MOCKUP · 2026-04-20 · Apollo · §03.4 Thinking Layer — Strategy carousel
§03.4 Thinking Layer
A sentence becomes a strategy. The strategy runs every day. Never moves capital.
LIVE STRATEGY LIBRARY 6 strategies · 4 core · 1 custom · 1 agent-authored
strategies/01-pre-market-regime-read.md
Core
trigger
daily · 07:30 CET
connector
shared-bus + slack
model
Sonnet
Prompt 01 / 03

Every morning before markets open, tell me whether today feels risk-on or risk-off — and why.

Execution 02 / 03

Read FX overnight (DXY, EUR/USD, USD/JPY), Asia close (Nikkei, Kospi, HSI), equity futures (ES, NQ, DAX), bond futures, and last-12h macro news. Score regime across three axes: risk (on/off), dollar (strong/weak), vol (expanding/contracting). Produce a 3-line brief: the call, the evidence, the biggest risk to the call.

Customize 03 / 03

Each morning review whether the call held up and where it missed. After a week of misses on FX, widen the evidence set. After a clean run, tighten the threshold to reduce noise. Swap the output channel when your routine changes.

01 / 06 · Pre-Market Regime
strategies/02-fomc-ecb-decode.md
Core
trigger
event · statement release · <60s
connector
slack + shared-bus
model
Sonnet
Prompt 01 / 03

A central bank statement has just dropped. Pull the latest statement from the Fed or ECB and compare it line by line against the prior meeting. Flag every language change — forward guidance, inflation framing, balance-of-risks shifts, dot-plot moves. Classify each as hawkish, dovish, or neutral with an evidence quote. Map the overall delta to a directional bias for equities (ES, NQ), bonds (TY, Bund), USD (DXY), and front-end rates. Ship to Slack within 60 seconds of release.

Execution 02 / 03

Webhook fires on statement release. Strategy pulls the document, runs the diff, writes the structured brief to the shared bus, and routes a formatted summary to #macro-desk in Slack with evidence quotes attached. The broker workflow picks up the bias scores on next read cycle.

Customize 03 / 03

After each meeting, track how well the cross-asset call played out against the first-hour move. Reweight the language-tag sensitivities over time — the model gets better the more meetings it sees.

02 / 06 · FOMC/ECB Decode
strategies/03-positioning-scan.md
Core
trigger
weekly · Sun · 19:00 CET
connector
shared-bus
model
Sonnet
Prompt 01 / 03

Every Sunday evening, pull the latest CFTC COT data for equity index futures, currency futures, and commodity futures. For each contract, compute current net positioning (commercials, large specs, small specs) as a percentile against the trailing 52-week range. Flag any position at or beyond the 95th or 5th percentile. Pull the latest SPX and VIX option skew — flag any skew more than 2σ from its 30-day mean. Pull weekly ETF flows — flag any concentrated in or out of a single sector. Cross-reference every flag against my current book. Output: crowded longs, crowded shorts, stretched skews, flow divergence — each with evidence.

Execution 02 / 03

Cron fires Sun 19:00 CET. Strategy pulls from CFTC API, option-data vendor, and ETF-flow source in parallel. Joins against current book snapshot on the shared bus. Writes the positioning brief to #positioning plus a JSON record for Monday's pre-market review.

Customize 03 / 03

Track which flagged positions actually unwound in the following weeks and which held. Add instruments that keep surprising you. Prune flags that never translate into moves — they are noise.

03 / 06 · Positioning Scan
strategies/04-earnings-pre-read.md
Custom
trigger
quarterly · first Monday of season
connector
shared-bus
model
Sonnet
Prompt 01 / 03

Before each earnings season, walk my watchlist and tell me what analysts will actually be looking for.

Execution 02 / 03

For each watchlist name with earnings in the next 14 days, pull consensus EPS / revenue / growth. Pull last 3 broker notes; extract: guidance concern, margin watch, segment focus, capex trajectory, buyback pace. Build a prioritized calendar ranked by book exposure × print-date proximity.

Customize 03 / 03

After each season, review which focus items actually moved the stock. Drop analyst sources that add noise. Add the ones that keep calling the print right. Expand the window if too many names are missed.

04 / 06 · Earnings Pre-Read
strategies/05-risk-sentinel.md
Agent-Authored
trigger
weekly · Sat · 16:00 CET
connector
slack · private
model
Sonnet
Prompt 01 / 03

Watch my trading behavior weekly. Read the last seven days of trades from the trade log. Compute five patterns: trades per day vs 30-day baseline, average position size vs baseline, sector concentration delta, win-rate vs expectancy drift, and revenge-trade pattern — any new position opened within 30 minutes of a stop-out on the same or correlated instrument. Flag any metric that has drifted more than 1σ from its 30-day mean. For each flag, pull the specific trade IDs. Send a private note on Saturday afternoon with the flags ranked by deviation magnitude.

Execution 02 / 03

Cron fires Sat 16:00 CET. Strategy queries SQLite trade log, runs the five pattern computations, diffs against the 30-day baseline, ranks by deviation. Writes the behavioral note to a private Slack DM — never to a public channel. Logs the run to the strategy audit trail.

Customize 03 / 03

Review flags weekly. Drop patterns that keep firing without signal. Add patterns after painful lessons — each drawdown teaches what to watch next. Adjust thresholds as the book size grows.

05 / 06 · Risk Sentinel
strategies/06-signal-triangulation.md
Core
trigger
event · TradingView webhook
connector
shared-bus
model
Sonnet
Prompt 01 / 03

When TradingView fires a signal, don't surface it until you've cross-checked macro, positioning, and my current book.

Execution 02 / 03

On webhook: pull current macro regime (strategy 01), current positioning scan (strategy 03), current book exposure, and recent news for the ticker. Score alignment: signal agrees with macro? with positioning? with book? Produce a triage verdict — accept / hold / reject — with reasoning. Write to signal/{id}/triage.json. Analyst workflow reads this before opening a proposal.

Customize 03 / 03

Review accept / reject calls against eventual P&L monthly. If rejected trades would have made money, loosen a layer. If accepted trades didn't, tighten. Reweight the three layers as the book teaches you which one matters most.

06 / 06 · Signal Triangulation
Expand

The thinking layer is a set of single-purpose LLM sessions that reason over incoming data and produce structured output. It reads, reasons, writes. It never executes orders, never touches credentials, never stores state.

A strategy begins with what you tell it. A single sentence works — "tell me if today feels risk-on or risk-off." A detailed paragraph with explicit filters and thresholds also works. The shorter the prompt, the more the LLM fills in; the longer, the more control you keep. Either way, the system elaborates that instruction into a working strategy: data pulled, signals scored, brief formatted, channel routed.

Observe and adjust. A strategy is not a setting — it is a living instrument. Each week, review the outputs against what actually happened. Where did the call hold? Where did it miss? Tighten thresholds. Widen the evidence set. Drop a flag that never signals. Change the channel when your routine changes. The prompt stays short; the strategy improves.

Agent-authored strategies. The system also writes its own. When a new catalyst class appears — a central bank speaker, a recurring data release, a behavioral pattern in the trade log — an agent can draft a strategy, propose it to you, and register it on approval. The library grows without engineering.

Boundary. Strategies produce judgments, not transactions. A strategy that flags a high-confidence signal writes to the shared bus. The Analyst workflow reads it, the Risk workflow sizes it, the Operator signs it, Execution submits. The thinking layer ends at the signature gate — always.