diff --git a/.dotter/global.toml b/.dotter/global.toml
index 7e16296..d7be65d 100644
--- a/.dotter/global.toml
+++ b/.dotter/global.toml
@@ -73,6 +73,7 @@ depends = []
 
 [bin.files]
 "bin/discord-send" = "~/.local/bin/discord-send"
+"bin/claude-token-count" = "~/.local/bin/claude-token-count"
 
 [bin.variables]
 
diff --git a/bin/claude-token-count b/bin/claude-token-count
new file mode 100755
index 0000000..151b76e
--- /dev/null
+++ b/bin/claude-token-count
@@ -0,0 +1,95 @@
+#!/usr/bin/env bash
+#
+# claude-token-count — count tokens for text using Anthropic's
+# /v1/messages/count_tokens endpoint (the exact Claude tokenizer).
+#
+# The API key is pulled from 1Password at runtime via the `op` CLI; it is
+# never written to disk and is passed to curl via a config file on /dev/fd
+# so it stays out of the process argument list.
+#
+# Usage:
+#   claude-token-count "some text"          # count the argument(s)
+#   claude-token-count --file notes.md      # count a file's contents
+#   some-command | claude-token-count       # count piped stdin
+#
+# Options:
+#   --file PATH     read the text to count from PATH
+#   --model NAME    model whose tokenizer to use (default: $CLAUDE_TOKEN_COUNT_MODEL
+#                   or claude-sonnet-4-6)
+#   --raw           print the raw JSON response instead of just the number
+#   -h, --help      show this help
+#
+# Requires: op (1Password CLI, signed in), curl, jq.
+
+set -euo pipefail
+
+# 1Password reference for the Anthropic API key (from the item's share link).
+OP_ACCOUNT="EADC56ZW7NARRJTNZKSRBYRQNU"
+OP_REF="op://m6vn2ie66x7ysrjq73q45oiyeu/te6tuanr426ohu33t2qhqzjrb4/credential"
+
+MODEL="${CLAUDE_TOKEN_COUNT_MODEL:-claude-sonnet-4-6}"
+RAW=false
+FILE=""
+declare -a TEXT_PARTS=()
+
+usage() { sed -n '2,/^set -euo/{/^set -euo/!p}' "$0" | sed 's/^# \{0,1\}//'; }
+
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --file) FILE="${2:?--file requires a path}"; shift 2 ;;
+    --file=*) FILE="${1#--file=}"; shift ;;
+    --model) MODEL="${2:?--model requires a name}"; shift 2 ;;
+    --model=*) MODEL="${1#--model=}"; shift ;;
+    --raw) RAW=true; shift ;;
+    -h|--help) usage; exit 0 ;;
+    *) TEXT_PARTS+=("$1"); shift ;;
+  esac
+done
+
+for tool in op curl jq; do
+  command -v "$tool" >/dev/null 2>&1 || { echo "Error: '$tool' is required but not installed." >&2; exit 1; }
+done
+
+# Resolve the text to count: --file wins, then positional args, then stdin.
+if [[ -n "$FILE" ]]; then
+  [[ -r "$FILE" ]] || { echo "Error: cannot read file: $FILE" >&2; exit 1; }
+  TEXT="$(cat -- "$FILE")"
+elif [[ ${#TEXT_PARTS[@]} -gt 0 ]]; then
+  TEXT="${TEXT_PARTS[*]}"
+elif [[ ! -t 0 ]]; then
+  TEXT="$(cat)"
+else
+  echo "Error: no input. Provide text as an argument, --file, or via stdin." >&2
+  exit 2
+fi
+
+[[ -n "$TEXT" ]] || { echo "Error: input is empty." >&2; exit 2; }
+
+# Build the request body, encoding the text safely as a JSON string.
+BODY="$(jq -n --arg model "$MODEL" --arg text "$TEXT" \
+  '{model: $model, messages: [{role: "user", content: $text}]}')"
+
+# Fetch the key into a variable (local to this process) and hand it to curl
+# through a --config file via process substitution, keeping it out of argv.
+API_KEY="$(op read --account "$OP_ACCOUNT" "$OP_REF")"
+
+RESPONSE="$(curl -sS https://api.anthropic.com/v1/messages/count_tokens \
+  --config <(printf 'header = "x-api-key: %s"\n' "$API_KEY") \
+  -H "anthropic-version: 2023-06-01" \
+  -H "content-type: application/json" \
+  -d "$BODY")"
+unset API_KEY
+
+if $RAW; then
+  echo "$RESPONSE"
+  exit 0
+fi
+
+# Surface API errors instead of printing "null".
+if echo "$RESPONSE" | jq -e 'has("error")' >/dev/null 2>&1; then
+  echo "Error from Anthropic API:" >&2
+  echo "$RESPONSE" | jq -r '.error.type + ": " + .error.message' >&2
+  exit 1
+fi
+
+echo "$RESPONSE" | jq -r '.input_tokens'
diff --git a/claude/skills/cc-compact/SKILL.md b/claude/skills/cc-compact/SKILL.md
new file mode 100644
index 0000000..9cd2872
--- /dev/null
+++ b/claude/skills/cc-compact/SKILL.md
@@ -0,0 +1,64 @@
+---
+name: cc-compact
+description: Summarize ("compact") a past Claude Code session without ingesting its entire history. Resolves the session log, then extracts only the key signals — opening intent, recent prompts, most-edited files, and what the agent was doing last. Use, when asked, to resume long sessions without ingesting entire log.
+---
+
+You are reloading the context of a previous Claude Code session so you can
+**continue that work in this session** — like `/resume` followed by `/compact`, but pre-compacted. The
+bundled helper script extracts the key signals and prints a compact report.
+
+**Hard rule: you may NOT read the session `.jsonl` file by any means** — no
+`cat`, `head`, `tail`, `jq`, `grep`, `Read`, nothing. Run the helper script
+**once** and work solely from its output. That is the only thing allowed to
+touch the file.
+
+The helper script lives at `~/.claude/skills/cc-compact/compact_session.py`.
+
+## Step 1: Resolve the session
+
+The argument is one of two forms:
+
+- **A session id** — when invoked like `/resume claude --resume <id>` or
+  `cc-compact <id>`, the user already gave you the UUID. Pass it as `--id`.
+- **A session name / title** — free text. Pass it as `--title`; the script
+  matches it against the `ai-title` records inside the logs (case-insensitive
+  substring, newest match wins).
+
+Run the script exactly once with the matching form:
+
+```sh
+# By id:
+python3 ~/.claude/skills/cc-compact/compact_session.py --id <session-uuid>
+
+# By title:
+python3 ~/.claude/skills/cc-compact/compact_session.py --title "the session name"
+```
+
+## Step 2: Read the output
+
+It prints a bounded, XML-tagged report:
+
+- header: project, git branch, time span, record/prompt counts
+- the first few exchanges (user prompt + the agent's reply) — the original intent
+- a few exchanges sampled from the middle (non-overlapping) — the journey
+- the last few exchanges (user prompt + the agent's reply) — where things were heading
+- the most-edited files, ranked by edit count
+- the last several tool calls — what the agent was physically doing last
+- the final assistant message — what it was saying / waiting on last
+
+## Step 3: Pick up the work
+
+The output is context for *you* — treat it like a resumed session, not something
+to report on. From it, reconstruct what the work was, which files are in play,
+and what the agent was in the middle of.
+
+Compaction is lossy — the output may not make the next step unambiguous. Before
+diving in, decide whether you actually know how to proceed:
+
+- **If the next step is very clear, last few messages from user confirms what to do**, open with one line confirming what you're
+  resuming, then continue from where it left off.
+- **If it's ambiguous** (unclear what the user wants next, multiple plausible
+  directions, or the session ended mid-decision), **ask the user how to proceed**
+  before acting. Surface the few plausible next steps you inferred and let them
+  pick or correct you. If there is any doubt, opt to ask the user first.
+
diff --git a/claude/skills/cc-compact/compact_session.py b/claude/skills/cc-compact/compact_session.py
new file mode 100755
index 0000000..f5d6482
--- /dev/null
+++ b/claude/skills/cc-compact/compact_session.py
@@ -0,0 +1,262 @@
+#!/usr/bin/env python3
+"""Compact a Claude Code session log into a small, bounded summary.
+
+This reads a session JSONL file *carefully* — it never dumps the whole
+history. It extracts only the few signals needed to understand what the
+session was about and what the agent was doing at the end:
+
+  - header metadata (project, branch, time span, message counts)
+  - the first few exchanges (user prompt + the agent's reply, truncated)
+  - a few exchanges randomly sampled from the middle (non-overlapping)
+  - the last few exchanges (user prompt + the agent's reply, truncated)
+  - the most-edited files (top N by edit count)
+  - the final assistant text (what it was saying last)
+  - the last few tool calls (what it was doing last)
+
+Session resolution (pick one):
+  --file PATH        use this JSONL file directly
+  --id UUID          find <UUID>.jsonl under the projects dir
+  --title TEXT       find the session whose ai-title contains TEXT
+                     (case-insensitive substring; newest match wins)
+
+Sessions live at: ~/.claude/projects/<encoded-cwd>/<session-id>.jsonl
+"""
+
+import argparse
+import glob
+import json
+import os
+import random
+import sys
+
+PROJECTS_DIR = os.path.expanduser("~/.claude/projects")
+
+# Tools that change files on disk, and which input field holds the path.
+EDIT_TOOLS = {
+    "Edit": "file_path",
+    "Write": "file_path",
+    "MultiEdit": "file_path",
+    "NotebookEdit": "notebook_path",
+}
+
+
+def truncate(text, n):
+    """Collapse whitespace to a single line, then clip to n chars.
+    Use for compact one-liners like tool arguments."""
+    text = " ".join(text.split())
+    return text if len(text) <= n else text[: n - 1] + "…"
+
+
+def clip(text, n):
+    """Clip to n chars while preserving newlines/indentation.
+    Use for quoted messages where structure matters."""
+    text = text.strip("\n")
+    return text if len(text) <= n else text[:n].rstrip() + " …[truncated]"
+
+
+def esc(text):
+    """Escape XML metacharacters for tag values and attributes."""
+    return str(text).replace("&", "&amp;").replace("<", "&lt;").replace(">", "&gt;")
+
+
+def content_to_text(content):
+    """A message's .content is either a string or a list of typed blocks."""
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        parts = [b.get("text", "") for b in content if isinstance(b, dict) and b.get("type") == "text"]
+        return "\n".join(p for p in parts if p)
+    return ""
+
+
+def is_genuine_prompt(rec):
+    """A real human-typed prompt: a user message that isn't a tool result,
+    meta record, slash-command wrapper, or interrupt marker."""
+    if rec.get("type") != "user" or rec.get("isMeta"):
+        return False
+    content = rec.get("message", {}).get("content")
+    if not isinstance(content, str):
+        return False
+    s = content.lstrip()
+    if not s:
+        return False
+    skip_prefixes = ("<", "[Request interrupted", "Caveat:")
+    return not s.startswith(skip_prefixes)
+
+
+def resolve_file(args):
+    if args.file:
+        return os.path.expanduser(args.file)
+
+    if args.id:
+        matches = glob.glob(os.path.join(PROJECTS_DIR, "**", f"{args.id}.jsonl"), recursive=True)
+        if not matches:
+            sys.exit(f"No session file found for id {args.id} under {PROJECTS_DIR}")
+        return matches[0]
+
+    if args.title:
+        needle = args.title.lower()
+        candidates = []  # (mtime, path, title)
+        for path in glob.glob(os.path.join(PROJECTS_DIR, "**", "*.jsonl"), recursive=True):
+            title = None
+            try:
+                with open(path, encoding="utf-8") as fh:
+                    for line in fh:
+                        if '"ai-title"' not in line:
+                            continue
+                        try:
+                            rec = json.loads(line)
+                        except json.JSONDecodeError:
+                            continue
+                        if rec.get("type") == "ai-title":
+                            title = rec.get("aiTitle", "")
+                            if needle in title.lower():
+                                break
+                            title = None
+            except OSError:
+                continue
+            if title is not None:
+                candidates.append((os.path.getmtime(path), path, title))
+        if not candidates:
+            sys.exit(f"No session whose ai-title contains {args.title!r}")
+        candidates.sort(reverse=True)
+        if len(candidates) > 1:
+            sys.stderr.write("Multiple matches (using newest):\n")
+            for _, path, title in candidates:
+                sys.stderr.write(f"  {path}  —  {title}\n")
+        return candidates[0][1]
+
+    sys.exit("Provide one of --file, --id, or --title")
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    g = ap.add_mutually_exclusive_group(required=True)
+    g.add_argument("--file", help="path to a session .jsonl")
+    g.add_argument("--id", help="session UUID")
+    g.add_argument("--title", help="substring of the session's ai-title")
+    ap.add_argument("--first", type=int, default=5, help="how many opening exchanges")
+    ap.add_argument("--last", type=int, default=10, help="how many closing exchanges")
+    ap.add_argument("--middle", type=int, default=5, help="how many exchanges randomly sampled from the middle")
+    ap.add_argument("--top-files", type=int, default=10, help="how many most-edited files")
+    ap.add_argument("--maxlen", type=int, default=800, help="max chars per quoted message")
+    ap.add_argument("--seed", type=int, default=1, help="seed for the middle random sample")
+    args = ap.parse_args()
+
+    path = resolve_file(args)
+
+    turns = []            # conversational turns: {"prompt": str, "reply": [text,...]}
+    edit_counts = {}      # file path -> edit count
+    last_assistant_text = []  # final text blocks
+    recent_tools = []     # (name, short arg)
+    counts = {}           # record type -> count
+    first_ts = last_ts = None
+    cwd = branch = None
+
+    with open(path, encoding="utf-8") as fh:
+        for line in fh:
+            line = line.strip()
+            if not line:
+                continue
+            try:
+                rec = json.loads(line)
+            except json.JSONDecodeError:
+                continue
+
+            rtype = rec.get("type")
+            counts[rtype] = counts.get(rtype, 0) + 1
+
+            ts = rec.get("timestamp")
+            if ts:
+                first_ts = first_ts or ts
+                last_ts = ts
+            cwd = cwd or rec.get("cwd")
+            branch = branch or rec.get("gitBranch")
+
+            if is_genuine_prompt(rec):
+                turns.append({"prompt": rec["message"]["content"], "reply": []})
+
+            if rtype == "assistant":
+                blocks = rec.get("message", {}).get("content", [])
+                if isinstance(blocks, list):
+                    text_here = []
+                    for b in blocks:
+                        if not isinstance(b, dict):
+                            continue
+                        if b.get("type") == "text" and b.get("text", "").strip():
+                            text_here.append(b["text"])
+                        elif b.get("type") == "tool_use":
+                            name = b.get("name", "?")
+                            inp = b.get("input", {}) or {}
+                            if name in EDIT_TOOLS:
+                                fp = inp.get(EDIT_TOOLS[name])
+                                if fp:
+                                    edit_counts[fp] = edit_counts.get(fp, 0) + 1
+                            arg = inp.get("file_path") or inp.get("command") or inp.get("description") or inp.get("path") or ""
+                            recent_tools.append((name, truncate(str(arg), 300)))
+                    if text_here:
+                        last_assistant_text = text_here  # keep only the latest turn's text
+                        if turns:  # attach the agent's reply to the current turn
+                            turns[-1]["reply"].extend(text_here)
+
+    # Non-overlapping index sets: first wins, then last, then the middle is
+    # sampled only from what's left between them. If the regions collide
+    # (short session), the overlap is simply dropped.
+    n = len(turns)
+    first_idx = list(range(min(args.first, n)))
+    last_start = max(len(first_idx), n - args.last)
+    last_idx = list(range(last_start, n))
+    mid_pool = list(range(len(first_idx), last_start))
+    if args.seed is not None:
+        random.seed(args.seed)
+    mid_idx = sorted(random.sample(mid_pool, min(args.middle, len(mid_pool))))
+
+    # ---- emit a bounded, XML-tagged report (newlines preserved) ----
+    out = sys.stdout.write
+
+    out("<session-summary>\n")
+    out(f"  <file>{esc(path)}</file>\n")
+    out(f"  <project>{esc(cwd)}</project>\n")
+    out(f"  <git-branch>{esc(branch)}</git-branch>\n")
+    out(f'  <time-span start="{esc(first_ts)}" end="{esc(last_ts)}" />\n')
+    out(f"  <records>{esc(', '.join(f'{k}={v}' for k, v in sorted(counts.items())))}</records>\n")
+    out(f"  <user-prompts count=\"{len(turns)}\" />\n")
+    out("</session-summary>\n")
+
+    def emit_exchanges(label, indices):
+        if not indices:
+            return
+        out(f'\n<exchanges section="{label}">\n')
+        for i in indices:
+            t = turns[i]
+            reply = "\n".join(t["reply"]).strip()
+            out(f'  <exchange n="{i + 1}">\n')
+            out(f"    <user>\n{clip(t['prompt'], args.maxlen)}\n    </user>\n")
+            if reply:
+                out(f"    <agent>\n{clip(reply, args.maxlen)}\n    </agent>\n")
+            else:
+                out("    <agent note=\"no text reply — tool calls only\" />\n")
+            out("  </exchange>\n")
+        out("</exchanges>\n")
+
+    emit_exchanges("first", first_idx)
+    emit_exchanges("sampled-middle", mid_idx)
+    emit_exchanges("last", last_idx)
+
+    out(f'\n<most-edited-files top="{args.top_files}">\n')
+    for fp, c in sorted(edit_counts.items(), key=lambda kv: -kv[1])[: args.top_files]:
+        out(f'  <file edits="{c}">{esc(fp)}</file>\n')
+    out("</most-edited-files>\n")
+
+    out("\n<last-tool-calls>\n")
+    for name, arg in recent_tools[-8:]:
+        out(f'  <call tool="{esc(name)}">{esc(arg)}</call>\n')
+    out("</last-tool-calls>\n")
+
+    out("\n<final-agent-message>\n")
+    out(clip("\n".join(last_assistant_text), 2000) + "\n")
+    out("</final-agent-message>\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/claude/skills/new-work/SKILL.md b/claude/skills/new-work/SKILL.md
new file mode 100644
index 0000000..4af3b96
--- /dev/null
+++ b/claude/skills/new-work/SKILL.md
@@ -0,0 +1,28 @@
+---
+name: new-work
+description: Prepare the repo to start a new piece of work. Use this skill when the user asks to start a new task, begin new work, create a new branch, or start working on something new.
+---
+
+You are helping the user start a new piece of work in a git repository. Follow these steps in order:
+
+## Step 1: Check for uncommitted changes
+
+Run `git status --short` to check for uncommitted changes.
+
+If there are uncommitted changes:
+- Evaluate the changes:
+  - If there are build artifacts, etc. then ignore them. If they are tracked by git, note for user that they might want to gitingore these.
+  - If changes look important, such as code changes, new code files that are not empty, changes to build workflows etc., then ask user if they want to stash
+
+## Step 2: Switch to main and pull latest
+
+- Attempt to check out `main`.
+  - If that fails because the branch does not exist, check out `master` instead.
+- Then pull the latest changes:
+
+## Step 3: Create a new branch
+
+Based on the description of the work the user has provided, create a short, descriptive, kebab-case branch name (e.g. `add-user-auth`, `fix-checkout-crash`, `refactor-api-client`).
+Create and check out the new branch.
+Once branch is created, proceed with work as usual.
+
diff --git a/setup.sh b/setup.sh
index 20e08ab..d23d7e0 100755
--- a/setup.sh
+++ b/setup.sh
@@ -259,6 +259,7 @@ echo ""
 echo "Setting up directory symlinks..."
 DOTFILES_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 ensure_dir_symlink "$DOTFILES_DIR/fish/functions" "~/.config/fish/functions" "$FORCE_DEPLOY"
+ensure_dir_symlink "$DOTFILES_DIR/claude/skills" "~/.claude/skills" "$FORCE_DEPLOY"
 
 echo ""
 echo "✓ Setup complete!"