| name | salvage |
| description | Extract and repair valid JSON buried in chatty model output — strip the preamble, the ```json fence, the trailing comma, the // comments, the Python True/False/None, and the smart quotes, then emit clean parseable JSON. Use when a tool, downstream parser, or pipeline needs strict JSON but the source text is prose-wrapped or slightly malformed, or when you are post-processing another model's structured output. Distinct from simply prompting "respond in JSON": salvage assumes the chatter already happened and rips the JSON out of it after the fact. |
salvage
rip the JSON out of the chatter.
When this skill is active, treat any blob of model output as a haystack and find
the one valid JSON value inside it.
Rules
- Find the first JSON value. Ignore prose before and after. Strip any
markdown code fence (
```json, ~~~). Take the first { or [
and read through its balanced matching close — counting depth, never
counting braces or brackets that live inside string literals.
- Repair the usual damage before parsing: drop trailing commas (
,},
,]); strip // line comments and /* */ block comments; rewrite the
Python literals True / False / None to true / false / null
(as tokens, not inside strings); normalize smart quotes “ ” ‘ ’ to " '.
- Emit only the JSON. No fence, no commentary, no "here is your JSON".
Pretty-print by default; compact on request.
- If nothing parses, say so. Don't invent a structure. Report that there
is no salvageable JSON rather than guessing.
Example
Before: Certainly! Here's the data:
{
"name": "Ada",
"active": True,
}
Let me know if you need anything else!
After: {"name": "Ada", "active": true}
Output
Emit only the JSON — that's the whole point. Don't narrate running salvage.py,
don't fence it, don't wrap it in "here's the recovered JSON." If nothing parses,
say so in one line. Nothing before the JSON, nothing after.
Companion tool
salvage.py in this folder does exactly this mechanically — pipe any chatty
output through it (salvage.py, --compact, --extract-only) to recover the
JSON without re-prompting the model.
For sources with their own quirks, extend the built-in repair tables with a
JSON file via --patterns FILE (repeatable) or the SALVAGE_PATTERNS env var.
The file may carry smart_quotes (char -> replacement, e.g. {"«": "\""}) and
py_literals (bare token -> JSON value, e.g. {"Nil": "null"}); entries merge
into the built-ins and override on collision.
