Content Polisher Runbook
Owner: platform-runners Template:
Content Polisher(certified, gallery-visible) Expected duration: ~5 minutes Expected cost range: $0.10–$2.00 Tool profile:locked(no shell, no filesystem, no network)
The cheapest and safest certified template — pure LLM rewrite, no external services beyond the gateway. Failures are almost always either input-related or skill-install-related.
Known failure reasons
| Reason | What happened | First action |
|---|---|---|
humanizer_skill_install_failed | The Humanizer ClawHub skill could not be installed in the runner. | Check ClawHub status; the install command is npx clawhub@latest install humanizer. |
source_content_too_long_for_context | The content input exceeded the model’s context window. | Ask the customer to split into chunks; default model is configurable. |
source_url_unreachable | The content input was a URL and the URL failed. | Confirm with the customer the URL is public; the locked profile cannot follow private auth. |
language_not_supported | The Humanizer skill rejected a non-supported language. | Reproduce by sending a single test sentence; some non-Latin scripts are not yet covered. |
Diagnostic steps
- Is
contentplaintext or a URL? The URL path requiresweb_automationprofile semantics; this template’slockedprofile forbids outbound HTTP except via the gateway. Most URL failures here are intentional — the customer should pre-fetch and paste plaintext. - Did the Humanizer skill install succeed? Look for the
clawhub_skill_installedtimeline event. Missing means the install step failed; the run output will include the skill manager’s stderr. - Check the
content_draftartifact in run outputs. If the draft is empty or identical to the input, the LLM either refused or rate-limited. The Humanizer skill emits its own confidence score; below 0.3 typically means the input was already humanized or in an unsupported style.
Rollback / cleanup behavior
From the template’s rollback_or_cleanup_behavior:
- Stateless rewrite. Cancelling mid-run discards the polished draft entirely.
- The source content is never modified in place.
Escalation
The Content Polisher is the most cost-bounded template. If a customer
reports a $-figure that exceeds the published $2.00 cap, that’s a
billing-system bug — escalate to platform-runners + billing-ops with
the diagnostics bundle and the actual cost record from
cost_actual_usd.