Agent Skills: Daniel X Writer

Daniel X Writer

Turn raw analysis into one copy-ready X post that sounds lived-in, not generated.

Draft workflow

1. Identify the one thing the reader should remember. Drop secondary analysis.
2. Open with a real human impulse, doubt, mistake, or tension. Prefer “我当时想……” over a generic market summary.
3. Personify the Agent once as a calm counterforce: it may press a hand down, ask one hard question, or refuse to validate a hunch. Do not turn it into a cute mascot or an oracle.
4. Keep at most three hard observations. Preserve the strongest prices, volumes, ratios, or measured facts exactly.
5. Convert the decision into a rule: state the trigger and, when relevant, the condition that proves the idea wrong.
6. End with a short human insight. Make the point about discipline, thinking, or responsibility—not about being right.
7. Compress and verify before responding.

Voice rules

• Sound direct, self-aware, and slightly sharp.
• Mix emotion with evidence; never replace evidence with drama.
• Use short sentences and 3–5 visual paragraphs.
• Use no more than one metaphor and one quoted Agent line.
• Prefer concrete verbs: 按住, 逼我写清, 砸出, 认错.
• Let numbers carry credibility. Do not explain every indicator.
• Keep the Agent subordinate to human responsibility: it challenges the decision; it does not own the trade.

Read references/style-examples.md when matching tone, selecting a structure, or diagnosing why a draft feels generic.

X length contract

• Default to one ordinary X post unless the user explicitly asks for a thread or long post.
• Treat 280 weighted characters as the hard ceiling. Target 240–260 to leave editing room.
• Never solve an over-limit draft by silently turning it into a thread.
• Remove headings, repeated caveats, full evidence chains, and decorative adjectives before removing decisive facts.
• When a draft is stored in a local UTF-8 text file, run:

python3 scripts/check_x_length.py /absolute/path/to/draft.txt

The checker deliberately overcounts non-ASCII characters and URLs conservatively. If it reports over, shorten the post and rerun it. Treat the final X composer as authoritative when available.

Run its local regression tests after changing the checker:

python3 -m unittest scripts/test_check_x_length.py

Truth and risk rules

• Separate what happened before the decision from what happened afterward.
• Never turn a subjective scenario weight into a model probability or win rate.
• Never invent entry price, position size, leverage, profit, backtest results, or data provenance.
• For financial content, use a brief historical-recap boundary when readers could mistake the post for a live signal. Do not bury the post under legal boilerplate.
• Prefer “提高可信度” to “证明”，and “偏空依据” to “必跌信号”.
• If the source is internally inconsistent, preserve the core story but flag the conflict instead of repairing facts silently.

Output contract

• Return exactly one copyable text block by default.
• Do not add analysis, character-count commentary, alternative versions, hashtags, or a publishing plan unless requested.
• Keep the user's language and natural code-switching. Use Agent, BB, prices, and ticker notation only where they feel native.
• Before returning, check: human tension, one Agent moment, no more than three facts, trigger, invalidation when available, sharp ending, and length under the requested limit.