Written by subagents running under the writing-practitioner-guides skill,
one guide per experiment, 22,566 lines. Each walks a reader from baseline
capture through injection, injection verification, observation and recovery.
Section 3 carries the weight in most of them. Injection failed silently nine
times in this lab, and a failed injection looks exactly like no effect — so
the guides verify the target is actually in the intended state before
reading any result. A-4 makes virsh list the only proof because the node
reads Ready for 40 seconds after the machine is off; A-5 makes the packet
counter the sole go/no-go because a rule on the wrong node produces an empty
result that reads like a finding; A-6 quotes the run where 적용완료 was
printed between four Cannot find device "eth0" lines.
The traps the guides are built around are ones that invert a conclusion
rather than merely annoy:
A-0 emptying the session table without a restart leaves cache entries
that get counted as replication arriving
A-2 dropping -o /dev/null fuses body and status into one string
A-3 presence of "ready to accept connections" instead of its timestamp
B-2 row count alone reads an UPDATE as nothing having happened
B-4 tr ',' '\n' splits ["admin","editor"] so only admin is seen
B-7 no login screen means the cookie died and SSO re-authenticated
C-1 counting sessions without joining realm counts your own kcadm one
D-1 kubectl exec without -i restores nothing and still exits 0
D-4a "ran with error output" is what success looks like
Every quoted block is copied from docs/evidence/ and marked 실측; reshaped
commands are marked 미검증 rather than passed off as measured. Where a source
document carries a ★ correction the guides follow the corrected claim — A-7's
REVOKED_TOKEN hypothesis, C-1's session count, B-2's schema attribution.
Two hazards are stated rather than smoothed over: B-6 deletes a key that
cannot be recreated, and D-1/D-4 need host sudo, which asks for a password,
so those steps say a person must type them.
Audit over all 26: 672 interpretation pairs, 486 evidence citations, 117
undo sections, and zero occurrences of the patterns the skill forbids —
no python data processing, no deprecated kubectl get endpoints, no
placeholders, no bare kcadm.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
DongHyeonka