Fable 5.1 prompting, inside Claude Code
The published guide is the authority; this skill only says which half of it you can act on from a session. Where this file and the live guide disagree, the guide wins and this file is wrong.
Look it up live
WebFetchthe guide at the URL in Sources. Ask it for the one section you need.- There is no pinned copy on purpose: the guide moves, and a pinned copy with no scheduled refresh ages while still claiming to be current.
Symptom index
Owner H means the harness or the effort setting owns it; prompt text costs input tokens and changes nothing. Owner P means the person writing the prompt owns it. Owner Mixed means both: set the effort level and quote the guide's instruction.
| Symptom | Guide section | Owner |
|---|---|---|
| Unsure which effort level to run | Consider all effort levels | H |
| Goes quiet for minutes between tool calls | Ask for user-facing progress updates | H |
| One tool call per turn | Batch independent tool calls in agent loops | H |
bound to a different conversation, cache misses |
Keep the conversation history append-only | H |
| Prose runs long and dense | Writing density | P |
| Replies carry less structure than the content needs | Formatting in chat | P |
| Summaries reuse source wording unmarked | Quoting retrieved sources | P |
| Ends the turn describing what it would do next | Finish the whole task | H |
| Compaction drops constraints or decisions | Tell the model what to preserve in compaction summaries | H |
| Unrequested fixes, or too many committed test files | Keep changes and tests to what the task asks for | P |
| Answers from memory instead of searching | Search triggering at low effort | Mixed |
| Benign code request refused | Reduce safeguard false positives | P |
| Whole file rewritten for a small change | Prefer targeted edits over whole-file rewrites | P |
| Long deliverable is slow or truncated | Leave room for long outputs at xhigh and max effort | H |
| Lead agent idles while subagents run | Let the lead agent keep working while subagents run | H |
| Chart and image detail missed | Give vision work tools to crop and zoom | H |
For a P row, quote the guide's instruction into the prompt. For an H row, change the effort level or the harness, and add nothing to the prompt.
Plan rubric
A plan for a fresh Fable 5.1 session carries five headings, in order. A handoff
note uses the seven headings in docs/WORKERS.md instead; the catchup staleness
check reads its Written at hash.
- Goal and why. What the work is for and what it unblocks. The guide notes Fable 5.1 runs very long tasks well "especially when the goal is clear".
- Constraints. Files in scope, what must not be touched. Carry the scope-of-extras rule here: report a pre-existing bug as a follow-up rather than fixing it, and commit tests only where the task asks for them.
- Done check per task. An artifact and a way to check it, per task. Mark a missing one visibly rather than inventing a check.
- Session conduct. Attended or unattended, and whether the deliverable is prose. If it is prose, add "Please remove all mannered prose."
- Target model and effort. Name the model and the effort level. Prefer
highfor a single long document or code file; atxhighandmaxthe model can draft the deliverable once in thinking and again in the reply.
Never paste these into a Claude Code prompt
The autonomy block beginning "You are operating autonomously", the
# Delivering work block, the progress-updates line, and the tool-batching
nudge. Claude Code already injects all four as system or turn-scoped messages.
A second copy is paid input tokens for no behavior change.
