How to use the Elliott Wave Signal Engine agent skill

What does the elliott-wave skill make the agent do?

Elliott Wave Theory assumes markets move in repeating structures. A 5-wave impulse (waves 1, 2, 3, 4, 5) moves with the main trend, and a 3-wave corrective pattern (A, B, C) moves against it as a pullback. The skill teaches the agent these structures and the rules that define them.

The skill is built on a Zigzag swing-point detector and a pure pandas implementation. It checks three rules for impulse waves: wave 2 cannot retrace beyond the start of wave 1, wave 3 cannot be the shortest impulse wave, and wave 4 cannot enter the price territory of wave 1. It then compares wave lengths with Fibonacci relationships, such as wave 3 being about 1.618 times wave 1, with a tolerance of 15 percent.

• Impulse wave: 5 waves, trend-following, marks the main trend direction.
• Corrective wave: 3 waves, counter-trend, marks a pullback.
• Signal logic: a completed 5-wave advance gives a sell (trend top); a completed ABC pullback gives a buy (correction finished); wave 3 in progress gives no reversal signal.

How do you install the elliott-wave skill?

Download the elliott-wave ZIP from this page and unzip it. Then place the resulting elliott-wave folder in one of two locations: your personal skills folder at ~/.claude/skills/, or a project-specific .claude/skills/ folder inside the project you work in. A personal install is available to all your projects; a project install applies only to that project.

If you use the Claude apps instead of files on disk, you can upload the ZIP under Settings → Capabilities → Skills.

The skill lists these Python dependencies, which you install with the command below. The engine excerpt does not obviously use requests, so it is listed but may not be needed for the core engine.

• Personal folder: ~/.claude/skills/elliott-wave/
• Project folder: .claude/skills/elliott-wave/ inside your project
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

How do you trigger the skill with a natural request?

You do not need special syntax. Describe what you want in plain language that mentions Elliott Wave, wave counting, swing points, or whether a trend or pullback looks mature or finished. The agent matches your request against the skill description and loads it when it fits.

Give the agent your OHLC price data (open, high, low, close candles) in the conversation or as a file, because the skill does not bring its own market data. Name the instrument and timeframe so the output is easier to read.

• Try: 'Using Elliott Wave, does this daily price series look like a completed 5-wave advance?'
• Try: 'Run the Elliott Wave signal engine on these candles and tell me the latest signal.'
• Try: 'Check whether this pullback looks like a finished ABC correction.'

How do you read the signals?

The engine outputs one of three values for each bar: 1 for long, -1 for short, and 0 for stand aside. A swing point is only known a set number of bars after it forms (the swing_window parameter, default 10), so the signal is stamped on that confirmation bar, not on the swing bar itself. This means the output is not a look-ahead prediction of turning points.

The skill's default parameters are swing_window 10, fib_tolerance 0.15, and min_wave_bars 5 (the minimum number of candles per wave). Changing them changes how many patterns pass the checks.

• 1: long signal, after a corrective pattern is complete.
• -1: short signal, after a 5-wave advance is complete.
• 0: stand aside, including when wave 3 is still in progress.

What should you ask the agent to show you?

Ask for the reasoning, not just the number. A useful answer names the swing points found, the wave labels assigned, and which Fibonacci ratios passed or failed. If the agent cannot match a clean pattern, the skill is designed to stand aside rather than force a count.

For backtesting, ask the agent to run the engine across a historical series and report the signals it stamps, so you can compare them with your own rules and costs.

• Ask which swing points were confirmed and at which bar.
• Ask which wave ratios fell within the 15 percent tolerance.
• Ask how many bars each wave lasted against the minimum of 5.

What are the safety notes: scripts and network access?

The skill runs scripts. Its bundled example_signal_engine.py is a Python program that the agent can execute over your data, so read the code before you run it on sensitive files. The skill does not use the network, according to its listing, so it should not send your price data elsewhere.

Wave counting is highly subjective. The skill notes that multiple interpretations exist and uses a single simplest interpretation that would rather miss a signal than misclassify one. Treat its output as one input to your own analysis, not as a trading instruction.

• Review example_signal_engine.py before running it on private data.
• Check the code path yourself if you need to confirm there is no network activity.
• Keep your own risk controls; the skill is not a validated profit tool.

Tips for getting useful results

Supply clean, complete OHLC data with consistent timeframes. Gaps, duplicated bars, or mixed timeframes make swing detection unreliable. Read the reference files, references/波浪结构.md (wave structures) and references/Fibonacci浪间关系.md (Fibonacci relationships), to understand the rules behind each label.

If you want to understand the logic rather than only get signals, ask the agent to walk through SKILL.md and the reference files step by step. That is also the fastest way to learn how wave-counting can be coded with pandas.

• Use the reference files to check which ratio targets the engine uses, such as wave 2 at 0.5 to 0.618 of wave 1.
• Compare the signals with a simple benchmark before drawing conclusions.
• Expect 0 (stand aside) often; it is a designed outcome.