claude-code-hidden-features

A Claude Code skill that surfaces capabilities the product under-surfaces, then triages them against what you actually use — instead of leaving them drip-fed or buried in a binary.

Two modes, one thesis: the in-product surfaces are gimped (a small side-model drips tips one at a time; --help is truncated). A powerful main agent with your full history can mine the installed binary and rank what it finds for you.

• Mode A — the tip catalog. Claude Code's "daily tips" come from a fixed catalog (41 entries in 2.1.196) shown one-at-a-time by a small classifier that declines most of the time and only picks from a pre-filtered eligible subset. It can never tell you "here are the N you're missing." This extracts the whole catalog.
• Mode B — undocumented CLI flags. The binary defines ~74 flags; claude --help documents ~60. The rest (--thinking-display, --bare, --max-budget-usd, --plan-mode-instructions, …) are real but invisible. This diffs the binary's flag block against --help.

Both modes re-extract per version (nothing is hardcoded) and the output is always a triage, not a list — the 🔥 high-leverage few first, every item carrying a one-line evidence-grounded reason, the full set demoted to a collapsed reference.

Install

Drop the directory into your Claude Code skills folder:

git clone https://github.com/evnchn-agentic/claude-code-hidden-features \
  ~/.claude/skills/claude-code-hidden-features

Then ask Claude Code something like "extract all the Claude Code tips and screen the good ones" or "what undocumented flags is --help hiding?"

What's here

File	Purpose
SKILL.md	The skill: procedure, triage contract, tier semantics, and the binary-mining gotchas.
extract_catalog.sh	Mode A — dumps the tip catalog from the installed binary (stable-string anchors, version-robust).
extract_flags.sh	Mode B — dumps the CLI flag block and the --help-omitted delta.

Run either script directly (./extract_catalog.sh / ./extract_flags.sh); both auto-locate the running version's binary, or take a path as $1.

Gotchas it bakes in (each cost a real agent time)

• System grep may be ugrep → greedy .\{0,N\} windows abort with "exceeds complexity limits"; use strings + line-range sed.
• The binary embeds HTML/CSS-parser and bun/ripgrep help corpora → phrase-greps for "tip" drown in false positives. Anchor on feature_ids / the commander block.
• Modern Claude Code is a single ~220 MB Bun-compiled binary — there's no readable cli.js.
• Don't anchor on minified symbols (they change every build) — anchor on stable English strings.
• Self-contamination: the moment you dump extracted strings into chat, they land in the session log — so grepping transcripts for "usage" returns your own output. Exclude the current session; prefer shell history for true human-typed usage.

Provenance

Built and verified (RED/GREEN, both extractors confirmed against 2.1.196) on macOS/arm64. Reverse-engineering of one's own installed CLI via strings is standard archaeology; none of this is security-sensitive, and the community already mirrors more (e.g. the full system prompt). Re-run per Claude Code version.

---

Part of evnchn-agentic — agentic skills as reversible experiments.