grade-all

Grade All Submissions

Evaluate all student capstone submissions and produce a ranked comparison table.

Options

• --deep-dive — Show detailed technical deep dive per candidate before the summary table. Default is OFF (summary table only with per-candidate scorecard).

Step 1: Identify submissions

The user provides submissions in one of these formats:

Format A — Name + Git URL list

grade-all:
- "Hong Dao" https://github.com/HONG-DAO/Deep-Analyst-Research-Intelligence-Platform
- "Nguyen Thanh Anh Duy" https://github.com/ntad27/chat-app-capstone
- "Nguyen Trong Quy" https://github.com/nguyenquyking/Research-Intelligence-Platform

For each entry, clone the repo with git clone --depth=1 <url> into a temp directory or ext/[shortname].

Format B — Local paths (default)

If no URLs provided, look for submissions in ext/ relative to the capstone repo root. Ask the user which subdirectories to evaluate.

Format C — Mixed

Some entries may be local paths, others may be git URLs. Handle both.

For repos where code is on a branch other than main (e.g., develop), check git branch -a and checkout the branch with the most recent commits.

Step 2: Grade each submission

For each submission, run the full grade-submission skill workflow from ${CLAUDE_PLUGIN_ROOT}/skills/grade-submission/SKILL.md.

• If --deep-dive is set, pass it through to each grade-submission call.
• If --deep-dive is NOT set, still run the full evaluation internally but only output the scorecard summary (not the technical deep dive section).

Output each complete scorecard as you go — do not batch them all at the end.

Step 3: Compile the summary comparison table

After all scorecards are produced, output a ranked summary table with the columns defined in the rubric's "Summary Table Columns" section:

---

Capstone Cohort Ranking

Full Name	Repo	Score	SDK Used	parent_tool_use_id	ask_user	Tests	Design Doc	TypeScript	Verdict
...	owner/repo	XX/100	claude_agent_sdk ✅	✅ Correct	✅	N tests ✅	Complete ✅	✅	STRONG HIRE
...	owner/repo	XX/100	...	...	...	...	...	...	...

Sort by score descending. Use emoji indicators: ✅ (present/correct), ❌ (absent/wrong), ⚠️ (partial).

---

Step 4: Compile the dimension breakdown table

Also output a per-dimension breakdown:

Rank	Name	Total	D1 SDK	D2 Orch	D3 Decode	D4 Arch	D5 UI	D6 Code	D7 Delivery	Stretch	Verdict
1	...	X.X	X/4	X/4	X/4	X/4	X/4	X/4	X/4	+X.X	...

Step 5: Write a cohort summary

2–3 sentences on:

• What distinguishes the top submission from the rest
• The most common gap across the cohort
• Any auto-reject triggers and which rules they hit
• Which SDK each candidate used and how that affected their score