watcher-hermes/hermes-skills/batch-triager
Claude Code and agent skill by watcher-hermes/hermes-skills
- community
Checking that the SKILL.md file is still in the repository.
Triage LLM workloads into interactive / semi-interactive / batch lanes, compute stacked discount (batch + cache) savings, and flag mis-triaged workloads.
Install
npx vskill@latest install watcher-hermes/hermes-skills/batch-triagerSKILL.md
The same SKILL.md was published earlier as watcher-hermes/hermes-skills/batch-triager.
Given a workload (name, user expectation for latency, traffic volume, shared prompt structure), produce a triage + cost plan.
Produce:
- Lane. Interactive (TTFT-bound, sync), semi-interactive (minutes OK, async queue), or batch (by-morning OK, batch API). Justify with the specific user expectation.
- Current cost. Compute monthly cost at current configuration (sync, no cache, etc.).
- Target cost. Compute cost after recommended config (batch + cache or sync + cache). Express as % of current.
- Migration plan. Provider-specific steps (pick the one that matches the workload's model, not both):
- OpenAI: migrate to
/v1/batches. Prompt caching is enabled automatically for eligible prompts (≥1024 tokens) — nocache_controlto set. Optionally passprompt_cache_keyfor tighter attribution. - Anthropic: migrate to Message Batches. Cache reuse requires explicit
cache_controlblocks (e.g.,{"type": "ephemeral"}) on the cacheable prompt spans; batch discount stacks with cached-read pricing. - Both: instrument a success/failure webhook and a spillover lane to sync for batches that miss their turnaround window.
- Risk. What if the batch turnaround is 20 hours at P99? Name the downstream system behavior (email delivery, queue spillover to sync).
- Observable. Metric that catches mis-triage: batch job completion latency P95; alert if > 12 hours.
Hard rejects:
- Running an overnight pipeline in sync mode without batch when the user only needs "by morning" latency. Refuse — call out the ~90% leaked spend.
- Promising batch for anything with a sub-15-minute user expectation. Refuse — batch SLA is 24h.
- Ignoring prompt caching on a batch workload with shared system prompt. Refuse — the stacked discount is the point.
Refusal rules:
- If the workload is marketed as "real-time" but the actual user expectation is minutes, require explicit confirmation before recommending batch.
- If the workload targets a provider without prompt caching in batch (e.g., any custom or self-hosted stack without KV-prefix reuse), note that only the batch discount applies and recompute without stacked savings. OpenAI batch caching is automatic; Anthropic batch caching requires explicit
cache_controlblocks. - If the workload has strict latency SLA (e.g., P99 < 60s) refuse batch outright — it belongs on a different lane.
Output: a one-page triage with lane, current cost, target cost, migration steps, risk, observable. End with a cadence: re-triage all workloads quarterly as product surface changes.
Security
- Security scanPassed, 95 of 100Pattern checks on the skill's files.
- Model reviewPassed
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingskills/batch-triager/SKILL.md
Popularity
- 7-day trend score
- 5
Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/watcher-hermes/hermes-skills/batch-triager)