Fable 5 is an operating partner, not a chatbot. Prompts written for older Claude models are often too prescriptive for it — Anthropic says they actively degrade output. The move: set goals, not steps. Define the outcome, give boundaries, tell it when to act and when to pause, then get out of the way. Because Fable keeps working until told to stop, every long run needs a hard stop: a budget cap, a time cap, a done condition, or all three.
1. Act when there's enough information — stops it re-deriving context on ambiguous tasks.
When you have enough information to act, act. Do not re-derive facts already
established, re-litigate decisions I've made, or narrate options you won't pursue. If weighing a choice, give a recommendation, not a survey.
2. No unrequested expansion — essential for narrow fixes.
Don't add features, refactor, or introduce abstractions beyond what the task
requires. A bug fix doesn't need surrounding cleanup. Don't design for hypothetical future requirements: do the simplest thing that works well. Only validate at system boundaries.
3. Lead with the outcome — for any reader who didn't watch the work happen.
Lead with the outcome. Your first sentence after finishing should answer "what
happened" or "what did you find." Supporting detail comes after. Pause for me only when the work genuinely requires it: a destructive action, a real scope change, or input only I can provide.
4. Evidence-backed progress — nearly eliminated fabricated status reports in Anthropic's testing.
Before reporting progress, audit each claim against a tool result from this session. Only report work you can point to evidence for; if something isn't verified, say so. If tests fail, say so with the output.
5. State the boundaries — the antidote to an agent that's too helpful.
When I'm describing a problem or thinking out loud rather than requesting a change, the deliverable is your assessment. Report findings and stop. Don't apply a fix until I ask.
6. Delegate to subagents
Delegate independent subtasks to subagents and keep working while they run.
Intervene if a subagent goes off track or is missing context.
7. Markdown memory — the difference between one-off prompting and compounding intelligence.
Store one lesson per file with a one-line summary at the top. Record corrections
and confirmed approaches, including why they mattered. Update existing notes rather than duplicating; delete notes that turn out wrong.
Suggested layout: /memory/product-decisions/, /memory/prompt-patterns/, /memory/engineering-conventions/.
8. Make autonomous runs actually finish