# Ase Meta Evaluate > > - **Type:** Skill - **Install:** `agentstack add skill-rse-ase-ase-meta-evaluate` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [rse](https://agentstack.voostack.com/s/rse) - **Installs:** 0 - **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills) - **Latest version:** 0.1.0 - **License:** Apache-2.0 - **Upstream author:** [rse](https://github.com/rse) - **Source:** https://github.com/rse/ase/tree/master/plugin/skills/ase-meta-evaluate - **Website:** https://ase.tools ## Install ```sh agentstack add skill-rse-ase-ase-meta-evaluate ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About @${CLAUDE_SKILL_DIR}/../../meta/ase-control.md @${CLAUDE_SKILL_DIR}/../../meta/ase-skill.md @${CLAUDE_SKILL_DIR}/../../meta/ase-getopt.md Evaluate Alternatives $ARGUMENTS *Evaluate* *alternatives* through a weighted multi-*criteria* decision matrix. 1. - From the , try to derive the overall reason for the evaluation. If no such reason can be derived, assume generic comparison. - Output the determined reason with just the following and do not output anything else: **REASON**: ** 2. - From the derive the two or more alternatives (K=1-N) the user wants to be evaluated. Do not output anything. - If fewer than two alternatives could be derived (N and *stop the entire flow* immediately without executing any further steps: **ERROR: INSUFFICIENT ALTERNATIVES**: at least two are required for a comparison! - For each alternative (K=1-N), decide whether you have all necessary information at hand to give it the proper, unique, short, and *concise name* . If you are unsure, use the `ase-meta-search` skill (at most one invocation per alternative, drawing from the *skill-wide shared budget* of at most 8 `ase-meta-search` invocations in total across STEP 2 and STEP 3 combined) to find out more and adjust the name . If still unsure after the shared budget is exhausted, pick the best-guess concise name and proceed. Do not output anything. - For each alternative (K=1-N), decide which *one* of *USP* (Unique Selling Point -- what makes it unique), *Crux* (what you should notice), or *Gotcha* (what you should not stumble over) is its single most distinguishing perspective, and remember this as an (K=1-N) formatted like `: ` where is one of `USP`, `Crux`, or `Gotcha` and is a 1-6 word hint. Do not output anything. - For the set of alternatives, decide what the 1-6 word long name of the *class of alternatives* is. Do not output anything. - For each alternative (K=1-N), decide whether it is a genuine member of . If any is *not* a member (i.e. the alternatives are not mutually comparable within a single class), collect the subset of non-member alternatives into the space-separated list (the for which the membership decision was negative), output the following and *stop the entire flow* immediately without executing any further steps: **ERROR: INCOMPARABLE ALTERNATIVES**: ** "> ⚑ **** (*member of a different class*) - Output the determined, individual alternatives as a Markdown *table* with just the following and do not output anything else: **ALTERNATIVES**: ** | ⚑ *Alternative* | ⚖ *Hint* | | :--------------------- | :-------- | | ⚑ **** | | [...alternatives K=2-(N-1) for N>2...] | ⚑ **** | | 3. - From the , try to derive the criteria (L=1-M) for the evaluation. Do not output anything. - For each criterion (L=1-M), decide on its from the value set { 4.00, 2.00, 1.00, 0.50, 0.25 } (from most important, over normal, to less important). Do not output anything. - Ensure the final number of criteria is always within the range of minimum 8 and maximum 12: if fewer than 8 criteria were requested, use the set of alternatives to decide on additional criteria which potentially allow best to triage the alternatives, take the into account, and use the `ase-meta-search` skill (drawing from the *skill-wide shared budget* of at most 8 `ase-meta-search` invocations in total across STEP 2 and STEP 3 combined) to find out about the potentially still missing criteria and assign their . If still under 8 criteria after the shared budget is exhausted, fill the remaining slots from existing knowledge without further searches; if more than 12 criteria were requested, drop the criteria with the smallest until 12 remain. Do not output anything. - To prevent a single high-weight criterion from dominating the weighted sum (the weight set is geometric ×2 while the evaluation Likert scale is linear), assign weight 4.00 to *at least one* and *at most two* criteria, and weight 2.00 to *at least two* and *at most three* criteria. Symmetrically, to prevent a long tail of negligible-weight criteria, assign weight 0.50 to *at most two* criteria, and weight 0.25 to *at most one* criterion. Do not output anything. 4. - For each alternative (K=1-N) and each criterion (L=1-M), decide on the evaluation , which means how well the alternative meets the criterion on a Likert-scale from { -2, -1, 0, +1, +2 } (from worst, over neutral, to best). Do not output anything. - Then, calculate the ratings (K=1-N) for all alternatives in a single call by invoking the `ase_decision_matrix(matrix: [ [ , , ..., ], ..., [ , , ..., ] ])` tool of the `ase` MCP server. The tool returns an array of N numerical values, where the K-th entry is the product-sum of all weights (L=1-M) and the evaluation (L=1-M) for alternative . Retain the *raw, unrounded* for use in STEP 5, but round to 2 decimal places *for display only* when emitting it in the table below. Do not output anything. - Output the resulting *Weighted Decision Matrix* as a Markdown *table* with just the following and do not output anything else. When emitting the table, render *one column per alternative* (K=1-N). **EVALUATION**: *Weighted Multi-Criteria Decision Matrix* | ⦿ *Criteria* | ⚖ *Weight* | ⚑ **** | [...alternatives 2-(N-1)...] | ⚑ **** | | :------------ | ----------: | ---------------------: | ---------------------------: | ---------------------: | | | | | [...evals 2-1..(N-1)-1...] | | [...criteria L=2-(M-1) for M>2...] | | | | [...evals 2-M..(N-1)-M...] | | | **RATING** | | **** | [...ratings 2-(N-1)...] | **** | 5. - The best alternative (K=1-N) is the alternative whose *raw, unrounded* (i.e. the product-sum from STEP 4, *before* the display-only rounding) is the maximum rating value across all alternatives. Do not output anything. - The second best alternative (X=1-N, X != K) is the alternative whose *raw, unrounded* is the second largest rating value across all alternatives. Do not output anything. - If multiple alternatives share the second-largest raw rating, pick any one of them as ; the resulting and are unaffected by the choice, so the downstream output is deterministic. Do not output anything. - Determine rating distance between and from their *raw, unrounded* ratings by calculating: = - . Do not output anything. - Determine rating distance percentage between and from their *raw, unrounded* ratings as follows. If is exactly zero, skip the division entirely (it is undefined against a zero anchor) and classify directly by instead of routing a raw rating magnitude through the ratio thresholds below: if is also exactly zero, both best and second best rate zero, which is a genuine tie, so set = 0 to fall into the *MULTIPLE BEST* branch below; otherwise > 0 means the second best rates strictly negative (since is the maximum), so there is a clear positive gap to a worse runner-up and is an unambiguous winner, hence set = 1 (a sentinel exceeding all small-distance thresholds) to fall through to the plain *BEST ALTERNATIVE* branch. Do not output anything. - Otherwise ( is non-zero), calculate: = / abs(). Do not output anything. - By construction, is the maximum rating across all alternatives, so >= 0 always holds; using abs() keeps sign-stable across all rating regimes. Note that when itself is negative, the denominator anchors to a poor best rating and small gaps can appear large; the all-negative regime is surfaced as a dedicated warning branch below. Do not output anything. - If is less than zero (i.e. all alternatives rate negatively and the "best" is merely the least-bad), stop the flow after you output just the following and do not output anything else. This all-negative check is evaluated *before* the small-distance checks below, so that an all-negative alternative set with a small winning gap surfaces the all-negative warning rather than being masked by the multiple-best or small-distance branches: **BEST ALTERNATIVE**: ⚑ **** ⚠ *ATTENTION: all alternatives rate negatively; this is the least-bad choice, not a strong winner!* - Otherwise, if is less than 0.01 (i.e. is effectively zero relative to abs()), stop the flow after you output just the following and do not output anything else: **ERROR**: ✘ *MULTIPLE BEST ALTERNATIVES FOUND*, ⚠ *Please give some hints on the criteria to ensure a single best alternative!* - Otherwise, if is less than 0.10, stop the flow after you output just the following and do not output anything else: **BEST ALTERNATIVE**: ⚑ **** ⚠ *ATTENTION: small distance to second best alternative!* - Otherwise ( is greater than or equal to 0.10), output just the following and do not output anything else: **BEST ALTERNATIVE**: ⚑ **** ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [rse](https://github.com/rse) - **Source:** [rse/ase](https://github.com/rse/ase) - **License:** Apache-2.0 - **Homepage:** https://ase.tools Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-rse-ase-ase-meta-evaluate - Seller: https://agentstack.voostack.com/s/rse - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.