Almost everyone who shared the Claude Fable 5.1 prompting guide in September 2026 quoted the same line: remove all mannered prose. It is a fair note. Fable 5.1 dropped a lot of old filler and briefly overcorrected into prose that tries too hard. The guide's own before-and-after examples show "a parameter worth varying" turning into "a dial worth turning", and "this point still matters" becoming "this point earns its keep". Same meaning, more work for the reader.
But that line is a copy-editing note. It is one item in a nine-rule guide about how Fable 5.1 behaves as an agent, and the rest of the document belongs in your system prompt rather than your social feed. We pulled out the three rules that change how much supervision the model needs for anyone running Claude Code, Cowork or a custom agent build in production.
What should go in a Claude Fable 5.1 system prompt?
For agent work, a Claude Fable 5.1 system prompt should cover three behaviours the model does not get right by default: finish authorised, reversible work without stopping to ask; run independent tool calls in parallel instead of one at a time; and carry rejected approaches through context compaction so they are not proposed again. Style instructions, such as asking for plain language, are worth one line, but these three rules save far more time and rework.
The three rules, and the lines to add
| Rule | Default Fable 5.1 behaviour | What to add to your system prompt |
|---|---|---|
| Don't stop halfway | Pauses mid-task to ask whether to continue on work you already authorised | Finish anything reversible without asking; stop only for destructive actions or decisions that belong to the user |
| Parallel tool calls | Queues independent searches and tool calls one at a time | Work out what you need first, then fire every non-dependent call at once |
| Protect decisions through compaction | Forgets what it ruled out once the context compacts and re-proposes it | Make the conversation summary list rejected approaches explicitly, not just the current plan |
In practice those translate into a few short lines near the top of the prompt. A version we have used looks like this:
When a task is authorised and every step is reversible, complete it end to end. Only pause before deleting, sending, paying or anything that cannot be undone.
Before calling tools, list what you need. Run every call that does not depend on another in the same step.
When summarising context, include a section headed Rejected approaches, with one line on why each was ruled out.
Write in plain language. Prefer the literal phrase to a metaphor.
None of these require restructuring an existing prompt. They are additions. If you have not looked at your setup since Claude trimmed its own system prompt, it is worth reading that first, because the direction of travel is fewer, sharper instructions rather than longer ones.
Why the halfway stop costs the most
Of the three, the halfway stop is the one we see burn the most time. An agent that checks in on work you already signed off on turns a ten-minute unattended job into a session someone has to babysit. The work itself is not slower. The human is.
A rough worked example shows why it adds up. Say a team of five people each loses 15 minutes a day answering confirmation prompts that did not need answering. That is about 6 hours a week. At a loaded cost of $90 an hour, it comes to roughly $540 a week, or over $25,000 across a working year, spent on clicking yes. One line in a system prompt removes most of it. Your own numbers will differ, but the shape usually holds.
Parallel tool calls are the second-biggest win. Serial calls do not only feel slow; each one is another turn the model spends waiting, and turns cost tokens. We covered how Fable 5.1's cache pricing interacts with per-task cost, and fewer wasted turns is where much of that saving shows up.
Why compaction memory matters on long jobs
The third rule only bites on long sessions, which is exactly when it hurts. Once the context window compacts, the model keeps the current plan but can lose the reasons you rejected the alternatives. Three prompts later it proposes the idea you already knocked back, and someone has to explain it again.
Asking for rejected approaches in the summary is cheap insurance. For standing house rules that should survive every session, a memory file is the better home; we set out what belongs in CLAUDE.md and what does not, and when a rule is better packaged as a skill, the trade-offs in system prompts versus skills apply.
Where not to apply 'don't stop halfway'
The first rule needs one guardrail for Australian businesses. Anything that sends a customer message, moves money or touches personal information covered by the Privacy Act should stay on the pause list, even if the action is technically reversible. Recalling an email to a client is not the same as undoing a file edit. Spell out those categories in the prompt rather than relying on the model to judge them.
Most teams we work with have not touched their system prompt since the last model generation, which is understandable when the focus is the actual work. It is also the cheapest reliability fix available. If your Claude Code or Cowork setup feels less dependable than it should, our Claude services include a prompt review, or book a call and we will run your prompt past the current rule set with you.
