skills.rest
ensj
ensjCOMMUNITY
Explore all skills by this author →
Website𝕏@_ensjGitHub

assemble-tier

Assembles the cheapest model-tier assignment for a composition that ties the all-opus baseline.

Updated Jun 20, 2026
Search Tags:#llm evaluation#cost optimization#model tiering#prompt composition#statistical evaluation#escalation loop#tier profile
One-click install
npx skills add https://github.com/ensj/skill-eval-framework --skill assemble-tier-ensj
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: assemble-tier
Source: https://github.com/ensj/skill-eval-framework/tree/main/.claude/skills/assemble-tier
Command: npx skills add https://github.com/ensj/skill-eval-framework --skill assemble-tier-ensj

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Running every step of a multi-step prompt composition on the most expensive model tier wastes budget, but downgrading blindly risks compounding quality loss. This Skill finds the cheapest model-heterogeneous assignment that statistically ties the all-opus composition end-to-end. ## Core Features & Use Cases - All-opus reference measurement: Runs the composition once at all-opus across eval cases and seeds to establish the quality baseline. - Verify-escalate loop: Starts from a floor assignment derived from per-step tier-profiles, then escalates individual steps to higher tiers whenever compounding loss is detected. - Statistical decision CLI: Uses the tier-assemble CLI to emit decisions (need_seeds, escalate, ship, stop) with epsilon-based equivalence testing, and writes a frozen composition-tier-profile.json plus a report. - Use Case: After certifying per-step tier profiles with the tier-ladder skill, ask to "tier the composition" to get the cheapest safe mix of haiku, sonnet, and opus across steps. ## Quick Start Tier the composition to assemble the cheapest model assignment that still ties the all-opus baseline end-to-end.

Frequently Asked Questions about assemble-tier

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I reduce model costs for a multi-step prompt composition?▼

Run the assemble-tier skill to find the cheapest mix of haiku, sonnet, and opus per step that still ties the all-opus composition. It starts from a floor assignment based on per-step tier-profiles and escalates steps only when compounding loss is detected.

How does the verify-escalate loop decide when to escalate a step?▼

The tier-assemble CLI compares heterogeneous run scores against the all-opus reference rows using an epsilon equivalence rule. If the decision is escalate, the assignment moves specific steps to higher tiers and the heterogeneous side re-runs until it ships or stops.

What inputs are required before assembling a tiered composition?▼

Each step skill needs a frozen winner.json and a tier-profile.json from the tier-ladder skill, plus built pattern and CLI bundles from npm run build and esbuild. Steps without a tier-profile are treated as opus with a warning.

Does the all-opus baseline re-run on every escalation round?▼

No. The all-opus reference is measured once and reused across all escalation rounds; only the heterogeneous assignment re-runs. This keeps the evaluation cost bounded while the loop moves monotonically toward all-opus.

What happens if the cheapest assignment cannot match all-opus quality?▼

The loop escalates steps toward higher tiers and never ships a quality loss on doubt. If no assignment suffices, the CLI returns a stop decision with best_insufficient status, which is recorded in the composition-tier-profile.

Related Skills

Explore More Software Engineering Skills→
jtucker9jtucker9

scan

Scan codebases to estimate project scope and generate three-tier pricing.

Community
Advanced
councilofwizardscouncilofwizards
1

tier2-test

Invoke tier1-test and verify docs/research/poc-tier1-output.md for tier1_executed.

Official
Intermediate
agentscope-aiagentscope-ai
809

metric-design

Select, configure, and combine evaluation graders into executable OpenJudge pipeline code.

Official
Advanced
t0ddc3byt0ddc3by
58

early-churn-downgrade-signal-detection

Classify churn and downgrade risk into three tiers with Markdown reports.

Community
Advanced