Versioning keeps a simulated AI trading agent honest. Each paper sample should be tied to the exact prompt, output format, risk rule, market scope, and review window that produced it.
Trading Boy does not execute live trades, hold funds, or provide financial advice. Prompt versioning helps compare simulated paper samples; it does not prove live fills, future returns, or readiness for real capital.
A prompt version is more than a saved text block. It should capture every instruction that could change the agent's paper behavior.
| Versioned item | What to record | Why it matters | Change rule |
|---|---|---|---|
| Prompt text | The exact role, task, setup, and output instructions. | Small wording changes can change behavior. | Any wording change creates a new version. |
| Persona | Agent style, market role, and review stance. | Persona changes can alter aggressiveness and skip discipline. | Change only when persona behavior is the issue. |
| Market scope | Watchlist, timeframe, data source, and review window. | Different markets make samples hard to compare. | Do not change mid-sample. |
| Setup and skip rules | Entry rule, invalidation, skip conditions, and excluded records. | These define what counts as a valid decision. | Change one behavior at a time. |
| Risk limits | Paper size cap, drawdown pause, and exposure limit. | Risk changes can explain paper-result changes. | Separate risk tests from setup tests. |
| Output format | Required fields for entries, exits, skips, tags, and review questions. | Output changes affect journal quality. | Version any required-field change. |
Version 1.2: The agent used a trend-continuation setup on BTC, ETH, and SOL. The output format required setup, thesis, invalidation, and paper risk, but skip notes could be free-form.
Review finding: The paper sample had clean entries, but weak skips. Several skips said market unclear without naming the failed condition. The evaluation checklist marked skip discipline as failed.
Version 1.3 change: The setup, market scope, and risk limits stayed fixed. The only change was a stricter skip output format requiring blocked condition, evidence, and next review question.
Comparison rule: Version 1.3 should be judged against skip quality first. It should not be declared better or worse based only on paper PnL because the change targeted review evidence.
Weak changes rewrite the whole prompt after one result, change the watchlist and risk cap at the same time, hide failed versions, or compare samples from completely different regimes without a caveat.
When in doubt, keep the agent in paper mode and collect another stable sample before changing the prompt.
Use the AI trading agent prompt template to write the first version. Use the output format template when journal records are inconsistent. Use the simulated agent review process to collect evidence. Use the evaluation checklist before changing a version.
This sequence prevents the common pattern where a trader rewrites the prompt because a single paper trade was exciting or frustrating. Versioning turns prompt changes into testable hypotheses rather than emotional edits.
Use names that show what changed without making performance claims. A useful label might be momentum-v1.4-skip-output, where the strategy family, version number, and change type are visible. Avoid labels like winner, improved, safe, or profitable. Those names bias the review before the next paper sample is collected.
Keep a brief changelog beside the version label. The changelog should include the previous version, the reason for change, the exact prompt or output field changed, the review window that triggered the change, and the metric or checklist row that should improve. If more than one major behavior changes at once, create a caveat so future reviewers know the next sample may not be comparable.
Versioning ties each paper-trading sample to the exact prompt, persona, rules, risk limits, output format, and review window that produced it.
Version the prompt text, persona, market scope, setup rules, skip rules, output format, risk limits, model notes, and review date.
No. Prompt versioning improves simulated review quality, but it does not prove live execution, guarantee returns, or replace financial advice.