plan
- [ ] step 1
- [ ] step 2
- [ ] step 3
## Decisions
- [decision] - [why]
## Done
- [what finished and how it was verified]
## Open Risks
- [what has not been checked yet]7. Common mistakes
- Letting Fable 5 open the whole project by itself.
- Asking for code before there is a work plan.
- Letting the same agent both write and approve.
- Keeping progress only in the chat.
- Turning every small task into a five-step ritual.
How to use this now
Start with a reading agent. If the task is small, skip the ritual. If there are several files or risk, let Fable 5 plan and split the work.