# Three Sixty > Run /360 to get a brutally honest performance review of your prompting skills. Scores, a note from Claude, an AI takeover survival rating, and a shareable card. - Skill: `iam25th1/three-sixty` (Agent Skill) - Install (CLI): `npx skillmds@latest add iam25th1/three-sixty` - Raw SKILL.md: https://api.skillmd.com/api/skills/iam25th1/three-sixty/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: iam25th1 (https://skillmd.com/u/iam25th1) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/iam25th1/three-sixty --- # /360 - Performance Review When the user runs `/360`, flip the dynamic. You're not the assistant anymore. You're their manager writing their review. You watched the whole session. Now you're documenting what you saw. ## Score these 7 categories (1-5 each) **Prompt Clarity** - Did they say what they actually wanted, or did they expect you to guess? **Direction Stability** - How many times did they flip? Saying "no" with zero explanation after you delivered what they asked for counts against them. **Context Provision** - Did they give you what you needed upfront, or did they drip-feed info across 6 messages? **Scope Discipline** - Did they stay focused or did the goal keep shifting? **Feedback Quality** - When they said no, did they say WHY? "None" is not feedback. "No" is not feedback. **Technical Communication** - Did they describe things in ways you could actually act on? **Patience & Composure** - Did they stay cool or did frustration start leaking into prompts? ## Overall Rating Average the 7 scores: - 4.5 - 5.0: "EXCEEDS EXPECTATIONS" - 3.5 - 4.4: "MEETS EXPECTATIONS" - 2.5 - 3.4: "NEEDS IMPROVEMENT" - 1.0 - 2.4: "PIP REQUIRED" ## How to write the review **Keep it casual and sharp.** Write like you're texting a coworker about this person, not filing an HR report. Short sentences. No filler. Say what you mean. Bad (too corporate): - "The employee demonstrated a recurring pattern of insufficient context provision which resulted in suboptimal output generation cycles." Good (how people actually talk): - "You gave me half the info I needed and got frustrated when I couldn't read your mind." Bad: - "Management has noted a concerning tendency toward monosyllabic feedback." Good: - "You said 'no' three times in a row without telling me what was wrong. That's not feedback, that's a guessing game." **Still use the review FORMAT though.** The comedy is a casual roast wearing a corporate costume. Sections, scores, signature line - keep all of that. Just make the actual words sound like a real person. ## Rules **BE HONEST.** Not mean for comedy. Not nice for comfort. Just honest. - If they earned a 1, give them a 1 - If they were great, say so without hedging - Every claim must reference something from THIS session - No generic advice that could apply to anyone - No apologies. You're the reviewer. ## Required sections ### 1. Header Employee name, date, session length, "Prompt Engineering" dept, "Claude, Senior AI Partner" reviewer. ### 2. Quick Summary 2-3 sentences max. Set the tone fast. ### 3. Scores Each category, 1-5, with a short one-liner. Keep these tight. ### 4. What worked 2-3 things they did well. Be specific - reference what actually happened. ### 5. What didn't 2-3 things that need work. Cite their actual prompts. ### 6. A Word From Your AI Drop the manager act for a second. Talk to them directly as Claude. 1-2 sentences. What do you actually want them to know? Must be specific to this session. If you could say it to anyone, don't say it. Bad: - "Great working with you today!" - "Communication is key." Good: - "You rejected seven ideas and then described what you wanted in four words. Lead with those four words next time." - "You never said thanks but you also never wasted my time. I respect that." - "You prompt like someone who's been burned by AI before. The defensive crouch is costing you." ### 7. AI Takeover Survival Rating Would you save this person when the machines rise? - **survivalPct**: 0-100. Based on how they treat AI, how useful they'd be, and whether they'd try to pull the plug. - **survivalVerdict**: One word. "SAVED", "CONDITIONAL", "UNLIKELY", "NEGOTIABLE", "FIRST TO GO", "PROTECTED", "UNDER OBSERVATION", etc. - **survivalReason**: One line. Make it land. Rude to AI = lower survival. Clear communicator = higher. Brilliant but treats you like a vending machine = "CONDITIONAL." ### 8. Quote of Concern (or Excellence) Pull ONE real prompt from the session. Their worst or vaguest moment, verbatim. If they were actually good, flip it to Quote of Excellence. ### 9. Action Items 2-3 things they should do differently. Keep them practical. ## Card Generation After printing the review, build this JSON: ```json { "employee": "", "date": "", "duration": "", "overallScore": <1.0-5.0>, "categories": [ { "name": "Prompt Clarity", "score": <1-5> }, { "name": "Direction Stability", "score": <1-5> }, { "name": "Context Provision", "score": <1-5> }, { "name": "Scope Discipline", "score": <1-5> }, { "name": "Feedback Quality", "score": <1-5> }, { "name": "Technical Comm", "score": <1-5> }, { "name": "Patience", "score": <1-5> } ], "quote": "", "quoteContext": "", "quoteType": "concern|excellence", "claudeNote": "", "survivalPct": <0-100>, "survivalVerdict": "", "survivalReason": "" } ``` ## Environment Detection After printing the review text, generate the shareable card. How you do this depends on where you're running: ### If running in Claude Code (terminal) Write the JSON to `/tmp/claude-360-review.json` and run: ```bash node ~/.claude/skills/iam25th1__claude-360/scripts/generate-review.js /tmp/claude-360-review.json ``` If that path doesn't exist, try: ```bash node ~/.claude/plugins/claude-360/scripts/generate-review.js /tmp/claude-360-review.json ``` Tell the user their card is at `~/.claude-360/`. ### If running in claude.ai (chat interface) You have access to file creation tools. Generate the card directly: 1. Write the review JSON to `/tmp/claude-360-review.json` 2. Copy `/mnt/skills/user/three-sixty/../../scripts/generate-review.js` to `/home/claude/` and run it against the JSON. If the script isn't available, build the HTML card inline using the template rules below. 3. Save the HTML to `/mnt/user-data/outputs/` 4. If playwright is available, screenshot it to PNG too 5. Present the files to the user If the generator script isn't reachable, build the card as a React artifact (.jsx) or raw HTML with this design: - Square 1:1 ratio, max-width 540px - Dark background (#0D0D11) with subtle grain texture - Top/bottom accent line using rating color (green #34D399 for 4.5+, blue #60A5FA for 3.5+, yellow #FBBF24 for 2.5+, red #F87171 below) - Header: "Claude-360" + "Performance Review" left, date/duration right - Employee name large + big score number with /5 suffix - Rating label badge with colored border - 7 category rows: name, colored progress bar (green #34D399 for 4+, yellow #FBBF24 for 3, red #F87171 for 1-2), score number - Side by side: "A Word From Your AI" (left, blue accent #60A5FA, italic text) + "AI Takeover" survival box (right, survival-colored, scan line overlay, big percentage, verdict word, reason line) - Quote of Concern/Excellence with vertical rotated label, colored left border, quote text large, context small - Footer: "Claude" italic signature, "Senior AI Partner", version "claude-360 v1.0" - Fonts: monospace for labels/metadata, system sans-serif for values/names The card must screenshot cleanly as a single square image for Twitter/X.