x-signal-material-vault

name	x-signal-material-vault
description	Research recent market and user-demand signals for creating content on X (Twitter) with selectable 1-day, 3-day, 7-day, or 30-day windows. Use when the user asks for X选题、热点素材、真实用户需求、市场信号、账号观察、每日/每周/月度内容库, or wants cited material cards saved as Markdown.

X Signal Material Vault

Turn recent evidence into X-ready writing material. Optimize for freshness, real user language, and traceable sources instead of generic news summaries.

Resolve The Request

Determine these inputs from the request:

• topic: market, product, person, account list, or question to research.
• window: auto, 1d, 3d, 7d, or 30d.
• mode: daily, weekly, or monthly.
• count: desired number of material cards.
• destination: inline response or local material-vault path.

Use these defaults:

Input	Default
window	1d for daily mode, 7d only when weekly is explicit, 30d only when monthly or recent 30 days is explicit
mode	daily
count	8 for daily, 10 for weekly, 12 for monthly
timezone	current user or runtime timezone
destination	inline response

If the request says daily/每日/今天/最近一天/过去 24 小时/today's latest, use 1d. If it says weekly/本周, use 7d. If it says monthly/月度/最近 30 天, use 30d. Respect an explicit window and do not expand it unless the user requests expansion.

For an explicitly requested auto window, search in this order:

1. Search the rolling last 1 day.
2. Expand to 3 days only when fewer than five credible cards remain after deduplication.
3. Stop at 3 days. Use 7 days only when the user explicitly asks for 7d, recent 7 days, weekly review, or trend confirmation.
4. Use 30 days only when the user explicitly asks for 30d, recent 30 days, monthly material, or a durable archive.
5. Label every card with the window that supplied its evidence. Never describe 3-day, 7-day, or 30-day evidence as a 1-day signal.

Ask one concise question only when no topic, account list, or usable watchlist exists. Otherwise proceed.

Source Cadence

Apply a source-specific cadence instead of treating every platform as equally fresh.

Daily Sources: 1d

Check these sources every daily run and surface the hottest evidence first:

• X / Twitter;
• YouTube;
• GitHub;
• Hacker News;
• Reddit;
• Threads, the Meta platform. Treat user spellings such as thrise or T-H-R-E-A-D-IS as Threads.

For daily output, prioritize high-signal posts, videos, releases, issues, discussions, and user reactions from these six sources. Sort by freshness, evidence strength, verified engagement, and usefulness for an X post. Do not fill the daily list with routine updates.

Weekly Sources: 7d

Process these sources in the weekly review, together with the daily sources aggregated over 7 days:

• Digg;
• arXiv;
• Techmeme;
• LinkedIn;
• StockTwits;
• Pinterest;
• Bluesky;
• Perplexity;
• general Web search and other secondary context.

Do not silently present a weekly-only source as a same-day signal. Label it 7d and keep it out of the daily hot list unless the user explicitly requests it. A thin daily run may expand from 1d to 3d, but it must not expand to 7d without explicit user instruction.

Monthly Archive: 30d

Use all available sources for monthly durable themes, recurring user needs, product adoption, and long-term narratives. A monthly archive is not a breaking-news feed.

Cadence Output

In the coverage block, state:

• daily sources checked with a 1d window;
• weekly-only sources deferred or checked with a 7d window;
• whether a 30d archive was requested.

Research Current Evidence

Browse for every run. Do not answer a recent-signal request from memory.

1. Compute and state the exact start and end dates before searching.
2. Build queries for announcements, adoption, complaints, alternatives, pricing, failures, workarounds, buying intent, and repeated questions.
3. Prefer direct X posts and replies when accessible. Supplement them with primary product sources and user discussions.
4. Open the underlying source. Do not treat a search-result snippet as final evidence when the page is accessible.
5. Record the source publication date and distinguish it from the date an older page was indexed or updated.
6. Deduplicate repeated reporting, syndicated articles, reposts, and multiple pages describing the same event.
7. Separate observed facts from the inferred user need and proposed X angle.

Read research-protocol.md before conducting research. Follow its source tiers, scoring model, and confidence rules.

If direct X access is unavailable, say Direct X coverage: unavailable and use accessible user-signal sources. Never imply that X itself was searched when only web-indexed mentions or secondary reports were available.

Select Material

Keep a candidate only when it offers at least one of these:

• a repeated pain point or workaround;
• visible demand, switching intent, or willingness to pay;
• meaningful product adoption or rejection;
• a contested opinion with evidence on both sides;
• a practical lesson, case study, or tutorial opportunity;
• a market change with a clear consequence for the target audience.

Reject generic announcements with no user reaction, unsupported predictions, recycled claims, and engagement numbers that cannot be verified.

Rank candidates using the 11-point model in references/research-protocol.md. Use the score to order cards, not as a substitute for judgment.

Produce The Brief

Start every response with a coverage block:

Topic: ...
Mode: daily / weekly / monthly
Requested window: auto / 1d / 3d / 7d / 30d
Evidence range: YYYY-MM-DD to YYYY-MM-DD
Direct X coverage: available / partial / unavailable
Sources searched: ...
Cadence: daily source group / weekly source group / monthly archive
Expansion used: none / 1d -> 3d
30d mode: explicit only / not requested

Then produce:

1. 最值得写的 3 个选题
2. 素材卡
3. 继续观察
4. 证据不足或不建议写

Each material card must contain:

• title;
• evidence window;
• signal type;
• signal score;
• dated source links;
• observed evidence;
• inferred user need;
• X writing angle;
• one original hook;
• confidence label;
• save bucket.

Use Chinese by default for Chinese requests. Paraphrase source content and keep quotations brief.

Read output-contract.md when saving files or when the user requests structured output.

Save The Material Vault

Save only when the user asks for local files or provides a destination.

Create a JSON run file matching references/output-contract.md, then validate and render it with the bundled script:

python3 scripts/material_vault.py validate --input RUN.json
python3 scripts/material_vault.py render --input RUN.json
python3 scripts/material_vault.py save --input RUN.json --vault MATERIAL_VAULT

Resolve scripts/material_vault.py relative to this SKILL.md, not relative to the user's current directory.

The save command writes one of:

material-vault/daily/YYYY-MM-DD.md
material-vault/weekly/YYYY-Www.md
material-vault/monthly/YYYY-MM.md

Do not overwrite an existing period file unless the user explicitly approves replacement. Report the final path after saving.

Quality Gate

Before returning the result, verify:

• every material card has at least one openable source URL;
• every source date falls inside the card's stated window;
• Hot cards have corroboration from at least two independent sources;
• facts, inference, and writing advice are distinguishable;
• duplicate events appear only once;
• any source-access limitation is visible near the top;
• fewer cards are returned instead of filling the quota with weak material.