Address Feedback
Take a pile of feedback, fix what is objectively broken, verify it, and report
back briefly. Lean toward fixing, but don't agree with every suggestion or
bundle unrelated cleanups. If no link or text was given, ask for it. Use the
matching connector (Notion, Drive, Linear, gh, Slack) rather than scraping
authenticated pages. Slack reactions and replies follow review-latest-feedback.
1. Read it and check for repeats
Read the whole source, including comments, screenshots, and linked recordings. The error text in a screenshot is often the diagnosis.
Then check whether it was reported before; the same defect gets filed weeks apart by different people, and one came back fifteen times. Search three months back: the source, Sentry, Analytics errors, and merged PR titles, using the feature name, error text, and surface separately.
- No earlier report: carry on.
- Earlier report, no fix: say how long it has been open.
- Earlier report, fix landed: a regression. Find why the first fix stopped holding and leave a test of the user's path that fails if it breaks again.
2. Sort each item
Checklist each item: symptom, expected behavior, evidence, owning surface.
- Bug: broken behavior, crash, wrong or lost data, dead link, API mismatch, captured exception, and objective visual defects such as clipped or overlapping text, bad contrast, or a broken layout. Separate the observed failure from the reporter's proposed remedy; fix the failure even if the remedy is out of scope.
- UX proposal: adding, moving, hiding, or restyling controls, or anything
"hard to find". These are product calls. Don't build them on your own; write
the proposal (what changes, why, the tradeoff) and flag it for a person. A
proposal a maintainer has explicitly approved (such as an
:upvote:) is a request to build the smallest version. - Unclear: you've read everything and one concrete detail still blocks a safe fix.
- Out of scope: praise, chatter, other repos, intentionally unsupported. Note it and move on.
For data, permissions, or lifecycle feedback, check the whole capability before calling it UX: does the action exist when a control is missing, and do owner, shared, and non-owner access behave? Use safe fixtures, never another user's production data. Possible cross-user exposure is a security bug.
3. Ask only when you must
Before asking the reporter anything, read the full thread and every accessible attachment, and list what is already known: surface, repro, exact error, run IDs. Never ask for something already there. If someone in the thread already named the cause, fixed it, or is fixing it, verify that and continue instead of asking again. If a run ID is given, inspect that run first.
When one specific detail still blocks the fix, ask one plain question, and only one pending question per thread. Don't ask a subjective question just to choose between plausible designs.
4. Check error sources
When it smells like an error, search Sentry (org builder-io) and Analytics
list-error-issues / get-error-issue by route, stack symbol, and error text.
Cite the match and its source. An unreadable source is "unavailable", never
"no match"; don't invent matches or repro steps.
5. Choose the fix altitude
Name the smallest invariant that explains the report and the boundary that owns it.
- One isolated report: fix the local seam and add a focused regression check.
- Repeated or cross-surface reports: look at the shared primitive or contract, once you've confirmed they fail the same way.
- Missing capability or wrong tool: fix discovery, the registry, or the action contract rather than coaching the agent around the gap.
- Source works but live differs: check build, deploy, and version state first.
Never turn one data point into a global agent rule.
6. Fix and verify
Reproduce first where you can, keep each fix narrow and mapped to an item, and
add a focused test when the risk warrants it. Verify as verifying-changes describes: a test that fails before and passes
after, or the flow working in the running app. For UI, look at the real screen
with a browser tool you already have; don't script browser automation for a
small fix. Reading the code is not verification.
For approved UX work, make the smallest change that clarifies the current state
or next step: remove competing elements before adding controls or text, and put
secondary actions in a menu or popover. Read frontend-design.
7. Report
Reply with a status, not an implementation report: one plain line per item, no file paths, PR numbers, or test counts unless asked:
## Feedback Status
- <item> - Fixed. It will be on beta shortly after it merges.
- <item> - Needs a person: <what's needed, from whom>.
- <item> - Needs your answer: <one question>.
- <item> - Not planned.
Don't say Fixed until it's verified.