How to analyze XAU/USD with AI — a complete step-by-step workflow (2026)

This post gives you two workflows:

1. The basic workflow — what you can do today with ChatGPT, Claude or Gemini. Free. Works. Has a ceiling.
2. The professional workflow — how the process changes once you use a purpose-built multi-agent system like Analiza.LH.

Both produce a trade plan in roughly 10-15 minutes. The difference is in consistency, purity, and how much trust you can place in the output under pressure.

Step 0 — Know your methodology cold

Before any AI tool enters the picture, you need to own a methodology. If you don’t know SMC, ICT, Wyckoff or Elliott at the level where you can verify what the AI says, AI will hurt you more than help.

The reason: an AI that stays framework-pure 95% of the time will drift 5% of the time. If you can’t spot the drift, you’ll trade on drifted analysis.

This is not negotiable. Do not skip.

Step 1 — Mark the higher timeframes manually

Open the 1D and 4H charts. Mark:

• Last major structure break (BOS or CHoCH).
• Unmitigated order blocks on HTF.
• Equal highs and equal lows (liquidity pools).
• Key levels with 2+ touches.

Do this yourself, with the chart. No AI yet. This step is the difference between AI that helps you and AI that leads you.

Why: when you feed the AI context later, you’re going to cross-check its read against yours. If you didn’t do this step, you have no reference point.

Step 2 — Identify the current session killzone

XAU/USD moves hardest during London open (07:00-10:00 UTC) and NY AM (13:00-16:00 UTC). Analysis during these windows matters differently than analysis during Asia.

Note the session, note the time to next killzone. This is context the AI needs.

Step 3 — Run the AI read

This is where the basic and professional workflows diverge.

Basic version (single-LLM chat)

Open ChatGPT, Claude or Gemini. Write a prompt that includes:

• The methodology you want applied (SMC, ICT, Wyckoff, Elliott — pick one, don’t mix).
• The timeframes to analyze (we recommend 15m/1H/4H composite for a day trade).
• The structure of the output you want (bias, key levels, setup if any, trigger condition, invalidation).
• A rule to state “no setup” when nothing clean is there.

Paste whatever candle data you have and ask for the analysis.

Limitations of this version: you are responsible for data quality, staleness, consistency, and cross-checking. Level precision will wobble. Methodology purity depends on how tightly you prompted. Re-runs will produce slightly different outputs. You accept this in exchange for $0 cost.

Professional version (multi-agent system)

With a system like Analiza, Step 3 collapses to: pick your expert (SMC, ICT, Wyckoff or Elliott), click analyze. The system delivers a structured read — bias, levels, setup, trigger, invalidation — with methodology purity that a single prompt cannot reach.

What you don’t see, but benefits from:

• Specialized agents — one per framework, not a single model trying to hold all four in its head.
• Deep learning over market structure, not over chat context.
• Multi-layer orchestration — market-state, methodology, and decision layered separately.

You’re not wrestling with prompts. You’re reading the output.

Step 4 — Cross-check against your HTF read

The AI read is a draft. Your job now:

• Do the AI’s structure calls match what you marked in Step 1?
• Are the unmitigated OBs the same ones you flagged?
• Is the bias consistent with HTF?

If yes on all three: you have a verified read. Proceed. If no: stop. Either the AI drifted, or you missed something. Figure out which before going further.

Cross-checking is more important in the basic workflow because drift is more frequent. It’s still essential in the professional workflow — any AI read is a draft until you’ve verified it.

Step 5 — Define risk before reading the setup

Here is the psychological trap: if you read a “LONG with entry 4801, stop 4790, target 4830” before you’ve defined your risk, you will unconsciously size up because the setup looks good.

Reverse the order:

1. Define risk per trade (0.5% for challenge, 1% max for live personal).
2. Define daily stop (-1.5% balance → close terminal 24h).
3. Define session stop (-1% → pause until next session).

Write these numbers down. Then read the AI’s setup.

Step 6 — Validate the trigger condition

This is where most “AI-assisted trades” go wrong. The AI gives you an entry, but the trigger is critical.

A good trigger is specific: “15m close above 4801 with rejection wick to 4798.” A bad trigger is vague: “when price reaches 4801 and shows strength.”

If the trigger isn’t specific enough to wait for, either tighten it yourself or skip the trade. Specificity is your defense against FOMO entries.

Step 7 — Set the exit plan before entry

Most traders set entry + stop + TP, then manage by feel. Don’t.

Before entry, write:

• Partial at 1R (close 50%, move stop to breakeven).
• Runner to 2R or 3R depending on HTF targets.
• Invalidation: if the market shows X, close regardless of price.

Having this written makes emotional mid-trade decisions unnecessary.

Step 8 — Execute the trade

Send the order. Limit if possible, market only if you missed the level and the structure is still intact.

Once in the trade: do not re-query the AI for reassurance. This is anchoring bias in action. The plan was made, execute the plan.

Step 9 — Journal the trade immediately

Open your journal before you close the terminal. Log:

• Entry reasoning (quote the AI’s read, quote your cross-check).
• Risk taken.
• Trigger that fired.
• How you felt at entry (confidence, doubt, neutral).
• Expected outcome vs actual.

AI can help structure the journal entry — paste the context and ask for a post-mortem prompt. But you author the feelings part. No AI does that honestly.

Step 10 — Post-session review

At end of session, run one more AI pass:

“I took this trade at [time]. Entry [price], stop [price], outcome [result]. Here’s the 4H structure at the time. What did I miss or execute well?”

The AI critique is not absolute truth — it’s a second pair of eyes. Accept what matches evidence, discard what feels like hallucination.

The common mistakes

Skipping Step 1. If you feed the AI a chart you haven’t read yourself, you can’t cross-check. This is the #1 reason AI-assisted trading fails.

Confirmation loops. Asking the AI to “re-check” until it agrees with your bias. Every time you do this, you’re training yourself to override evidence.

Ignoring session context. Analysis during Asia is worth less than during London or NY. AI doesn’t know what session we’re in unless you tell it.

Staying on the basic workflow past its ceiling. Single-LLM prompts work for learning and casual analysis. Once you’re trading meaningful size, the consistency gap between a prompt and a multi-agent system matters. That’s when traders graduate.

When to upgrade from basic to professional

Signs you’ve hit the single-LLM ceiling:

• You’ve spent hours tuning a prompt and you’re still getting inconsistent reads.
• Re-runs on the same chart produce noticeably different levels.
• The model drifts between frameworks when the setup is ambiguous.
• You find yourself cross-checking against another LLM to “validate” the first.

At that point, a purpose-built system stops being optional. The cost of one bad trade from a hallucinated level is 10x the cost of a proper tool.

Try the professional workflow free

Analiza.LH is purpose-built for XAU/USD technical analysis. Multi-agent architecture, specialized experts per methodology, deep learning under the hood. Pick your framework, run your first analysis free. Compare it to whatever prompt you’re using today.

Related reading

• AI applied to trading in 2026: what actually works, what doesn’t
• ChatGPT vs Claude vs Gemini for technical analysis — real XAU/USD test