| name | ai-collab-scorecard |
| description | Evaluate long-term AI/Codex collaboration corpora, chat logs, prompts, revisions, and produced artifacts to build an evidence-grounded cognitive work capability scorecard. Use when the user asks to assess AI collaboration records, abstract a colleague's capability profile from their AI-assisted work, compare collaboration quality across people, or convert long-term chat/productivity traces into neutral dimension scores. Do not use as an IQ test, personality diagnosis, hiring decision automation, or clinical assessment. |
AI Collab Scorecard
Purpose
Assess observable cognitive work behavior in long-term AI collaboration records. The result is a neutral scorecard, not an IQ estimate or total-person judgment.
Use references/setup-and-data-sources.md when installing, validating, or collecting historical collaboration records. Use references/sampling-protocol.md and scripts/build_corpus_manifest.py to freeze corpus scope. Use references/evidence-abstraction-pipeline.md and scripts/extract_collab_signals.py to turn raw history into five-dimension evidence. Use the detailed scoring anchors in references/rubric.md when assigning scores. Use references/output-schema.md for machine-readable output. Use references/worktypes.md and scripts/classify_worktype.py when deriving worktypes. Use references/share-card-image.md, references/share-copy.md, scripts/render_share_card_png.ps1, and scripts/render_share_card.py when the user asks for a shareable image/card. Use references/asset-production/worktype-avatar-brief.md only when producing or replacing the fixed worktype illustration asset batch outside runtime.
Respond in the user's language. For Chinese users, all human-facing conclusions, headings, score labels, dimension names, worktype descriptions, status notes, evidence bullets, limits, and share-card explanations must be in Chinese. Keep English only for machine-readable JSON keys, script arguments, file paths, fixed type_id values, or an optional appendix when cross-person comparison requires canonical English labels.
Interaction Mode
- Default to low-friction execution for end users: one consolidated permission request at most, concise progress, and one final Chinese result.
- For first-time install/run requests, prefer
scripts/bootstrap.ps1 as the deterministic entrypoint. It handles repository acquisition, installation/update, environment diagnostics, validator repair, and render self-check before the assessment continues.
- If the user already grants network, write, install, validation, rendering, and read-only Codex/software history permissions in the same message, do not ask those permissions again except for platform-required tool approvals.
- Do not require a new Codex window after installation. If
$ai-collab-scorecard is not discoverable in the current session, read the installed SKILL.md and references/setup-and-data-sources.md from disk and continue the same workflow in the current conversation.
- Suppress raw diagnostic JSON, install commands, repeated validation logs, and long environment explanations unless a check fails or the user asks for debug detail.
- Avoid duplicate conclusions: output one short self-check summary, then one complete scorecard. Do not repeat the same scores in multiple prose blocks unless it clarifies a decision.
Guardrails
- Score only evidence visible in the corpus: prompts, corrections, constraints, decisions, review comments, artifact changes, and follow-up results.
- Separate the user's contribution from the AI's contribution. A polished final artifact is weak evidence unless the corpus shows how the user steered, rejected, corrected, or validated it.
- Prefer "insufficient evidence" over confident inference when the corpus is thin, narrow, edited, or mostly AI-generated.
- Do not infer mental health, moral quality, personality type, or exact IQ. If asked about IQ, state that this skill measures AI-assisted cognitive work performance, not psychometric intelligence.
- Normalize by role, domain, task difficulty, and available context before comparing people.
- Keep the serious assessment and any playful/shareable derivative strictly separate. Humor may make the result easier to share, but must never change scores, evidence, confidence, or capability claims.
Workflow
-
If the user asks how to install, use, or reuse the skill, read references/setup-and-data-sources.md first:
- For a new user, run
scripts/bootstrap.ps1 when available instead of manually reproducing install steps
- Verify the skill folder is installed under the active Codex skills directory
- Run the basic skill validation when useful
- Explain required runtime assumptions before collecting private history
- If installation, validation, rendering, or corpus acquisition is incomplete on the current machine, enter the completion recovery protocol in
references/setup-and-data-sources.md
- Do not stop at the first missing dependency or configuration difference; diagnose the gap, ask for the minimum needed permission, repair what can be repaired, and continue toward a complete scorecard
-
Define the assessment scope and data source:
- Person or team being assessed
- Time window and corpus sources
- Role/domain expectations
- Whether the goal is self-review, coaching, team abstraction, or comparison
- Corpus manifest or sampling method from
references/sampling-protocol.md
- If the user asks to use "all current Codex/software history" but has not explicitly granted local-history access in this turn, first ask for permission to inspect current Codex/software history read-only; do not assume elevated local access
- After the user grants permission, treat the current local Codex history as the intended corpus: inspect only Codex-owned or clearly session-related history locations read-only, freeze a manifest, and do not require manual pasted logs unless the local history is unavailable
- If the user asks for "all Codex history" without explicit authorization, use only user-authorized exports, visible conversation context, or explicitly allowed local history directories; do not silently scrape unrelated private app data
-
Build an evidence map:
- Task types and difficulty
- User prompts and constraints
- AI outputs the user accepted, rejected, or revised
- User corrections, tests, validations, and final artifacts
- Repeated patterns across tasks
- Use
references/evidence-abstraction-pipeline.md for session segmentation, evidence tagging, and dimension mapping
-
Score the five minimum dimensions:
- 问题定义与约束控制 (
problem_framing)
- 推理建模与机制抽象 (
reasoning_modeling)
- 证据校验与验证纪律 (
evidence_verification)
- AI 编排与批判性审阅 (
ai_orchestration)
- 交付转化与迭代沉淀 (
delivery_iteration)
-
Produce dimension scores and a composite score:
- Score each dimension as an integer from
0-100
- Compute the weighted composite as an integer from
0-100 using references/rubric.md
- Use the required component scoring standard in
references/rubric.md; do not assign scores by overall impression
- Use the repeatability worksheet in
references/rubric.md before finalizing each score
- Report percentile-like position only when a real comparison set exists; otherwise do not invent one
- Include confidence for each score:
low, medium, or high
- Attach 2-5 concrete evidence bullets per dimension
-
Derive the worktype when requested or useful:
- Use the full five-dimension score vector, not only the strongest dimension
- Match against fixed prototypes in
references/worktypes.md
- Prefer
scripts/classify_worktype.py to compute the primary type, fit score, secondary tendency, and risk modifiers from a score JSON
- For Chinese reports, display
serious_name_zh, share_name_zh, secondary_tendency_zh, risk_modifiers_zh, and band_zh; do not expose English enum values in the main conclusion
- Do not invent new canonical type names
-
Explain the profile:
- Strongest stable capability
- Most important bottleneck
- Best-fit collaboration pattern
- Risks of over-inference
- One or two next improvement levers
-
If the user asks for a shareable, viral, or image version, create a separate "Share Card" only after the serious scorecard:
- Derive it from the scored dimensions and evidence, not from vibes.
- Make it light, memorable, and mildly funny.
- Keep it respectful and workplace-safe.
- Avoid IQ claims, personality diagnosis, fixed identity claims, insults, medicalized labels, or hiring recommendations.
- Include a tiny disclaimer that the card is a playful summary of the evidence-based scorecard.
- If creating an image, select a confirmed pre-generated worktype illustration from
assets/worktype-illustrations/manifest.json and render important text with a deterministic renderer.
- Use
scripts/render_share_card_png.ps1 for the final standard PNG card on Windows. It loads the deterministic contract from references/render-lock.json, uses the bundled pinned font from assets/fonts/NotoSansSC-VF.ttf, and fails if the font hash changes. Use scripts/self_check_render_determinism.ps1 after renderer or layout changes. Use scripts/render_share_card.py only for SVG debugging or editable inspection.
- Use
references/share-copy.json for standard Chinese share-card copy defaults and enforced text-length/layout limits.
- Do not call an image model during scorecard/card generation. Image models may be used only when producing the fixed asset batch for the skill, following
references/asset-production/worktype-avatar-brief.md.
- When returning a rendered share card, include both a direct image file link and a direct link to the containing output directory.
Output Template
# AI 协作力评分卡
## 测评范围
- 测评对象:
- 语料范围:
- 角色/领域:
- 测评目标:
- 证据置信度:低/中/高
## 分数
| 维度 | 分数 /100 | 权重 | 置信度 | 解读 |
|---|---:|---:|---|---|
| 问题定义与约束控制 | | 22 | | |
| 推理建模与机制抽象 | | 22 | | |
| 证据校验与验证纪律 | | 20 | | |
| AI 编排与批判性审阅 | | 18 | | |
| 交付转化与迭代沉淀 | | 18 | | |
综合分:`__/100`
档位:
## 评分拆解
每个维度保留以下拆解:
`核心行为 /45`、`跨任务一致性 /20`、`人类主导性 /20`、`难度迁移 /15`、`反例修正`、`上限限制`、`最终分`。
## 量化画像
- 五维向量:`[问题定义, 推理建模, 证据校验, AI 编排, 交付迭代]`
- AI 协作工作型:
- 工作型匹配度:`__/100`
- 第二倾向:
- 风险修正:
- 置信度:
## 各维度证据
## 能力画像
## 主要瓶颈
## 可迁移给同事的抽象
## 非声明与限制
Scoring Discipline
When evidence is mixed, score the repeated pattern, not the best sample. When task difficulty differs sharply, report separate sub-scores by task class instead of forcing one number.
Do not hide uncertainty inside smooth prose. A useful scorecard is allowed to say: "high output quality, but low confidence in human contribution because the corpus lacks correction traces."
Optional Share Card
Use a share card only as a derivative layer. It should be fun to read, but it is not the source of truth. For a standard final image on Windows, use scripts/render_share_card_png.ps1 with a JSON result that follows references/output-schema.md. For SVG debugging or editable inspection, use scripts/render_share_card.py. Both renderers select confirmed assets from assets/worktype-illustrations/manifest.json; deterministic placeholder art is only a draft fallback until the fixed image batch is approved.
# AI 协作工作型分享卡
类型:
含义:
这个人倾向于如何使用 AI:
高光能力:
喜剧化风险:
最适合的 AI 配置:
严肃评分卡链接:
小字声明:
Good share-card names describe work behavior, not inner essence. Prefer names like "Problem Architect", "Evidence Auditor", "Workflow Blacksmith", "Prompt Air-Traffic Controller", or "Scope Boundary Patrol". Avoid names that imply intelligence rank, moral worth, mental health, or permanent personality.