Found while canonicalising this skill from claw (culturebotai-claw#342). CultureMech carries
the same defect, filed as CultureBotAI/CultureMech#418.
.claude/skills/schema-gap-analysis/SKILL.md carries a table headed "CommunityMech-specific
state (as of 2026-05-17 pass)":
| Surface |
Records claimed |
Errors claimed |
kb/communities/*.yaml |
261 |
0 |
data/isolates/**/*.yaml |
5 |
4 ('id' is a required property) |
Measured against origin/main on 2026-09-04:
kb/communities/*.yaml — 325 files, not 261.
data/isolates/*.yaml — 4 files, not 5.
- Isolates missing a top-level
id: — 0, not 4.
Every number is wrong, and the prose that follows them is actively misleading: it instructs
a reader to "just add the missing id: to each isolate YAML and re-run step 2" for a
defect that no longer exists in any of the four files.
Nothing flagged it, because prose has no test. It has read as current fact for three and a
half months.
The fix already exists in the fleet
TraitMech's audit-schema-gaps lists this as an anti-pattern — "Do not present a snapshot
count as current repository state" — and its schema-gap-analysis alias deletes its own
snapshot in favour of the commands that regenerate one.
Suggested: keep the shape of the section — which surfaces exist, which root class each
validates against, what kinds of error each tends to produce — and replace the counts and the
remediation instruction with the command that produces current ones.
The claw template for this skill now makes "do not write down the state" a managed region,
with these two measurements as its evidence.
Found while canonicalising this skill from claw (culturebotai-claw#342). CultureMech carries
the same defect, filed as CultureBotAI/CultureMech#418.
.claude/skills/schema-gap-analysis/SKILL.mdcarries a table headed "CommunityMech-specificstate (as of 2026-05-17 pass)":
kb/communities/*.yamldata/isolates/**/*.yaml'id' is a required property)Measured against
origin/mainon 2026-09-04:kb/communities/*.yaml— 325 files, not 261.data/isolates/*.yaml— 4 files, not 5.id:— 0, not 4.Every number is wrong, and the prose that follows them is actively misleading: it instructs
a reader to "just add the missing
id:to each isolate YAML and re-run step 2" for adefect that no longer exists in any of the four files.
Nothing flagged it, because prose has no test. It has read as current fact for three and a
half months.
The fix already exists in the fleet
TraitMech's
audit-schema-gapslists this as an anti-pattern — "Do not present a snapshotcount as current repository state" — and its
schema-gap-analysisalias deletes its ownsnapshot in favour of the commands that regenerate one.
Suggested: keep the shape of the section — which surfaces exist, which root class each
validates against, what kinds of error each tends to produce — and replace the counts and the
remediation instruction with the command that produces current ones.
The claw template for this skill now makes "do not write down the state" a managed region,
with these two measurements as its evidence.