3.0 KiB
Daily Patch Review Loop Policy v1
This policy exists because a pack can pass schema checks and still be boring, weird, or too generated.
The daily pack must be fixed in patch mode unless a mass rewrite is explicitly justified.
Core Rule
Review the full pack. Mark only questions that fail. Fix only the marked question IDs.
Do not rewrite passing questions just because a rewrite is happening.
Passing questions are frozen unless they later fail a specific rule.
Required Inputs
Before editing daily questions, the writer must have:
- the current JSON file
- the current daily guide files
- the previous review findings, if any
- a marked fixes CSV or JSON from the review pass
Required Review Pass
Each question must be checked for:
- schema validity
- weekday tag correctness
- prompt clarity
- option fit
- fun-first value
- therapy or wellness voice
- household admin drift
- awkward phrasing
- repeated opener or repeated mechanic
- too-random silliness
Mark Before Fixing
Every failed question must be marked with:
- question ID
- weekday
- current prompt
- failure reason
- fix scope
- whether metadata may change
Use these fix scopes:
prompt_onlyoptions_onlyprompt_and_optionsmetadata_onlydelete_or_replace
Fix Only What Failed
When patching:
- preserve the original ID unless the ID itself is wrong
- preserve
typeunless the type is wrong - preserve
accessunless the count split is wrong - preserve
depthunless the import code or schema requires a change - preserve weekday tags unless the tag is wrong
- preserve compatibility tags unless app code no longer needs them
- change only the prompt and options for normal content fixes
Mass Rewrite Exception
A mass rewrite is allowed only when more than 60 percent of a weekday or pack fails for the same root reason.
The rewrite report must state:
- failure percentage
- shared root cause
- why patching individual IDs would be worse
- what will be preserved
- how the new pack will be reviewed
Without that report, mass rewriting is not allowed.
Loop Stop Conditions
The loop stops only when:
- every marked question has been fixed
- every fixed question has been reviewed again
- remaining hard flags are 0
- duplicate checks pass
- option-answer checks pass
- two random samples per weekday pass cleanly
- the review findings file says
production_ready
Do not call the pack production ready with known unresolved content flags.
Fun But Grounded Rule
Fun does not mean random nonsense.
Reject prompts or options that feel like a carnival generator, such as too many fake awards, too many snack drafts, too many mascot jokes, or cute objects that do not actually fit the prompt.
A daily question should feel playful and usable by adults, not like a children's party game accidentally shipped inside a couples app.
Required Outputs
Every daily patch must output:
- updated JSON
- validation report
- marked fixes file
- patched fixes file
- remaining flags file
- review findings file
- short apply instructions