CONCEPT The classic two-part trick
Before structured outputs existed, the standard way to coax clean JSON out of Claude with no feature at all combined two Messages-API mechanics:
- Prefill the opening brace. End your request with an
assistantturn containing just{. Claude continues from that character, so it never emits a "Here's the JSON:" preamble — the very first token is already inside the object. - Stop on the closing brace. Pass
stop_sequences: ["}"]. The moment Claude would emit}, generation halts withstop_reason: "stop_sequence". The stop text itself is not included in the output, so you re-add the braces yourself when parsing.
stop_sequences is a plain Messages-API parameter (a list of strings). When one
fires, stop_reason is "stop_sequence" and the matched string appears in
response.stop_sequence — and is stripped from the generated text.CONCEPT Prefill is now model-gated — know the caveat
assistant message. On current models, use tool use (7.3) or the
structured-outputs feature (7.4) instead.The still-portable no-feature approach on any model is the second flavor: ask for JSON only
and parse. Instruct "reply with only valid JSON, no prose", optionally add a stop sequence to trim any
trailing chatter, then json.loads() the text with a try/except and a retry on failure.
DIAGRAM How the two mechanics combine
assistant prefill?"} PRE -->|"older model: yes"| A["🟢 Prefill assistant turn with '{'
+ stop_sequences=['}']"] A --> OUT["Claude emits object body,
halts at '}' (stripped)"] OUT --> WRAP["Re-add braces → json.loads()"] PRE -->|"current model: no (400)"| B["🟠 Ask for JSON-only + parse
(or escalate to 7.3 / 7.4)"] B --> WRAP classDef ok fill:#d4edda,stroke:#28a745,color:#155724; classDef gate fill:#ffe8cc,stroke:#ff9f40,color:#9a4a0a; class A ok; class B gate;
CODE The classic trick vs. the portable fallback
# ── Classic prefill + stop (OLDER models only — 400 on 4.6/4.7/4.8/Fable 5) ── resp = client.messages.create( model="claude-3-5-sonnet-20241022", # a model that still allows prefill max_tokens=1024, stop_sequences=["}"], # halt at the closing brace messages=[ {"role": "user", "content": "Extract name and email as JSON."}, {"role": "assistant", "content": "{"}, # prefill the opening brace ], ) # stop_reason == "stop_sequence"; braces are stripped, so wrap them back: raw = resp.content[0].text # e.g. "name": "Jane", "email": "j@co.com" data = json.loads("{" + raw + "}") # ── Portable fallback (ANY model): ask for JSON-only, then parse defensively ── resp = client.messages.create( model="claude-opus-4-8", max_tokens=1024, system="Reply with ONLY a valid JSON object. No prose, no code fences.", messages=[{"role": "user", "content": "Extract name and email as JSON."}], ) try: data = json.loads(resp.content[0].text) except json.JSONDecodeError: data = retry_or_repair(resp) # re-ask, or escalate to 7.4
assistant turn is rejected before generation even starts.TRADE-OFFS No-feature JSON vs. the real feature
✅ Prefill + stop / JSON-only
- Zero setup — no schema, no beta, no compile step
- Prefill kills the preamble on models that allow it
- Portable JSON-only variant works everywhere
⚠️ The cost
- Not guaranteed — output can still be malformed
- You own parsing, validation, and retries
- Prefill path is dead on current models
- Nested
}in strings can trip a naive["}"]stop
EXAM TRAPS
response.stop_sequence. That's why you re-add
the closing brace.stop_sequences=["}"] on nested JSON. → It fires on the
first }, which may be inside a nested object — truncating the response. The trick suits flat
objects; for nested shapes, prefer the JSON-only + parse variant.PRACTICE
Q1 · Describe the classic no-feature JSON trick in two moves.
Prefill the assistant turn with { so Claude skips the preamble and
starts inside the object, and set stop_sequences=["}"] so generation halts at the closing brace.
Because the stop string is stripped, you re-wrap the result: json.loads("{" + text + "}").
Q2 · Why can't you use the prefill half on Claude Opus 4.8?
Assistant-turn prefill is not supported on Fable 5 and the 4.6/4.7/4.8 family — a trailing
assistant message returns a 400. Use tool use, structured outputs, or a JSON-only system
instruction instead.
Q3 · What is the portable, any-model version of this technique?
Instruct the model to "reply with only valid JSON, no prose", optionally add a stop sequence
to trim trailing text, then json.loads() defensively with a try/except and a retry or repair path.
It's still best-effort — escalate to structured outputs for a hard guarantee.