eval-draft
Phase 1: Generate an initial evaluation draft and trace for a resource by navigating its website and scoring each criterion.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
Phase 1: Generate an initial evaluation draft and trace for a resource by navigating its website and scoring each criterion.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Phase 2: Critique a draft evaluation by searching the resource website for evidence that contradicts or complicates the draft's conclusions.
Phase 3: Reconcile a draft evaluation with its critique, producing a final trace and updated YAML.
Full three-phase evaluation pipeline: draft → critique → harmonize. Runs each phase as a separate agent for context isolation. Individual phases can also be run manually via /eval-draft, /eval-critique, /eval-harmonize.
Cross-check an existing data-sources evaluation by generating a fresh trace from the resource website and reporting differences. Use when you want to verify or re-verify an existing evaluation.
| name | eval-draft |
| description | Phase 1: Generate an initial evaluation draft and trace for a resource by navigating its website and scoring each criterion. |
| disable-model-invocation | true |
| argument-hint | ["resource-url"] |
Generate an initial evaluation and trace for the resource at $ARGUMENTS.
This is the first phase of a three-phase evaluation pipeline. Your output will be reviewed by a critic agent, so be thorough and precise — but do not over-reach. Stick to what you can find via normal navigation from the root.
Read project context.
docs/criteria.md for the full evaluation rubric.scripts/source.schema.yaml for the YAML schema.data-sources/*.yaml files to pick up idioms
(field formatting, commentary style, description style).Determine the resource ID.
Derive an ID from the resource name following existing conventions:
lowercase, hyphenated, concise (e.g. "UniProt" → uniprot,
"Gene Ontology Annotations" → go-ann). Check that
data-sources/{id}.yaml does not already exist. If it does, stop and
tell the user — they may want /cross-check instead.
Navigate the resource website and build the trace.
source-link).Write the draft trace file.
Write data-drafts/{id}/draft.md using the standard trace format from
CLAUDE.md. Every source URL must have a navigation path from the root.
Quotes must be verbatim.
Write the draft evaluation YAML.
Write data-drafts/{id}/{id}.yaml with all required and applicable fields:
id: the resource ID from step 2source: human-readable resource namesource-link: the provided URL (use https if available)status: completedescription: use the resource's own description of itself, quoteddata-field: the broad domain (e.g. biology, chemistry, genomics)data-type: the specific data type (e.g. pathway, variant, expression)data-categories: list of data category tagsdata-access: download and/or API entries with URLslicense: SPDX identifier, or unlicensed/custom/all rights reserved/etc.license-type: per the mapping in CLAUDE.mdlicense-link: URL to the license pagelicense-issues: list of {criteria, comment} for any failed criteria.
Use [] if none.license-commentary: list of strings noting findings. Follow the style
of existing evaluations — quote relevant passages, note only findings
and inconsistencies, do not confirm expected/unremarkable details.was-controversial: "false"provisional: "true"last-curated: today's date (YYYY-MM-DD)contacts: resource contact email if foundValidate.
Temporarily copy the YAML to data-sources/{id}.yaml, run make check,
then remove it. The draft YAML stays in data-drafts/ until a human
curator promotes it. Fix any validation errors before reporting success.
Report.
Summarize: resource name/ID, license found, star score, notable findings.
Remind the user this is a draft (provisional: "true") pending critique.
make check before reporting success.