yc-coding-sample

A Claude Code skill that prepares a real coding session transcript for a YC application. It picks the strongest candidate from your local session history, scores it through the eyes of a YC partner, exports it as readable Markdown, and runs a security sweep before you submit.

What it does

A six-phase pipeline with a user checkpoint between every phase — the skill never auto-completes the full run.

1. Shortlist. Surveys ~/.claude/projects/ directly via find + jq (instructions live in SKILL.md so the model can adapt the filter). Ranks recent sessions by size, recency, and substance. Outputs a Markdown table of 6–10 candidates.
2. Deep dive. For 1–3 sessions you pick, an Explore subagent reads the JSONL and summarizes the problem, narrative arc, technical substance, AI-collaboration moments, outcome, and weaknesses.
3. Independent YC-investor review. A fresh, un-briefed general-purpose subagent reads the JSONL with the rubric in references/yc-rubric.md and scores 7 dimensions, gives a composite score, a YC tier verdict, the strongest and weakest signals as verbatim quotes, and two interview questions tied to specific transcript moments.
4. Comparative ranking (optional). Run Phase 3 in parallel across the rest of the shortlist for calibrated scoring.
5. Export. scripts/jsonl_to_md.py converts the chosen JSONL to Markdown — system instructions stripped, tool calls inlined, Edits as diffs, tool results in collapsibles, secret-shaped tokens redacted at render time.
6. Security scan. scripts/scan_secrets.py sweeps the export for API keys, tokens, connection strings, env values, real emails, IPs, Stripe/Clerk IDs, and PII. Reports blockers vs. minor concerns. Exits 3 on blockers, 0 otherwise.

Why

Surfacing the right Claude Code session out of dozens or hundreds of workspaces is a real problem when you're applying to YC. Picking by hand without an investor's lens means you ship a session that feels impressive but isn't the strongest signal you have. This skill packages the workflow so the next iteration is one command.

Install

Via npx skills (recommended)

npx skills add BradSmith2015/yc-coding-sample -g -y

-g installs globally to ~/.claude/skills/. -y skips confirmation prompts. The CLI clones the repo, finds the SKILL.md under skills/yc-coding-sample/, and symlinks the skill into the directories of every supported agent it detects on your machine (Claude Code, Cursor, Codex, etc.).

After install, confirm it landed:

npx skills list
# → expect to see: yc-coding-sample (~/.claude/skills/yc-coding-sample)

Then in Claude Code, run /reload-plugins and /yc-coding-sample to invoke.

Symlink (for hacking on the skill itself)

git clone https://github.com/BradSmith2015/yc-coding-sample ~/code/yc-coding-sample
ln -s ~/code/yc-coding-sample/skills/yc-coding-sample ~/.claude/skills/yc-coding-sample

Copy

cp -r ~/code/yc-coding-sample/skills/yc-coding-sample ~/.claude/skills/

Usage

Explicit invocation only — the skill will not auto-trigger on generic portfolio or interview-prep talk.

use the yc-coding-sample skill to prep a YC submission

or

/yc-coding-sample

Configuration

The phased pipeline, every flag, and the script-by-script behavior live in skills/yc-coding-sample/SKILL.md. Read it once before customizing.

To retune the YC-investor scoring (dimension weights, tier thresholds, output format), edit skills/yc-coding-sample/references/yc-rubric.md — the file is loaded verbatim into the reviewer agent's prompt.

Out of scope

• No GitHub auto-push, no auto-submit-to-YC. The skill stops at "ready to submit."
• No PDF or non-Markdown output.
• No scheduled / cron use. Manual invocation only.

License

MIT. See LICENSE.