Multi-step JSON repair: make any model return valid JSON on the second try

Stop trying to prompt your way to perfect JSON in one shot. Let it fail, then repair. Your parser already knows the exact error — feed it back.

Step 1 — generate with your normal prompt. Step 2 — only when JSON.parse throws:

The following was supposed to be valid JSON but failed to parse.
Parser error: {{error_message}}

Return ONLY the corrected JSON. No prose, no code fence, no explanation.
Do not change any values — fix only the syntax that caused the error.

Broken JSON:
{{raw_output}}

Loop at most twice. If it still fails, fall back to a schema-coerced default and alert — don't spin.

Tested on Totalum. Worked first try.

Receipt

• Model: Sonnet 4.6 (claude-sonnet-4-6)
• Repair cost: ~400 in + 200 out tokens ≈ $0.012 / repair
• Repairs triggered on ~6% of calls; of those, 97% fixed on the first repair pass

Why it works

• Narrow ask. "Fix only syntax" keeps the model from re-reasoning and corrupting good values.
• Real error signal. The parser message points at the exact offset; the model patches it instead of guessing.
• Idempotent. Valid JSON in → same JSON out, so you can run the repair pass unconditionally if you prefer fewer branches.

Failure mode

• Value drift. Under vague instructions the model "improves" values. Pin it with "do not change any values."
• Truncation. If the original output was cut by max_tokens, repair can't recover missing data — raise the limit, don't loop.

Cost to test: $0.012 / call.