Back to search
agent-team-skills
MarketplaceWorks with Claude CodeGitHub
Skills for leading a team of Claude Code agents without paying for them to remember.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Checked today
Badge for your README
Markdown
[](/listings/coracle-us--agent-team-skills)HTML
<a href="/listings/coracle-us--agent-team-skills"><img src="/api/listings/coracle-us--agent-team-skills/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
In Claude Code:
/plugin install agent-team-skills --marketplace coracle-us/agent-team-skills
On older versions, two steps:
/plugin marketplace add coracle-us/agent-team-skills /plugin install agent-team-skills@agent-team-skills
Then check your own week: ask Claude "where did my tokens go this week?" and the `audit-your-spend` skill runs.
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 17 lines
- {
- "name": "agent-team-skills",
- "description": "Skills for leading a team of Claude Code agents without paying for them to remember.",
- "owner": {
- "name": "Jon McBee (Coracle)",
- "email": "jon@coracle.us"
- },
- "plugins": [
- {
- "name": "agent-team-skills",
- "description": "Skills for leading a team of Claude Code agents without paying for them to remember.",
- "version": "0.1.0",
- "source": "./",
- "author": { "name": "Jon McBee", "email": "jon@coracle.us" }
- }
- ]
- }
README.md · README · 97 lines
- # agent-team-skills
- **In agentic coding you pay for agents to remember, not to work.**
- In October 2026 I used 25.5 billion tokens through Claude Code, most of it leading teams of agents that built a multiplayer
- browser game in 12 days. Only **0.24%** of those tokens were the model writing anything: code, prose, commands.
- **96.4%** was the model re-reading context it already had, from the cache, on every single call.
- So I audited every transcript to find where the re-reading went, and turned what I found into seven Claude Code
- skills. Each one fires at the moment the leak happens, for whoever is **leading a team of agents**.
- ## Install
- In Claude Code:
- ```text
- /plugin install agent-team-skills --marketplace coracle-us/agent-team-skills
- ```
- On older versions, two steps:
- ```text
- /plugin marketplace add coracle-us/agent-team-skills
- /plugin install agent-team-skills@agent-team-skills
- ```
- Then check your own week: ask Claude "where did my tokens go this week?" and the `audit-your-spend` skill runs.
- ## What I found, and the skill for each
- From the 12-day build (Oct 18–30: 21.9B tokens, 292 agents plus 2 lead sessions), as `audit-your-spend` counts it. Cost shares are estimates from API
- list-price ratios.
- | Skill | Fires when | The leak it stops |
- |---|---|---|
- | `briefing-agents` | you're about to spawn an agent | Agents were **86%** of tokens. One agent kept alive for **117 hours** made 1,320 calls at ~500k context each: **12%** of all cost by itself. |
- | `waiting-for-free` | a build, CI run or deploy will take minutes | over **2,000** sleep-and-check loops, and **1,154** cache rebuilds after idle gaps: about **35%** of all cost. |
- | `context-hygiene` | a session is getting heavy, or CLAUDE.md is growing | Calls carrying over 400k context were **46%** of calls and **70%** of cost. CLAUDE.md grew from 10 KB to 255 KB, and a new agent's starting context went from **32k to 113k** tokens. |
- | `looking-cheaply` | you're about to take screenshots | More than **6,500** screenshots, each re-read on every later turn. |
- | `merging-in-batches` | you're merging a stack of PRs, or a test is flaky | One CI run for a resolved stack, not one per PR; a flaky test gets one logged retry, not a restart. |
- | `rulings-become-tests` | the lead makes a call that must keep holding | Fresh agents don't know yesterday's decisions. A test does. |
- | `audit-your-spend` | you ask where the tokens went | Runs locally on your own logs and shows all of the above for your week. |
- ### `briefing-agents`
- Agents are disposable: one job, one result, then the agent ends. The skill gives the shape of a brief (goal, where to
- look, what's ruled out, done when) and a footer to paste into every brief: stop your servers, never end a turn
- waiting on a job, no sleep loops, state goes in files.
- ### `waiting-for-free`
- Waiting should cost nothing. Let the platform wait (`gh pr merge --auto`), or run one background waiter that wakes the
- agent once. Includes the three waiter traps that bit us: `pgrep -f` matching its own wait loop, a "done" grep firing
- on the previous run's log line, and completion notices that don't wake an idle agent.
- ### `context-hygiene`
- You pay for the window you carry, not the window you have. Near 250k tokens, write a handoff file and start fresh.
- Keep CLAUDE.md an index under about 20 KB; detail goes in docs agents read only when their task touches it.
- ### `looking-cheaply`
- Capture everything, pixel-diff against the last run, view one labelled contact sheet, and zoom only where something
- changed. Or hand the looking to a short-lived reviewer that reports in words.
- ### `merging-in-batches`
- Run the slow check once on the tip of a resolved stack, not once per PR. A flaky test gets one logged retry; two
- failures stop the queue; then fix the root (pin the clock and the seed).
- ### `rulings-become-tests`
- When the lead says "note that" about a rule, write the guard test in the same turn, with the ruling quoted in its
- failure message, and one line in CLAUDE.md pointing at it.
- ### `audit-your-spend`
- A standard-library Python script that reads `~/.claude/projects` and prints counts and percentages: tokens by day and
- week, the mix, heavy calls, cache rebuilds, the costliest sessions and how long they lived, starting context by day,
- polls and images. It's read-only, makes no network calls, and never prints message text, project names or prices.
- Run it directly from a clone of this repo:
- ```bash
- python3 skills/audit-your-spend/scripts/audit.py --since 2026-11-01 --until 2026-11-07
- ```
- ## How the skills were tested
- Each behaviour skill has a scenario in `tests/scenarios/`: a fresh agent answers it once without the skill and once
- with it (`zsh tests/run-scenarios.sh <out-dir> [skills-root]`). Without the skills, a fresh agent views all 12 screenshots
- at full size, keeps working at 300k context ("300k of 1M is not a problem"), and runs CI five times on a stack it
- could check once. With them, it doesn't. The audit script and the repo have unit tests
- (`python3 -m unittest discover -s tests`), and on my own logs the audit reproduces my October total to the token.
- ## Licence
- MIT. By Jon McBee, [Coracle](https://github.com/coracle-us).
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.