Record the user's correction while it is fresh. This skill is a receipt printer, not an analyzer: save the feedback, emit a stable marker, and stop.
Storage
Default root:
~/.agents/capture-feedback/
The shared ~/.agents location lets Claude and Codex use the same inbox. To
override it, set CAPTURE_FEEDBACK_HOME.
Each note is one JSON file:
~/.agents/capture-feedback/inbox/<id>.json
Keep notes local. They may contain private paths or user-provided sensitive text.
Minimal Schema
{
"schema_version": "capture_feedback_note/v1",
"id": "cf_20260602T183012Z_7f3a",
"created_at": "2026-06-02T18:30:12Z",
"marker": "capture-feedback:cf_20260602T183012Z_7f3a",
"cwd": "/Users/andypai/Projects/example",
"user_words": "You should have run browser QA too after changing frontend behavior."
}
Do not add classification, severity, proposed fix targets, thread ids, git metadata, model names, or rollout hints during capture. Later review can derive those from traces, git state, or repeated notes. The marker is the bridge back to the transcript.
Capture Flow
- Use the user's correction verbatim. Do not rewrite it.
- If no correction text was provided, ask one short question: "What should I capture?"
- From this skill directory, run the bundled script:
python3 scripts/capture_feedback.py capture -- "<verbatim feedback>"
During repo development, the script path is
capture-feedback/skills/capture-feedback/scripts/capture_feedback.py.
- Reply with only the marker and, when useful, the local file path.
Capture is complete when the script exits successfully, the marker is returned,
the JSON file exists under inbox/, and user_words matches the user's
correction verbatim.
Example reply:
Captured capture-feedback:cf_20260602T183012Z_7f3a
Read Existing Notes
When the user asks to list or show captured feedback, use the same script for lightweight retrieval:
python3 scripts/capture_feedback.py list
python3 scripts/capture_feedback.py show cf_20260602T183012Z_7f3a
list shows recent notes. show prints the raw JSON.
Out Of Scope
- Do not cluster notes.
- Do not inspect traces.
- Do not decide whether to update a skill, memory, or
AGENTS.md. - Do not create evals.
- Do not mutate global rules from a single note.
Those are separate distillation or review workflows.