ThreadCloud SOPs

For the ThreadCloud team only

AI & Models

The control surface for everything AI does in ThreadCloud. Parse, enrich, classify, describe — all of it routes through this configuration.

Only the Owner edits this
The team should never change prompts, models, or API keys. If something looks off, refresh the page or email ThreadCloud Support at hello@threadcloud.io.

What lives here

Settings → AI & Models has four sections:

1. API key

The Anthropic API key. Stored encrypted, never echoed back in the UI after save.

To rotate:

  1. Get new key from console.anthropic.com
  2. Paste into the field
  3. Click Save
  4. Old key auto-revokes after 24h

2. Model selection

Which Claude model handles each workstream:

WorkstreamDefault modelWhy
Parse (Tier-1)claude-haiku-4-5Fast, cheap, structured extraction
Enrich (Tier-2)claude-haiku-4-5Same — accuracy-sensitive but bounded
Descriptionclaude-sonnet-4-6Prose quality matters more than speed
Classify (Legacy Enrich)claude-haiku-4-5Volume-sensitive, structured output

Override per-workstream if needed. Generally don’t.

3. Brand voice + description rules

The editorial voice used for storefront descriptions:

  • Brand voice — paragraph describing the tone (e.g. “Understated luxury. Inform, don’t sell. Lead with fabric and construction.”)
  • Structure — what to lead with (fabric → construction → fit → COO → occasion)
  • Avoid rules — words and phrases to never use (“elevate”, “timeless”, “effortless”, “perfect for any occasion”)

Changes apply to new descriptions and any re-generated descriptions. Existing descriptions stay as they were.

4. Prompt overrides (advanced)

Per-prompt overrides for specific extraction rules:

Prompt keyWhat it controls
description.styleDescription structure rules
description.cooFilterWhich countries get mentioned in description prose
description.bulletPointsWhether to use bullet points in descriptions
extraction.fabricContentCustom rules for fabric extraction
extraction.occasionRuleCustom rules for occasion classification

Edit at your own risk. Each override is a free-form prompt that gets injected at AI call time.

Sync controls

At the bottom of the page:

  • Sync sell-through now — triggers an immediate Shopify Order Sync (otherwise runs every 6 hours)
  • Reset AI cache — clears Anthropic prompt cache (rarely needed)
  • Re-run enrichment on all draft orders — useful after changing brand voice

Common patterns

After changing brand voice

  1. Save new brand voice
  2. Click “Re-run enrichment on all draft orders” — this regenerates descriptions for any orders in DRAFT status
  3. Existing live products on Shopify need a per-order re-push to pick up the new voice

When AI is throwing 429s (rate limits)

Anthropic rate limit hit. Two options:

  1. Wait — limits reset every minute, and the retry logic handles it
  2. Upgrade the Anthropic plan if you’re constantly hitting the wall

Don’t switch to a smaller model — that’s a different fix and usually wrong.

Anthropic outage

If status.anthropic.com shows an incident:

  • Don’t kick workers manually — they’re already retrying
  • Don’t change the API key
  • Don’t change the model
  • Just wait

Common mistakes

Pasting an invalid API key
Throws no error on save — only fails at next AI call. Test by triggering one classification on a small dataset after rotating the key.

Editing the description prompt before testing
Always test prompt changes against 5–10 styles before applying to a full order. Use a small draft order or the legacy enrichment dry-run.

Don’t switch to a non-Anthropic provider
The prompts are tuned to Claude’s structured-output behavior. OpenAI or Gemini will return malformed JSON and break extraction. Stick with Anthropic.

Next steps