Technology Job

One-line statement of the job this skill does.

Purpose

Use this when:

• Symptom or request that should route here.
• Another one.

Do not use this when:

• The problem is actually X — use other-skill instead.
• The change is routine and already covered by an existing runbook.

What this skill assumes: state what must be true (access level, tooling installed, etc.).

Operating Procedure

Work top to bottom. Do not skip to remediation.

1. Establish the environment. What cluster/account/host/version is this? Record it.
2. Collect evidence. Read-only commands only. Capture output before anything changes.
3. State the symptom precisely. "30% of requests return 503 since 14:20 UTC", not "it's broken".
4. Form hypotheses. Rank by likelihood given the evidence, not by familiarity.
5. Test the top hypothesis. Prefer a test that eliminates a whole class of causes.
6. Apply the smallest safe change. One change at a time.
7. Verify. See Verification. A command exiting 0 is not verification.
8. Document. What was observed, what changed, how to roll back.

Label every statement as Observed, Assumed, Recommended, or Unverified. When something is unknown, run an inspection command or ask one targeted question — do not guess identifiers, regions, or versions.

Investigation

Step 1 — <what this establishes>

command --with-flags <placeholder>

What it checks, the specific field or line to look at, and what each plausible value implies.

Step 2 — <what this establishes>

another-command

Same: signal, then meaning.

Decision Tree

Symptom
├── Condition A observed
│   ├── Sub-condition → likely cause, next check
│   └── Sub-condition → likely cause, next check
└── Condition B observed
    └── likely cause, next check

Common Causes

Safety

Read-only

Commands that only inspect. Safe to run unprompted.

• command --dry-run

Potentially disruptive

Restarts connections, evicts workloads, changes live configuration. Requires an explicit decision and a stated rollback.

• command --apply — blast radius: ..., rollback: ...

Destructive

Delete this subsection if risk is not destructive. If it stays, every entry needs the backup or confirmation that must happen first.

• command --delete — irreversible. Requires: verified backup, explicit human confirmation, blast-radius check naming every affected resource.

Verification

State the before and after in observable terms.

Before: <metric or status>
After:  <metric or status>

Verify at least two independent layers (e.g. the platform's own status and a client-side check).

Anti-Patterns

Bad: the shortcut people reach for. Better: what to do instead, and why the shortcut hides the cause.

Composition

• other-skill — when the investigation points at that layer.

References

• references/<file>.md — what it contains and when to read it.
• Official docs — what to look up there.