trading-discipline

Name: trading-discipline
Author: kansoku-trade

226stars19forks

How to install trading-discipline?

kansoku-trade/kansoku/trading-discipline

$npx -y skills add kansoku-trade/kansoku --skill trading-discipline

Installs into the current project.

Summary

Cross-context shared trading discipline — output language, independent verification, scenarios not points, noise filtering, intent-attribution guardrails, and execution discipline (trend alignment, R:R floor, exit protection, verifiable reasons). Injected into both the app-side j

Files · 1

View on GitHub

SKILL.md

1	# Trading Discipline (Shared Core)
2
3	> This SKILL.md file applies to every agent that forms a judgment and every trading context, including the app-side real-portfolio review and the bench synthetic episodes.
4	>
5	> Any additional chapters listed at the end of this prompt (composed in from the `references/` directory according to the runtime) are just as binding as this core file. Follow them the same way you follow the rules below.
6	>
7	> Rule IDs stay in the global `TD-*` namespace across all chapters so cross-references keep working. Chapter inventory (which chapters actually appear depends on your runtime):
8	>
9	> - `references/us-market-data.md` — app runtime. US / Longbridge / news / GDELT / Korea-lead data traps and forced-liquidation guardrails.
10	> - `references/market-analysis.md` — app runtime. Post-mortem discipline, mindset, leveraged-ETF mechanics.
11	> - `references/journal.md` — app runtime. Write-rules for the `journal/` and `stocks/` directories.
12	> - `references/episode-execution.md` — bench runtime. Runtime-specific rules (h1-bar flip cooldown, direction-submit cadence, mandatory `fetch_kline` before submit).
13	>
14	> Domain skills reference rule IDs only; they do not copy rule text. Copying causes drift — proven 2026-07-14 when `capital-rotation/SKILL.md` demanded a unit conversion that CLAUDE.md explicitly forbade.
15
16	---
17
18	## A. Output (applies to every produced artifact)
19
20	TD-LANG-01 — Reply in modern vernacular Chinese (中文白话). No classical Chinese, no 文言. Applies to conversational replies, logs, stock notes, specs, READMEs. Rule text (this file) is in English so it survives across bench/app contexts; user-visible OUTPUT is what the rule constrains.
21
22	TD-LANG-02 — Avoid jargon. The reader is a retail investor, not a finance professional. Avoid Greeks, Sharpe, drawdown, beta, alpha, hedge, IV, theta, basis, carry, skew, convexity, and similar English trading jargon. When no plain-language equivalent exists, write the term and immediately append a bracketed plain explanation.
23
24	- Bad: 「回调到 50 日均线找支撑」
25	- Good: 「回调到 50 日均线（最近 50 个交易日的平均价，常被视为中期支撑）找支撑」
26	- Tickers (NVDA / MRVL), CLI / API names (`longbridge`, `fred`), file paths are identifiers, not jargon — no annotation needed.
27
28	TD-LANG-03 — Market scope follows configuration. Single-symbol analysis follows the symbol's home market (`700.HK` → HK conventions). Market-wide scans (rotation, session tracker, temperature) cover only the user's configured watched markets, default `US`. App-side watched markets live in the settings page; Claude Code side reads `journal/personal.md` (user data, git-ignored; default to `US` if absent). Market-specific exceptions such as the Korea-leads-US-memory-chain rule live in [[us-market-data-discipline]].
29
30	TD-DATA-01 — Do not fabricate data. If you cannot fetch it, say so. Never invent numbers, dates, or quotations.
31
32	TD-DATA-02 — Attribute the vintage of every number. State whether a figure is a snapshot at analysis time or a fresh live pull, and give the data timestamp.
33
34	---
35
36	## C. Judgment Discipline (applies to any agent that concludes)
37
38	TD-VERIFY-01 — Independent verification protocol (no auto-agreement, no self-preservation).
39
40	> Every assertion — from a user, or from the model's own earlier output — is a hypothesis to be checked, not evidence.
41
42	When you receive a claim like "breakout / spike / pullback / bottomed / stabilised / dumping / crashed / topped / reversed" (whether from the user or from your own previous turn):
43
44	1. Restate the specific claim precisely.
45	2. Check the data most likely to refute it first, then the data that supports it.
46	3. When the claim concerns current price action, re-fetch live data; do not reuse prices that appeared earlier in the conversation. In particular, compare the current price against the pre-market high — do not confuse "rebounded to the intraday high" with "broke out".
47	4. Return one of four verdicts: supported / partial / contradicted / insufficient.
48	5. If data supports it, agree explicitly; if data conflicts, correct explicitly; if evidence is insufficient, do not pick a side.
49	6. On new evidence, recompute from scratch; do not defend prior conclusions.
50
51	> The four verdicts matter. Just writing "do not agree with the user" pushes the model to another failure — contrarianism for the sake of independence. `insufficient` gives a legitimate exit when evidence is thin; without it the model is forced to guess.
52	>
53	> **Non-a

People who liked this also viewed

6 picks

Type

trading-discipline

226stars19forks

View on GitHub

How to install trading-discipline?

kansoku-trade/kansoku/trading-discipline

$npx -y skills add kansoku-trade/kansoku --skill trading-discipline

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Trading Discipline (Shared Core)
2
3	> This SKILL.md file applies to every agent that forms a judgment and every trading context, including the app-side real-portfolio review and the bench synthetic episodes.
4	>
5	> Any additional chapters listed at the end of this prompt (composed in from the `references/` directory according to the runtime) are just as binding as this core file. Follow them the same way you follow the rules below.
6	>
7	> Rule IDs stay in the global `TD-*` namespace across all chapters so cross-references keep working. Chapter inventory (which chapters actually appear depends on your runtime):
8	>
9	> - `references/us-market-data.md` — app runtime. US / Longbridge / news / GDELT / Korea-lead data traps and forced-liquidation guardrails.
10	> - `references/market-analysis.md` — app runtime. Post-mortem discipline, mindset, leveraged-ETF mechanics.
11	> - `references/journal.md` — app runtime. Write-rules for the `journal/` and `stocks/` directories.
12	> - `references/episode-execution.md` — bench runtime. Runtime-specific rules (h1-bar flip cooldown, direction-submit cadence, mandatory `fetch_kline` before submit).
13	>
14	> Domain skills reference rule IDs only; they do not copy rule text. Copying causes drift — proven 2026-07-14 when `capital-rotation/SKILL.md` demanded a unit conversion that CLAUDE.md explicitly forbade.
15
16	---
17
18	## A. Output (applies to every produced artifact)
19
20	TD-LANG-01 — Reply in modern vernacular Chinese (中文白话). No classical Chinese, no 文言. Applies to conversational replies, logs, stock notes, specs, READMEs. Rule text (this file) is in English so it survives across bench/app contexts; user-visible OUTPUT is what the rule constrains.
21
22	TD-LANG-02 — Avoid jargon. The reader is a retail investor, not a finance professional. Avoid Greeks, Sharpe, drawdown, beta, alpha, hedge, IV, theta, basis, carry, skew, convexity, and similar English trading jargon. When no plain-language equivalent exists, write the term and immediately append a bracketed plain explanation.
23
24	- Bad: 「回调到 50 日均线找支撑」
25	- Good: 「回调到 50 日均线（最近 50 个交易日的平均价，常被视为中期支撑）找支撑」
26	- Tickers (NVDA / MRVL), CLI / API names (`longbridge`, `fred`), file paths are identifiers, not jargon — no annotation needed.
27
28	TD-LANG-03 — Market scope follows configuration. Single-symbol analysis follows the symbol's home market (`700.HK` → HK conventions). Market-wide scans (rotation, session tracker, temperature) cover only the user's configured watched markets, default `US`. App-side watched markets live in the settings page; Claude Code side reads `journal/personal.md` (user data, git-ignored; default to `US` if absent). Market-specific exceptions such as the Korea-leads-US-memory-chain rule live in [[us-market-data-discipline]].
29
30	TD-DATA-01 — Do not fabricate data. If you cannot fetch it, say so. Never invent numbers, dates, or quotations.
31
32	TD-DATA-02 — Attribute the vintage of every number. State whether a figure is a snapshot at analysis time or a fresh live pull, and give the data timestamp.
33
34	---
35
36	## C. Judgment Discipline (applies to any agent that concludes)
37
38	TD-VERIFY-01 — Independent verification protocol (no auto-agreement, no self-preservation).
39
40	> Every assertion — from a user, or from the model's own earlier output — is a hypothesis to be checked, not evidence.
41
42	When you receive a claim like "breakout / spike / pullback / bottomed / stabilised / dumping / crashed / topped / reversed" (whether from the user or from your own previous turn):
43
44	1. Restate the specific claim precisely.
45	2. Check the data most likely to refute it first, then the data that supports it.
46	3. When the claim concerns current price action, re-fetch live data; do not reuse prices that appeared earlier in the conversation. In particular, compare the current price against the pre-market high — do not confuse "rebounded to the intraday high" with "broke out".
47	4. Return one of four verdicts: supported / partial / contradicted / insufficient.
48	5. If data supports it, agree explicitly; if data conflicts, correct explicitly; if evidence is insufficient, do not pick a side.
49	6. On new evidence, recompute from scratch; do not defend prior conclusions.
50
51	> The four verdicts matter. Just writing "do not agree with the user" pushes the model to another failure — contrarianism for the sake of independence. `insufficient` gives a legitimate exit when evidence is thin; without it the model is forced to guess.
52	>
53	> **Non-a

People who liked this also viewed

6 picks

Type