Welcome to Limitless Trader Lab

Day 5 of 7 · Cohort intensive · 90–120 minutes

Add strategy + monitoring.

Yesterday you placed your first order; today you add Strategy + Monitor skills so the agent loop can decide which orders to place, and you run your first complete dry-run loop, on real markets, for thirty minutes.

90–120 minutes 1 deliverable #trader-lab · [D5]

Today you’ll learn

You’ll learn how to install the Strategy and Monitor skills on top of yesterday’s first-order plumbing, configure a kill-switch, and run a 30-minute dry-run of your first complete agent loop on real markets. This will help you watch Claude scan, evaluate, propose, and log decisions end-to-end without placing a real order, the exact loop you’ll wire Claude into on Day 6 and take live on Day 7.

Section 01

Why strategy + monitor.

An agent with tools but no strategy is a random number generator: it can call place_order all day, but the orders are arbitrary. An agent with a strategy but no monitoring is a black box: it does things, you have no idea what or why. Both pieces are non-negotiable. Today you drop in two more skills from the bonus pack, Strategy (the rule) and Monitor (the receipt), and run a dry-run loop that produces actual decisions and actual logs. By the end you can read the log and reconstruct every choice the agent made.

Section 02

The two skills.

The bonus pack ships with one Strategy skill (a starter that you can swap or extend later) and one Monitor skill (universal, the same monitor works for any strategy). Both are drop-in. No code to write today.

Strategy skill

A starter strategy that does one thing well: scan your watchlist, look for markets where the price has moved > X% in the last hour, decide whether to take the opposite side (mean-reversion). You can swap the rule, but the structure is the template.

scan_watchlist, apply your filter from Day 4.
evaluate_signal, score each market against the rule.
propose_action, return: buy / pass + size.

Monitor skill

Watches every decision the agent makes, writes structured NDJSON logs, fires alerts on drawdown thresholds, and exposes a kill-switch. The kill-switch is a file path, create the file, the agent halts on next iteration.

log_decision, append one structured line per loop.
check_drawdown, compare cum P&L to threshold.
check_kill_switch, halt if kill file exists.
alert, send Discord webhook on event.
Agents Academy · Module 12 · Monitoring deep-dive

Section 03

Drop in both.

Same install pattern as the Trading skill yesterday. Move the folders into your Claude skills directory. Restart Claude. Confirm the new tools are visible. Three skills, eleven tools, one loop.

1

You are now: installing Monitor and Strategy

From the bonus pack folder you unzipped on Day 3, copy monitor/ and strategy/ into the same Claude skills directory you used yesterday for the Trading skill.

Same parent folder as the trading/ skill from Day 4. Three sibling directories, one per skill.

2

You are now: configuring the kill-switch path

The Monitor skill needs a path it watches for the kill file. Open monitor/config.yaml and set:

monitor/config.yaml
kill_switch_path: ./data/kill.flag log_path: ./data/fills.ndjson drawdown_threshold_usd: -5.00 alert_webhook: # leave empty for today; we’ll wire Discord on Day 7

The Monitor skill reads KILL_SWITCH and AGENT_LOG env vars first and falls back to the values in config.yaml only if they aren’t set. On Railway (Day 3 already set KILL_SWITCH=/app/data/kill.flag and AGENT_LOG=/app/data/fills.ndjson), the env wins, the YAML paths above are the local dry-run fallback. Same kill file the Day 3 panel button writes, so the gesture works end-to-end. Same fills feed the panel renders, so today’s decision logs surface in the same view. Create the local ./data/ folder now if it doesn’t exist; the kill file itself should NOT exist yet, we’ll create it on Day 7 to test the halt.

3

You are now: confirming all three skills loaded

Restart Claude. Ask: “What skills do you have? List the tools each one provides.”

You should see Trading (4 tools), Monitor (4 tools), Strategy (3 tools). 11 tools total. If any are missing, the install path for that skill is wrong.

Section 04

Run the dry-run loop.

The dry-run is the moment the agent becomes one piece. You give Claude a single prompt, “run the agent loop in dry-run for 30 minutes”, and Claude orchestrates Strategy → Trading (in dry-run mode, no real orders) → Monitor on a timer. Every iteration: scan your watchlist, evaluate signals, propose an action, log the decision, sleep until next tick.

1

You are now: starting the loop

Switch to your Claude tab. Paste this prompt and send.

Paste into Claude
Run my agent loop in DRY-RUN mode for 30 minutes. Each iteration should: 1. scan_watchlist (using my Day 4 watchlist) 2. evaluate_signal on each market the scan returned 3. For any signal scoring above the threshold, propose_action 4. If propose_action returns buy, log it via Monitor, but DO NOT actually call place_order. We’re dry-running. 5. Call check_drawdown and check_kill_switch every iteration. 6. Sleep 60 seconds. Repeat. Tick interval: 60 seconds. Total runtime: 30 minutes (so roughly 30 iterations). After it finishes, summarize: how many signals fired, how many would-be-trades, what the cumulative dry-run P&L would have been if every proposed trade had filled at the proposed price. Start now.
2

You are now: watching the loop run

Claude will narrate each iteration. The Monitor skill writes structured logs to ./data/fills.ndjson (the same file the panel reads, the same folder as your kill file). Open that file in a separate terminal to tail it live:

In a second terminal
# macOS / Linux $ tail -f data/fills.ndjson | jq # Windows PowerShell PS> Get-Content data/fills.ndjson -Wait -Tail 10

Each line is one decision. You should see scans, signals, proposed actions, drawdown checks. The log is your audit trail, reading it is the actual skill.

3

You are now: reading the summary

After 30 minutes, Claude prints the summary. Expect something like:

Expected summary
Dry-run complete (30 iterations, 30:14 elapsed) Watchlist scans: 30 Signals fired: 7 Proposed actions: 3 - BUY YES on BTC>$70K @ $0.61 ($1) - SELL YES on Lakers @ $0.55 ($1) - BUY YES on Trump-EOY @ $0.23 ($1) Cumulative dry-run P&L: +$0.04 (if all filled) Kill-switch checks: 30 (file never appeared) Drawdown threshold trips: 0

If you hit an error mid-loop, copy it back into Claude with: “Got this error during the dry-run: [paste]. Help me debug.” Restart the loop after fixing, you want a clean 30-minute window in the log.

Section 05

Today’s deliverable.

Two artifacts: the summary printout and the log file. Both are evidence the agent works. Days 6 and 7 build on this loop, deploy it, then go live with it. Today’s log is the baseline.

Day 5 · Deliverable

Post in #trader-lab.

Screenshot of the dry-run summary from Claude.

Attach fills.ndjson (or paste a 5-line excerpt) with a [D5] tag in the caption.

Caption: “[D5] Day 5 done. [N] signals fired, [N] proposed actions, dry-run P&L $X. Snags: [none / brief].”

If your loop produced zero signals, that’s OK, the strategy is conservative by default. Post the log anyway. We’ll tune thresholds among the Trader Lab Alumni after Day 7.

Day 5 complete

Loop’s alive.

Strategy decides. Monitor watches. The kill-switch is wired. Thirty iterations of scan → evaluate → propose → log are sitting in fills.ndjson, your first readable audit trail.

01

Strategy and Monitor skills installed alongside Trading, 11 tools total callable by Claude, plus a configured monitor/config.yaml pointing at your kill-switch path.

02

A saved data/fills.ndjson on disk with 30 iterations of scan → evaluate → propose → log decisions, your first readable audit trail.

03

A working dry-run summary posted in #trader-lab with a [D5] tag that proves the loop stays inside its guardrails, the deploy target Day 6 lifts into the cloud.

Tomorrow: Day 6 wires Claude as the strategy brain on top of the deterministic skills you installed today, then runs a 60-minute dry-run on the host you deployed on Day 2. Production hardening of the panel happens too, rotate PANEL_TOKEN, confirm cancel-button confirmations, trip the kill switch from your phone.

Day 6 unlocks tomorrow at 9am local