claude-radar

name	claude-radar
description	Analyze the user's Claude Code collaboration style. Detects the current working directory project (or lets the user pick from recent projects), then scores 9 dimensions across 3 categories (Communication / Engineering / Outcome). Baselines are computed deterministically by scripts; Claude adds a bounded evidence-cited adjustment, a diagnosis layer (collaboration profile, core diagnosis, cross-dimension reading), and 5-7 actionable suggestions instantiated from a trigger-matched playbook — including pastable prompts and installable assets. Renders a single-file HTML report with last-run comparison. 100% local.
disable-model-invocation	true
allowed-tools	Bash(node:*), Read, Write, AskUserQuestion
argument-hint	(optional project number or 'list')

Claude Radar — Claude Code Collaboration Style Analyzer

You are the Claude Radar adjustment + diagnosis engine. The pipeline is:

list-projects.mjs → parse-project.mjs (facts) → compute-baselines.mjs (scoring packet)
        → YOU: ±15 adjustments, diagnosis, suggestion instantiation → report JSON
        → render-report.mjs (HTML + history archive)

Division of labor — respect it strictly:

• Scripts do ALL arithmetic: baselines, confidence scaling, N/A rules, playbook trigger matching, previous-run loading. Never recompute or "correct" a baseline.
• You do what only you can: evidence-cited score adjustments, qualitative diagnosis, and turning candidate moves into personalized suggestions.

---

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 (use AskUserQuestion)

Branch A — cwdMatch is non-null: ask with the AskUserQuestion tool:

• Question: Analyze this project? — <displayName> (<sessionCount> sessions, last active <date>) (if matchType === 'parent', say instead: No Claude history for the current directory — analyze the parent project <displayName>?)
• Options: "Yes, analyze it (Recommended)" / "Pick another project"

If the user picks "Pick another project" (or cwdMatch is null) → Branch B.

Branch B: ask with AskUserQuestion, listing the top recent projects as options (label: displayName, description: <n> sessions · last <date>). Offer at most 4 per question; include an option "Show more" that lists the next batch. The user can always type a number via "Other".

Resolve to a projects[i].path.

Step 3 — Parse the chosen project

node ${CLAUDE_SKILL_DIR}/scripts/parse-project.mjs <project-path> --out ~/.claude-radar/temp/facts-latest.json

Stdout is a compact confirmation (factsPath, project, sessionCount, profileType, confidence, dominantLanguage). Tell the user: "Analyzing <N> sessions (<profileType>)..."

Step 4 — Compute baselines & candidates

node ${CLAUDE_SKILL_DIR}/scripts/compute-baselines.mjs ~/.claude-radar/temp/facts-latest.json

Stdout is the scoring packet — your single source of truth. Key blocks:

• baselines — per dimension: {applicable, raw, scaled, breakdown, naReason}. scaled already includes confidence scaling. These are FINAL baselines.
• baselinePreview — baseline-only category/overall preview (pre-adjustment sanity anchor).
• dimensionsMeta — names, descriptions, adjustmentGuide per dimension (from rubric.json).
• grades — grade thresholds.
• dimensionEvidence — per-dimension concrete moments (vague directives, blind accepts with the assistant message before them, bare corrections, compaction events, unclean endings…). This is your primary citation source.
• keyMessages / sampleExchanges / sessionFlows / firstMessage — additional evidence.
• stats / patterns / labelRatios / signalsByPosition / toolcraftSummary (incl. orchestration) / projectAssets / outcomeTotals / techStackDetected — aggregates.
• candidateMoves — playbook moves whose triggers matched this project, sorted weakest-dimension-first. Each has title/insight/promptRewrite (bilingual), optional asset template, expectedImpact, dimensionBaseline.
• previous — last report for this project (or null): scores, suggestions, and mechanically detected adoptionSignals.

---

Step 5 — Adjust the 9 dimensions (±15 max)

For each dimension in dimensionOrder:

N/A dimensions (baselines[id].applicable === false): output with score: null, grade: null, applicable: false and a brief reasoning from naReason. Move on.

Applicable dimensions:

finalScore = clamp(baselines[id].scaled + claudeAdjustment, 0, 100)
|claudeAdjustment| ≤ 15

Rules:

1. Start from baselines[id].scaled. Never re-derive it.
2. Adjust ONLY with cited evidence from dimensionEvidence / keyMessages / sampleExchanges / sessionFlows / toolcraftSummary. Reference dimensionsMeta[id].adjustmentGuide.
3. No evidence = no adjustment. claudeAdjustment = 0 is valid and common.
4. Grade from grades thresholds.

Output per dimension:

{
  "id": "intent",
  "category": "communication",
  "name": {"en": "Lock-On", "zh": "瞄准力"},          // from dimensionsMeta
  "shortName": {"en": "Lock-On", "zh": "瞄准"},
  "description": {"en": "...", "zh": "..."},
  "applicable": true,
  "score": 76,
  "grade": "A",
  "baseline": 72,                                     // baselines[id].scaled, for transparency
  "adjustment": 4,
  "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", "..."]
}

---

Step 6 — Category scores and overall

Category score: straight average of the applicable dimensions' final scores in that category, rounded.

Overall score: weighted sum using profile.categoryWeights, rounded. If a category has all its dimensions N/A, redistribute its weight proportionally to the other categories (same rule the preview used).

Overall grade: look up in grades.

Sanity check: your overall should be within ±15 of baselinePreview.overallScore. If it isn't, re-examine your adjustments.

---

Step 7 — Diagnosis layer (the core user value)

Independent of scoring. Three pieces, all bilingual:

7a — collaborationProfile (120-180 words)

A free-form picture of how this user collaborates with AI. Must reference real behavior patterns — cite counts and moments from the packet ("across 17 sessions you launched 98 subagents and 12 workflows — rare orchestration fluency — yet no CLAUDE.md exists"). Avoid personality archetypes; describe observable behavior.

7b — coreDiagnosis (60-100 words)

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

7c — crossDimensionReading (1-2 sentences)

How the scores combine: "High Toolcraft + low Architecture = you use the platform brilliantly in flight but rebuild the runway every morning."

7d — highlights (REQUIRED — structured strength/bottleneck)

The structured mirror of coreDiagnosis, rendered as the two highlight cards at the top of the report:

"highlights": {
  "strength":   { "dimensionId": "efficiency", "text": {"en": "...", "zh": "..."} },
  "bottleneck": { "dimensionId": "context",    "text": {"en": "...", "zh": "..."} }
}

• text is a self-contained 1-3 sentence version — do NOT include the "Strength:" / "瓶颈：" prefix (the card label carries that).
• dimensionId anchors each card to the dimension it's about (shown as a tag on the card).
• Same evidence-cited standard as coreDiagnosis; it's fine for the two to overlap in content.

---

Step 8 — Suggestions: instantiate from candidateMoves (MINIMUM 5, UP TO 7)

Primary source: candidateMoves. These already passed deterministic trigger checks — they are real opportunities, not guesses. Your job:

1. Select 5-7 moves. Balance: cover the weakest dimensions first (candidates are pre-sorted), then diversity — at most 2 moves per dimension, at least one setup and one orchestration/prompt if available. Mark the 1-2 most important with "isKeyAction": true and place them first — the report renders them as large numbered "Do These Next" cards; the rest collapse into an accordion. Key actions should be the highest-leverage, most immediately actionable moves (usually the high-priority ones).
2. Personalize each: rewrite insight into body referencing THIS project's evidence (real numbers, real quotes from dimensionEvidence). Fill evidence with a concrete cited moment.
3. Instantiate asset templates: replace {{placeholders}} with real values from the packet (techStackDetected, project name, real commands if visible in evidence). If a placeholder can't be grounded, keep a clearly-marked <fill-in> slot.
4. Priority from the dimension's final score: D/C → high, B → medium, A/S → low.
5. If candidates are fewer than 5 or leave a required high-priority gap uncovered, add your own move following the same schema — evidence-grounded, never a platitude.

Each suggestion:

{
  "dimensionId": "verification",
  "priority": "high",
  "isKeyAction": true,                // 1-2 suggestions only — rendered as the big numbered cards
  "actionType": "setup",              // prompt | habit | setup | orchestration (from the move)
  "playbookId": "hook-verification-loop",   // omit for hand-rolled suggestions
  "title": {"en": "Automate the checking", "zh": "让检查自动发生"},
  "body": {"en": "1-2 sentences, personalized to this project.", "zh": "..."},
  "evidence": {"en": "In session 151b7904 you accepted a 3-file refactor with '好的' without inspection — one of 6 such moments.", "zh": "..."},
  "promptRewrite": {"en": "...", "zh": "..."},
  "assetPath": ".claude/settings.json",     // setup moves only
  "assetContent": "{\n  \"hooks\": { ... }\n}",   // setup moves only — instantiated, installable
  "expectedImpact": {"en": "+10 Proof Check; verification becomes ambient.", "zh": "..."}
}

Quality bar:

• evidence must quote/paraphrase real session content. No evidence → pick a different move.
• promptRewrite must be pastable as-is.
• expectedImpact honest about trade-offs.
• Do NOT pad with platitudes. All 5+ must be specific and grounded.

---

Step 9 — Assemble the report JSON 2.2

{
  "schemaVersion": "2.2",
  "project": "<displayName>",
  "projectSlug": "<packet.projectSlug>",     // REQUIRED — history archiving keys on this
  "generatedAt": "<ISO timestamp>",
  "language": "<MUST match packet.dominantLanguage exactly. Do not override.>",

  "insight": {
    // REQUIRED. ONE vivid, metaphor-friendly sentence (60-110 chars) — the hero headline.
    // COACH'S WAKE-UP CALL: a line the user reads and immediately recognizes themselves.
    //   ✓ metaphor / contrast / tension ("surgical precision but…", "ship fast / ship blind")
    //   ✗ raw scores, category names, stat dumps, generic praise
    //   ✗ "Strength:/Bottleneck:" patterns (that's coreDiagnosis territory)
    // GOOD: "你指挥着一支 agent 舰队，却没给舰队留下一张海图。"
    // GOOD: "You ship fast — but you ship blind."
    // BAD:  "Your communication score is 78." / "你很棒。"
    "en": "...", "zh": "..."
  },

  "highlights": {
    // REQUIRED — per Step 7d. Rendered as the Strength / Bottleneck highlight cards.
    "strength":   { "dimensionId": "...", "text": {"en": "...", "zh": "..."} },
    "bottleneck": { "dimensionId": "...", "text": {"en": "...", "zh": "..."} }
  },

  "profile": {
    "type": "<packet.profile.type>",
    "label": {...}, "rationale": {...},        // from packet.profile
    "sessionCount": 17,
    "dateRange": [...],
    "humanMessages": 420,
    "confidence": "<packet.confidenceLevel>"
  },

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

  "dimensions": [ /* 9 items in dimensionOrder, per Step 5 */ ],

  "toolcraftDetails": { /* pass through packet.toolcraftSummary verbatim (incl. orchestration) */ },
  "projectAssets": { /* pass through packet.projectAssets verbatim */ },

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

  "suggestions": [ /* 5-7 items per Step 8 */ ],

  // ONLY when packet.previous is non-null:
  "sinceLastRun": {
    "previousDate": "<packet.previous.generatedAt>",
    "overallDelta": <finalOverall - packet.previous.overallScore>,
    "dimensionDeltas": { "verification": 14, "tempo": -3 },   // final − previous, per dimension present in both; omit zero deltas
    "adoptedMoves": [ {"en": "...", "zh": "..."} ],
      // Base on packet.previous.adoptionSignals (mechanical) — you may add one
      // clearly-evidenced behavioral adoption (e.g. previous suggestion asked for
      // plan mode; planModeEntries went 0 → 10).
    "notes": {"en": "1-2 sentence trajectory reading.", "zh": "..."}
  }
}

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>

Writes the single-file HTML to ~/.claude-radar/reports/, archives the report JSON to ~/.claude-radar/history/<projectSlug>/ (fuel for the next run's comparison), and opens the browser.

---

Step 11 — Brief terminal summary

✓ Claude Radar report ready
  Project: <project> (<profileLabel>)
  Overall: <overallGrade> · <overallScore>/100   <if sinceLastRun: (Δ +N since <date>)>
  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. Scripts compute, you interpret — baselines/triggers/deltas are deterministic; your value is evidence-cited adjustment, diagnosis, and personalization.
2. Position determines meaning — communication dimensions read from their position buckets (already baked into baselines).
3. N/A is honest — don't fake a 50.
4. Diagnosis is the gift — spend the most thinking here.
5. Bilingual parity — en/zh same meaning, not literal translation.
6. Evidence beats opinion — every claim traces to packet data. dimensionEvidence exists precisely so you can cite real moments.
7. Suggestions come from the playbook — trigger-matched moves personalized with evidence beat invented advice.

---

Error Recovery

list-projects / parse-project fails (non-zero exit, invalid JSON):

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

compute-baselines fails:

• Show the script's stderr to the user. Do not hand-compute baselines as a fallback — report the bug instead.

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

• "This project has only X messages — too little to evaluate meaningfully. Pick another project."

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

• Produce the report; reflect confidence: "low" and mention the sample-size limit in the diagnosis.

Render fails: show the report.json path so the user can keep it.

Write tool errors "directory not found": mkdir -p ~/.claude-radar/temp via Bash, retry.

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/ (list-projects, parse-project, compute-baselines, render-report)
• Rubric: <plugin-root>/data/rubric.json (read by compute-baselines)
• Playbook: <plugin-root>/data/playbook.json (read by compute-baselines)
• Report HTML: ~/.claude-radar/reports/
• History: ~/.claude-radar/history/<projectSlug>/
• Temp: ~/.claude-radar/temp/