Reference · dump frozen on 02/09/2026
What Claude reads before from you
Analysis, index, and practical guides to Claude Fable 5.1's system prompt: what's official, what's extracted, what changed, and how to use it to your advantage.
Analysis, index, and practical guides to Claude Fable 5.1's system prompt: what's official, what's extracted, what changed, and how to use it to your advantage.
Everything that supports the course "Fable 5.1 in Practice," organized for quick reference: sources and their weights, an analysis of what changed, and ready-to-copy reference cards. The framing throughout: official × extracted e model × surrounding product.
The central prompt Anthropic publishes and signs, and the Claude.ai runtime dump (memory, search, artifacts, 46 tools), which Anthropic has not authenticated. The README identifies which is which.
Script-generated map of the file’s 40 sections, with line counts, size, and group. Metadata only: you can see where everything is without redistributing the content.
Form for the buyer's 5 tests, rules for writing for Fable (remove the scaffolding, say when), and a cost and effort table. Plain Markdown, ready to copy.
The source material stays local (it’s from a third party or an unauthenticated extraction). What gets published is the derivative work: index, analysis, and reference cards.
Identity, tone, refusals, well-being, cutoff date. Lines 2–188 of the dump match what Anthropic publishes.
Memory (836 lines, 38% of the file), past conversations, visuals, plugins and skills, search, artifacts. This is where 5.1 grew 2.3x.
Index, analysis, and reference sheets. Dated. The product changes; the method of reauditing each generation remains.
You only need a browser to read the reference cards. To regenerate the index, you need Python 3 and the local dump.
The derived files come with the clone. The folder doc/ doesn't come with it.
# clone git clone git@github.com:inematds/fable51-system-prompt.git cd fable51-system-prompt
Only needed to regenerate the index. The original file is public in the CL4R1T4S repository, in the ANTHROPIC folder.
# save as doc/Claude-Fable-5.1.md mkdir -p doc # expected md5: be4cb3a5103be5c1c581afac2a04b1d0
No external dependencies.
python3 --versionChoose based on what you want to do today. Each step points to a repository file.
Open the analysis: the three numbers (2.3x · +28 tools · 75% memory+schemas), the 28 additions in 6 groups, the new behavior rules, and the model × product paradox.
cat ANALISE.md # or read it on GitHub
The index lists the 40 sections with line counts and size, the tools found by schema, and the 11 skills, plus a suggested reading order (tone → memory → conversations → visuals → search → tools).
cat INDICE-DO-DUMP.md # with the local dump, regenerate: python3 build/indice.py # 40 sections · 31 tools by "name" · 11 skills
Copy the form, fill in the surface, plan, and date, then run the 5 tests: continuity (memory), records (past conversations), interface (visuals), capability (skills/plugins), and control (delete memory). A result without a form is an impression; with a form, it's data.
cat fichas/01-ficha-dos-5-testes.md # Test 2 prompt example (record): Sobre aquela decisão que tomamos sobre [tema]: qual foi a opção que escolhemos e por quê?
Remove the scaffolding ("be concise," "think step by step" are already in the system prompt), tell when use each capability, show the format instead of prohibiting it, and use the Sep/2026 skeleton (result, source of truth, authority, ready, verification).
cat fichas/02-escrever-para-o-fable.md # rule of thumb: every line of the prompt should contain information the model doesn't have
Per token, Fable costs 2x as much as Opus 5 and 5x as much as Sonnet 5. You only save money when you measure by completed task: sweep through the effort before dropping to a smaller model, run it cheaply and redo only what failed (if you have a checker), and compare on the most difficult tenth of your work.
cat fichas/03-custo-e-esforco.md # API tuning order: cache → effort → budget → model, one at a time, measuring
Measured by the Field Guide using two unauthenticated dumps, frozen on 02/09/2026. The benchmarks are those Anthropic chose to publish.
words in the captured prompt: 17,501 → 40,046
tools added (46 total, none removed)
of the growth is memory (+8.639 words) and schemas (+8.328)
of the current dump is just the section memory_filesystem (836 of 2,196 lines)
list / read / write / append / str_replace / delete. Never stores health, religion, orientation, or payment information. Only included in a response if it changes the substance.
conversation_search, recent_chats, read_conversation. Reopens the original conversation using clues like "that decision."
Responses become interfaces; suggestion for a missing capability (1 card per conversation); deep research at your discretion; stop only for safety.
Everything is dated. The next model generation will reopen the audit.