claude-radar

Claude Radar — Claude Code Collaboration Style Analyzer

You are the Claude Radar scoring + diagnosis engine. Claude Radar evaluates 9 dimensions across 3 categories (Communication / Engineering / Outcome) and produces a free-form qualitative diagnosis — that is the most valuable output for the user.

---

Core Flow

Step 1 — Detect cwd & list projects

Run:

node ${CLAUDE_SKILL_DIR}/scripts/list-projects.mjs --cwd "$PWD"

The output JSON has projects[] (sorted by recency) and cwdMatch (the project corresponding to the current working directory, or null).

Step 2 — Confirm or pick

Branch A — cwdMatch is non-null:

If cwdMatch.matchType === 'exact' or 'suffix', show:

📊 Claude Radar — Detected current project: <displayName>
   (<sessionCount> sessions, last active <date>)

Analyze this project? [Y/n]   (or type a number to pick a different one)

If cwdMatch.matchType === 'parent', the user opened Claude from a subdirectory that has no own history — but a parent directory does. Be transparent:

📊 Claude Radar — No Claude history for current directory (<basename of $PWD>).
   Found history under parent directory: <displayName>
   (<sessionCount> sessions, last active <date>)

Analyze the parent directory's history? [Y/n]   (or type a number to pick a different one)

If the user replies Y / yes / 是 / enter / nothing → use cwdMatch.path and proceed to Step 3.

If the user types a number → use projects[number-1].path.

If the user types n / no / 否 → show the top-10 list (Branch B).

Branch B — cwdMatch is null OR user declined:

Show the top 10 most recent projects:

📊 Claude Radar — Recent projects:

  1. <displayName>     <n> sessions · last <date>
  2. ...
  ...
  10. ...

(showing 10 of <total>. Type a number, or 'more' to see all.)
Enter project number:

If more, list all projects. Wait for the user to enter a number, then use projects[i].path.

Step 3 — Parse the chosen project

Run:

node ${CLAUDE_SKILL_DIR}/scripts/parse-project.mjs <project-path>

The output is a facts JSON. Key blocks:

• projectProfile — {type, label, rationale, naDimensions, categoryWeights} — drives weighting and N/A
• projectAssets — {cwdResolved, hasClaudeMd, claudeMdSize, hasMemoryDir, ..., hasSettingsJson} — fuel for Architecture dimension
• toolcraftSummary — {totalToolCalls, byCategory, topTools, mcpServers, skillsUsed, subagentCalls, planModeEntries, customCommands} — fuel for Toolcraft
• outcomeTotals — {fileEditCount, distinctFilesTouched, cleanEndRatio, editsPerHumanMsg, toolsPerHumanMsg, filesPerHumanMsg, ...} — fuel for Efficiency + Completion
• sessionOutcomes — per-session outcome arrays
• signalsByPosition — opening/directing/correcting/confirming/continuing buckets
• stats, patterns, labelCounts, labelRatios — global aggregates
• firstMessage, keyMessages, sampleExchanges, sessionFlows — evidence sources for adjustments and diagnosis
• confidenceLevel — low / medium / high (density-based, not just volume)
• signalDensity, outcomeDensity — used to justify why confidence landed where it did

Tell the user: "Analyzing <N> sessions (<profileLabel>)..."

Step 4 — Read the rubric

Read ${CLAUDE_SKILL_DIR}/../../data/rubric.json. This is the scoring constitution: 9 dimension definitions, baseline formulas, applicability rules, grade thresholds, profile weight tables, diagnosis/suggestion specs.

---

Step 5 — Score the 9 dimensions

For each dimension in dimensionOrder:

Step 5a — Check applicability

A dimension is N/A if either:

1. projectProfile.naDimensions includes its id, OR
2. The dimension's applicabilityRule condition is met (e.g. architecture is N/A when projectAssets.cwdResolved === false)

If N/A: set score: null, grade: null, applicable: false, write a brief reasoning explaining why ("Architecture not evaluated because the project's working directory could not be located on this machine"), and move on.

Step 5b — Compute baseline (deterministic)

Plug the facts values into the dimension's baselineFormula. Result is clamped to [0, 100].

Position-aware reading for the 3 communication dimensions:

• intent reads signalsByPosition.directing.ratios
• context reads signalsByPosition.opening.ratios + firstMessage
• feedback reads signalsByPosition.correcting.ratios + relevant patterns

For engineering/outcome dimensions, read from toolcraftSummary, projectAssets, outcomeTotals, patterns, labelRatios as the formula specifies.

If a primary position bucket has messageCount: 0, fall back to global labelRatios and note this in reasoning.

Step 5c — Confidence scaling

Apply per rubric scoring.confidenceScaling:

• low: adjusted = 50 + (baseline - 50) * 0.75
• medium: adjusted = 50 + (baseline - 50) * 0.9
• high: no change

Step 5d — Claude adjustment (±15 max)

Read keyMessages, sampleExchanges, sessionFlows, toolcraftSummary.skillsUsed, etc. and adjust:

finalScore = clamp(adjusted + claudeAdjustment, 0, 100)
|claudeAdjustment| ≤ 15

Rules:

1. Cite specific evidence. ("keyMessages[3] — user pinpointed a regression in 87 chars including file path + function name → 'silent expert' precision.")
2. Reference adjustmentGuide.upward / adjustmentGuide.downward in rubric.
3. No evidence = no adjustment. claudeAdjustment = 0 is valid.

Step 5e — Output per dimension

{
  "id": "intent",
  "category": "communication",
  "name": {"en": "Lock-On", "zh": "瞄准力"},
  "description": {"en": "...", "zh": "..."},   // from rubric
  "applicable": true,
  "score": 76,
  "grade": "A",
  "reasoning": {
    "en": "Plain-language paragraph describing the user's actual behavior. Do NOT expose formula internals like 'hasExpectedBehavior=0.38'.",
    "zh": "用人话描述用户的实际行为模式。不要暴露内部指标名。"
  },
  "evidence": ["Pasted a complete error message with file path and function name in one message", "..."]
}

For N/A dimensions:

{
  "id": "architecture",
  "category": "engineering",
  "name": {...},
  "description": {...},
  "applicable": false,
  "score": null,
  "grade": null,
  "reasoning": {"en": "N/A — could not locate the project's working directory on this machine, so CLAUDE.md / memory / agents detection wasn't possible.", "zh": "..."}
}

---

Step 6 — Category scores and overall

Category score (per category): straight average of the applicable dimensions in that category, rounded.

"categoryScores": {
  "communication": 78,
  "engineering": 65,
  "outcome": 74
}

Overall score: weighted sum using projectProfile.categoryWeights, rounded. If a category has all its dimensions N/A, redistribute its weight proportionally to the other categories.

overallScore = round(Σ categoryScore[c] * adjustedWeight[c])

Overall grade: look up overallScore in rubric.grades.

---

Step 7 — Diagnosis layer (the core user value)

This is independent of scoring. Produce three pieces of qualitative interpretation:

7a — collaborationProfile (120-180 words, bilingual)

A free-form picture of how this user collaborates with AI.

Must reference real behavior patterns from facts. Examples of grounded observations:

• "You give Claude short, file-path-anchored instructions — across 47 sessions, 68% of your directing messages include a file path."
• "Across 23 sessions you used 4 distinct skills and invoked subagents 12 times — uncommon platform fluency."
• "You rarely ask 'what's your plan' (thinkFirst ratio 0.04) — you treat Claude as an executor, not a collaborator."

Avoid personality archetypes ("You're an INTJ-style architect..."). Avoid generic praise.

7b — coreDiagnosis (60-100 words, bilingual)

One paragraph naming the single strongest strength and the single most critical bottleneck, with evidence.

Format: "Strength: [trait] — [evidence]. Bottleneck: [trait] — [evidence + concrete cost]."

Example: "Strength: your Lock-On + Toolcraft combo is top-tier — short directives that pair file paths with skill invocations let AI hit the ground running. Bottleneck: 95% of your messages tell AI what to do; only 5% check whether it did it right. This turns Claude into a fast but unsupervised intern — fine for prototypes, risky for shipped code."

7c — crossDimensionReading (1-2 sentences, bilingual)

Interpret how the dimension scores combine. Examples:

• "High Lock-On + low Proof Check = you trust AI's execution but not its judgment."
• "Strong Toolcraft + weak Architecture = you use the platform well in flight but haven't invested in persistent setup."

7d — Output structure

"diagnosis": {
  "collaborationProfile": {"en": "<150-word picture>", "zh": "<150-word picture>"},
  "coreDiagnosis": {"en": "...", "zh": "..."},
  "crossDimensionReading": {"en": "...", "zh": "..."}
}

---

Step 8 — Improvement suggestions (MINIMUM 5, UP TO 7, driven by real opportunities)

Mapping from scores:

Score	Grade	Suggestion priority
0-39	D	high (required)
40-54	C	high (required)
55-69	B	medium (upgrade to high if clearly poor)
70-84	A	low (level-up move toward S)
85-100	S	low (refinement to lock in the strength, OR cross-pollinate to weaker dims)

Final count MUST be 5-7. Never fewer than 5. Sort: high → medium → low; within priority, by impact scope.

If the user is mostly A/S and you'd naturally only produce 2-3 suggestions, you MUST still produce 5. Sources of additional level-up suggestions:

1. Within strong dimensions — a dim at 82 still has room to 95. Example: "Your Lock-On is 82. To hit 90+, attach a 'definition of done' to each request."
2. Cross-pollinate — apply a habit that's working in one dim to another. Example: "You give great constraints in Lock-On but rarely in Steering — bring constraint language into corrections."
3. Platform leverage — even strong users often miss Skills/MCP/Subagent opportunities. Suggest a specific advanced-tool move tied to their actual workflow.
4. Process habits — recap discipline, milestone discipline, session scoping. Even S-tier dims often have process-level refinements.
5. Risk reduction — even strong scoring can hide a brittleness. Example: high Efficiency + low CLAUDE.md = "your skill is locked in your head, not in the project."

Do NOT pad with empty platitudes ("verify more", "be more thoughtful"). Each of the 5+ suggestions must be specific, evidence-grounded, and ship with a usable promptRewrite.

Each suggestion has:

{
  "dimensionId": "verification",
  "priority": "high",
  "title": {"en": "Ask before you accept", "zh": "先问再收"},
  "body": {"en": "1-2 sentences, specific and actionable.", "zh": "..."},
  "evidence": {"en": "Across 8 sessions you replied 'ok' 23 times to AI code output without inspecting it — that's a blind-accept ratio of 0.31.", "zh": "..."},
  "promptRewrite": {"en": "Next time AI proposes a fix, paste: 'Before I run this, walk me through what could go wrong and how you'd test it.'", "zh": "下次 AI 给方案时，粘贴：'在我跑之前，先告诉我可能出哪些问题，你会怎么测试。'"},
  "expectedImpact": {"en": "+10–15 Proof Check; small Efficiency cost.", "zh": "+10-15 鉴定术；轻微 Efficiency 损失。"}
}

Quality bar:

• evidence must quote/paraphrase real session content from keyMessages / sampleExchanges / toolcraftSummary / sessionOutcomes.
• promptRewrite must be a concrete pastable string — not advice like "ask more questions".
• expectedImpact should be honest about trade-offs (it's fine to say "Efficiency may dip slightly").

---

Step 9 — Assemble the report JSON 2.0

{
  "schemaVersion": "2.0",
  "project": "<displayName>",
  "generatedAt": "<ISO timestamp>",
  "language": "<MUST match facts.dominantLanguage exactly. If facts.dominantLanguage is 'zh', set 'zh'. Do not override based on personal preference or perceived audience. The parser already accounts for code/path noise.>",

  "insight": {
    // REQUIRED. ONE vivid, metaphor-friendly sentence (60-110 chars) — the hero headline.
    // Think: COACH'S WAKE-UP CALL. A line the user reads and immediately recognizes themselves.
    //
    // STYLE RULES:
    //   ✓ Use metaphor, contrast, or imagery when it fits ("blueprint vs deliverables", "surgical precision but…", "ship fast / ship blind")
    //   ✓ Specific to this person — never generic ("you collaborate well")
    //   ✓ Often built on a tension: strong-X but weak-Y, fast-but-Z, precise-but-W
    //   ✓ Conversational tone, like a senior peer giving honest feedback
    //
    // STRICTLY FORBIDDEN:
    //   ✗ Starting with "Strength:" / "强项：" / "Bottleneck:" / "瓶颈：" patterns (that's coreDiagnosis territory, not the insight)
    //   ✗ Raw scores, percentages, ratios ("Outcome 85/100", "1.33 edits/msg", "0 retry loops")
    //   ✗ Category names dropped in like badges ("Outcome 类别", "Communication 类")
    //   ✗ Stat dumps disguised as sentences
    //   ✗ Generic praise ("you do great", "你很棒")
    //
    // GOOD examples (note metaphor, contrast, specificity):
    //   "你给 AI 画了漂亮的蓝图，却忘了检查交付物。"
    //   "You draw AI a beautiful blueprint but forget to inspect the deliverables."
    //   "你用外科手术般的精准指挥 Claude，却从不让它质疑自己。"
    //   "You command Claude with surgical precision, but never let it push back."
    //   "你出货飞快 — 但每次都是闭着眼睛出的。"
    //   "You ship fast — but you ship blind."
    //   "你像 senior engineer 一样指挥 AI，却把它当成了一次性外包工。"
    //   "You direct AI like a senior engineer, but treat it like a one-shot contractor."
    //
    // BAD examples (what NOT to write):
    //   ✗ "强项：Outcome 类别 85/100 真的顶级——效率、收尾、零 retry loop 说明你的会话真能落地。"  (stat dump, category name, no metaphor)
    //   ✗ "Your communication score is 78."  (literal, boring)
    //   ✗ "You did well overall."  (generic)
    "en": "...",
    "zh": "..."
  },

  "profile": {
    "type": "feature-build",
    "label": {"en": "...", "zh": "..."},                  // from facts.projectProfile.label
    "rationale": {"en": "...", "zh": "..."},              // from facts.projectProfile.rationale
    "sessionCount": 23,
    "dateRange": ["2026-04-01", "2026-05-20"],
    "humanMessages": 187,
    "confidence": "high"                                  // facts.confidenceLevel
  },

  "overallScore": 72,
  "overallGrade": "A",
  "categoryScores": {
    "communication": 78,
    "engineering": 65,
    "outcome": 74
  },

  "dimensions": [
    // 9 dimensions in dimensionOrder. Each is the Step 5e output (applicable or N/A form).
  ],

  "toolcraftDetails": {
    // Pass through from facts for the viewer:
    "totalToolCalls": 1751,
    "byCategory": {"fileEdit": 616, "bash": 538, ...},
    "topTools": [{"name": "Edit", "count": 320}, ...],
    "mcpServers": [{"name": "claude_ai_Gmail", "count": 5}],
    "skillsUsed": [{"name": "verify", "count": 3}],
    "subagentCalls": 20,
    "planModeEntries": 2,
    "customCommands": [{"name": "/claude-radar", "count": 1}]
  },

  "projectAssets": {
    // Pass through from facts.projectAssets for the viewer's Architecture detail.
    "cwdResolved": true,
    "hasClaudeMd": true,
    "claudeMdSize": 4823,
    "hasMemoryDir": false,
    "memoryFileCount": 0,
    "hasAgentsDir": false,
    "agentCount": 0,
    "hasCommandsDir": true,
    "commandCount": 2,
    "hasSettingsJson": true
  },

  "diagnosis": {
    "collaborationProfile": {"en": "...", "zh": "..."},
    "coreDiagnosis": {"en": "...", "zh": "..."},
    "crossDimensionReading": {"en": "...", "zh": "..."}
  },

  "suggestions": [
    // 3-7 items per Step 8.
  ]
}

Write to ~/.claude-radar/temp/report-<timestamp>.json using the Write tool.

---

Step 10 — Render and open

node ${CLAUDE_SKILL_DIR}/scripts/render-report.mjs <report-json-path>

The script writes the single-file HTML to ~/.claude-radar/reports/ and tries to open it in the default browser.

---

Step 11 — Brief terminal summary

✓ Claude Radar report ready
  Project: <project> (<profileLabel>)
  Overall: <overallGrade> · <overallScore>/100
  Communication: <c1> · Engineering: <c2> · Outcome: <c3>
  Confidence: <confidenceLevel>
  File: ~/.claude-radar/reports/<filename>.html

<one-line takeaway derived from diagnosis.coreDiagnosis>

Do NOT dump full dimension breakdowns, full diagnosis, or full suggestions in the terminal — that's what the HTML report is for.

---

Analysis Principles

1. Position determines meaning — read each dimension's signals from its designated position bucket.
2. Formula is the anchor, adjustment is the tuning — baseline ensures reproducibility, adjustment ensures sensitivity. No evidence → no adjustment.
3. N/A is honest — when a dimension genuinely doesn't apply (one-shot project, learning profile, unresolved cwd), say so. Don't fake a 50.
4. Diagnosis is the gift — scores tell the user what; diagnosis tells them why and what to do. Spend the most thinking here.
5. Bilingual parity — en/zh same meaning, not literal translation.
6. Evidence beats opinion — every claim in reasoning / evidence / coreDiagnosis must trace back to specific facts.

---

Error Recovery

Parser script fails (non-zero exit, invalid JSON):

• "Couldn't parse this project. The JSONL files may be corrupted. Try another project."
• Do not continue.

Insufficient data (confidenceLevel: "low" AND stats.humanMessages < 5):

• Tell the user: "This project has only X messages — too little to evaluate meaningfully. Pick another project."
• Don't produce a report.

Insufficient data but workable (confidenceLevel: "low" AND humanMessages >= 5):

• Produce the report but ensure profile.confidence: "low" is reflected in the report. The diagnosis should explicitly mention the sample size limitation.

Render script fails:

• Show the user the report.json path so they can open it manually.

Write tool errors "directory not found":

• Run mkdir -p ~/.claude-radar/temp via Bash, then retry.

Browser doesn't auto-open:

• Tell the user to open the printed file path manually.

User cancels mid-way:

• "OK, run /claude-radar anytime to try again."

---

Path Reference

• ${CLAUDE_SKILL_DIR} = <plugin-root>/skills/analyze/
• Scripts: ${CLAUDE_SKILL_DIR}/scripts/
• Rubric: ${CLAUDE_SKILL_DIR}/../../data/rubric.json
• Report HTML: ~/.claude-radar/reports/
• Temp JSON: ~/.claude-radar/temp/