Research and triage learning materials; read and explain sources for "请你读" / "精读"; process "素材:" intakes and "读完" closure. Explicit reading follows the source's argument, not career or project ranking.
69%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Critical
Do not install without reviewing
Fix and improve this skill with Tessl
tessl review fix ./.codex/skills/research-material-scout/SKILL.md请你读 / 精读 / 帮我读一下: read the reading protocol before source work. It owns source coverage, explanation, figures and delivery checks for articles, papers, chapters, repos and bundles. Use the source routes below as needed; skip discovery ranking and intake workflows. Explain the author's argument before project applications. Do not edit notes, indexes or lifecycle/ranking; cache only reading artifacts.素材 / 调研 / 继续调研: follow discovery, source verification and managed intake below. Decision Context guides what to select, not what an explicitly requested reading may omit.整理笔记 / 读完: follow repository note-integration rules and detail preservation below; only user-backed completion authorizes archiving.Build a high-signal learning and career material pipeline for the user.
Use the user's current, explicitly confirmed Decision Context to select the next learning action. Read its current version and distinguish user facts, source evidence, and assistant proposals. Career themes are defaults, not permanent weights: new responsibilities, a decision deadline, or a completed reading can change the order. Keep private context out of public queries and skill text.
Before recommending a first read, check existing Notes, managed material records, and the user's latest read-completion statements. Material importance and remaining reading effort are separate: a central paper already understood may need only one missing experiment, a design delta, or a later revisit. Do not invent a completed/archive transition; use the active lifecycle contract and an explicit disposition.
For user-directed note integration, the named target and source's primary domain take precedence over career mapping. Common discovery lanes include agent runtime/evaluation/memory, model learning, serving and RL systems, product workflows, and career signals; select among them from the current task rather than a fixed percentage split.
For decision-driven X discovery or benchmark reading recommendations, read the focused scout protocol.
This skill is the CS-Notes source-discovery and exact-reading adapter. It owns source recall, primary-source verification, reader maps, domain scoring, and the CS-Notes public/private landing decision.
When the connected project or goal explicitly activates Material Lifecycle, use the
project-local managed loopx-material skill for generic store inventory, candidate/archive
transitions, lossless migration, ranked-entry rebuild, bounded rerank,
owner-gated apply, rollback, and audit. Pass exact-read evidence and the
project-specific score into that workflow; do not duplicate or weaken its
authority and losslessness gates here.
For CS-Notes, a managed authority pointer under
.local/material-lifecycle/authority/current.json means the old candidate and
archive Markdown files are immutable legacy content backing. Never append a
candidate, archive an item, or edit the old Top30 in those files. Use the
project-local loopx-material workflow and the private managed intake adapter.
An explicit 素材: request authorizes candidate intake, a ranking
settlement, and reviewed public-field export for verified public sources
in the same user workflow. Keep intake and ranking as separate
receipts and, when ranking changes, separate authority revisions so CAS,
preview, rollback, and audit remain independent.
For verified public sources, finish version-bound field review and public queue prepare/apply/check in that workflow; do not wait for another request to publish the local queue. Keep original content, private annotations and audit evidence local. Internal or uncertain sources remain private. A public-looking domain alone is insufficient evidence, and public export never bypasses digest or owner-gate checks.
Before recommending a first read or refreshing after Git pull, reconcile explicit
owner read-completion evidence and verified remotely merged completion records
with the managed authority. Archive completed material through the source owner;
retain original sources/backing. If retained as a read reference, label the revisit
purpose and keep it in backlog. Assistant read_scope is not owner completion.
Every intake must end with one explicit disposition: top_window,
ranked_backlog, or no_change. Determine high value from exact-read evidence,
S/A/B tier, current Decision Context, overlap, and artifact convertibility rather
than tier alone. A high-value material must gain verified ranked membership in
the Top30 or ranked backlog; it may not remain only an unranked candidate.
no_change requires a reason and is reserved for lower-value or substantially
duplicative material. If the Top30 is full, move displaced entries into the
ranked backlog without loss. Do not move protected anchors without a changed
Decision Context or explicit owner instruction. When auditing or reshaping the
Top30 against active deliverables, stratify by horizon wave and reading effort
(note-system/materials/material_wave_ranking.py), verify load-bearing primary
citations are settled ahead of secondary commentary, and keep orthogonal trend
items in ranked_backlog rather than crowding active decision waves.
The installed project skill does not itself activate Material Lifecycle. Without an active project- or goal-scoped profile, declared source authority, and current native owner authorization, this skill may research or read material but must not rewrite the managed store. An already authorized ordinary project source does not require a new Goal or a borrowed agent identity.
Prefer primary sources for technical conclusions:
Use social media, 微信公众号, 小红书, X/Twitter, 微博, and aggregators as discovery signals, not final truth.
When extracting mechanisms from a repo, quickstart, prompt, or code file, include file-level links in the user-facing answer and persisted notes/archive. For GitHub sources, prefer commit-pinned permalinks and record the commit read; avoid writing only a repo name or bare filename when a concrete source file drove the claim.
Choose currently callable tools by the question and source type:
Codex remains responsible for verification, ranking, summarization, and local persistence. No specific aggregator is a prerequisite. A publicly listed skill or downloadable client does not establish that its backend is publicly usable; verify service access and dependencies before adopting one. Continue through available sources when an optional tool is absent, and report only coverage gaps that affect the answer.
For Agent Harness / agent infra exploration, use implementation-first catalogs as a source-discovery layer before broad web search when available.
Current primary catalog:
Agent Harness Engineering implementation-first catalog: https://github.com/Picrew/awesome-agent-harnessUse it as an index, not as evidence by itself:
For arXiv papers, do not jump straight to PDF extraction unless HTML is unavailable.
Resolution order:
https://arxiv.org/abs/<id>https://arxiv.org/pdf/<id>https://arxiv.org/html/<id>vNhttps://arxiv.org/html/<id>vN;https://arxiv.org/html/<id>v<version>;https://arxiv.org/html/<id> if versioned HTML is not obvious.精读, method-heavy papers, or cases where HTML lacks needed appendix, math, algorithm, prompt, or caption detail, try the arXiv TeX source before PDF fallback:
https://arxiv.org/src/<id> into .local/paper-cache/<id>-src.tar.gz;.local/paper-cache/<id>-src/;.tex file, usually main.tex or the file containing \documentclass;.tex, .bib, figure captions, tables, algorithm blocks, appendices, and prompt/templates..local/paper-cache/ and explicitly say that the read path was PDF fallback.Why this matters:
请你读 / 精读.请你读 / 精读, always provide the user-facing HTML link when it exists, even if Codex also used the PDF for extraction.For recurring or paper-heavy exploration, use daily-paper-reader as a design reference, not as a dependency to install. Its useful increment is the pipeline shape: intent profiles -> multi-lane recall -> fusion/rerank -> evidence scoring -> deep/quick/carryover selection.
Adopt these patterns:
arXiv, OpenReview, and major venue lanes (NeurIPS, ICLR, ICML, ACL, EMNLP, AAAI) as high-value paper sources.bioRxiv, medRxiv, and ChemRxiv as separate lanes.deep, quick, background, and carryover:
deep: user should personally read or Codex should deep-read next;quick: Codex summary is enough now;background: useful but not active;carryover: high-signal but not yet processed, keep it visible for the next exploration pass instead of letting daily freshness bury it.Do not copy these parts by default: GitHub Actions / Pages deployment, Supabase schema, remote public embedding/rerank services, front-end panels, API keys, or the exact prompt text. Keep the local skill lightweight and source-agnostic.
For paper-heavy exploration, absorb daily-paper-reader's paper data-source coverage rather than its hosted search/deployment stack. The key improvement is to search several paper-source lanes in parallel and then merge them with Codex's existing source stack.
Default paper lanes:
arXiv: fast preprint and recent-paper recall.OpenReview: ICLR / NeurIPS / ICML / AAAI submissions, public reviews, decisions, and withdrawn-public papers when visible.NeurIPS, ICLR, ICML, ACL, EMNLP, AAAI; use official venue pages, OpenReview, ACL Anthology, AAAI/OJS, proceedings pages, or targeted web search as appropriate.bioRxiv, medRxiv, ChemRxiv; use only when the topic is bio / medical / chemistry / scientific-agent adjacent.Combine these with non-paper lanes:
Operational rule:
If this skill has scripts/multi_source_paper_explore.py, use it as the first-pass paper-source recall helper for paper-heavy tasks:
python3 .codex/skills/research-material-scout/scripts/multi_source_paper_explore.py \
--query "<topic>" \
--query "<alternate wording or known title>" \
--sources arxiv,openreview,openalex,biorxiv,medrxiv,chemrxiv,venue-hintsRepeat --query for intent-query expansion. Narrow --sources to topic-fit lanes when domain preprint servers are likely irrelevant.
Use a dual-track exploration flow for real paper-heavy work:
The script is only a recall layer; it complements rather than replaces the older exploration stack.
Use public metadata when sufficient; for incomplete posts, replies, or signed-in search, read the installed ego-browser skill and follow its current TaskSpace and API contract. Do not copy stale browser methods into this skill. Reuse the user's sign-in; inspect visible posts and original links, and close only task-owned surfaces.
Capture relevant author, date, post URL, primary-source links, read scope, and the claim to verify. Avoid unrelated recommendations, DMs, account statistics, or a bulk feed dump. Inspect needed images through supported browser/media tools; do not download remote media to work around display restrictions. Store private evidence under .local/. A social post can be a first-party statement of its author's claim, but reported experimental results still require the paper, code, or data protocol.
Agent-Reach (https://github.com/Panniantong/Agent-Reach) is an optional local scaffolding layer for source-level coverage. Check its installed capabilities before relying on it.
Use it as a design reference or optional install, not the default primary source.
What it adds:
yt-dlp,gh,feedparser,agent-reach doctor style capability diagnostics.Use it when an installed reader adds needed coverage for a specific URL, video, repo, or thread, or when search results are too summarized to verify. Prefer an existing native reader when it already covers the source.
Do not install Agent-Reach automatically unless the user asks. It may install many dependencies and configure cookies/proxies. If installed, keep secrets/cookies local and never commit them.
Directive convention:
素材:<link/text> means intake. Read, classify, summarize, preserve the original link, and append one managed candidate through exact-read evidence, immutable content backing, authority CAS, readback, and receipt.调研:<question/topic> means active research. Use broad recall plus source verification, then write high-signal candidates and recommendations.整理笔记:<link/text> or "整理笔记 + named note/theme" means direct note integration. The output surface is Notes/, not the candidate library by default. Named target and source-domain taxonomy win over career priority.请你读 / 精读: use the reading-only route at the top; do not run intake.读完:<material id/link/title + user notes> means close the reading loop. Update the best Notes/ landing before archiving unless the content is private-only. Treat user-highlighted points as retention requirements: concrete prompts, env flags, schema fields, tool/API names, figures/tables, failure cases, doubts, and comparison phrases should be preserved in the public note when safe and conceptually useful; .local may keep raw/private/full detail, but must not be the only landing for points the user explicitly asked to remember.继续调研 means continue the latest active-research theme, but only if adding new sources or a new decision-relevant synthesis.For each user-provided material:
整理笔记: use the repository note-integration workflow. If the user named a file/section, inspect that target first; if not, classify the source's primary contribution such as model algorithm, training method, inference system, agent runtime, product strategy, or career signal, then find the matching note. Do not default to Agent infra just because the material is AI-related.素材: intake to the current managed material authority. Do not write the legacy candidate Markdown.调研: prioritize by the user's current confirmed Decision Context.请你读 / 精读: use the reading protocol, not the intake steps below.When the user provides a numbered/bulleted readout, quoted phrase, prompt snippet, schema, env var, failure case, or says "这个值得作为专题section / 概念级别 / 这个点要记", treat it as first-class source material rather than optional color.
The note-writing rules below apply to 整理笔记 / 读完. During 请你读 / 精读, cover the user's details in the illustrated conversation answer; do not turn detail preservation into unsolicited note edits.
Notes/, in .local with a privacy/version reason, or intentionally skipped with a reason reported to the user.Notes/: short prompt excerpts or paraphrases, tool names, env flags, schema fields, benchmark names, mode names, key tables/figures, and caveats. Do not over-compress them into only an abstract framework..local..local is for raw cache, private URLs, full prompt copies, and sensitive/internal detail. It supplements public notes; it is not a substitute for durable Notes/ synthesis.wechat-article-readerxiaohongshu-readerlark-doc / lark-wikighego-browser authenticated snapshot / DOM extraction if neededUnread and ask for pasted text, screenshot, export, or accessible copy..local/material-lifecycle/README.md to intake one candidate or revise its existing stable ref. Verify an intake adds exactly one record, or a revision preserves identity, lifecycle and membership, then create the separate Decision Context-backed ranking revision in the same workflow. Verify the disposition, ranked membership when required, authority readback, readable projection, audit receipt, and rollback path before reporting completion.When proactively finding materials:
NeurIPS, ICLR, ICML, ACL, EMNLP, AAAI) when relevant, and add domain preprint lanes (bioRxiv, medRxiv, ChemRxiv) only when the topic warrants them. Then query across at least two source types when possible: paper/code/docs/social.deep / quick / background / carryover; carryover items must have a reason and a next trigger.For discovery/intake (reading-only checks live in the reading protocol):
读完 / note integration, run a user-highlighted point audit: every explicit bullet, numbered item, prompt snippet, schema field, env flag, comparison phrase, or doubt from the user is either present in Notes/, present only in .local with a privacy/version reason, or intentionally skipped with a reason reported.Reject or demote materials that are:
For intake and note edits, report concisely what changed, where, and the key judgment. Reading explanations follow the reading protocol; concise status reporting does not set their depth.
When writing into Notes/:
$$...$$.Notes/AI-Applied-Algorithms/) and link it with a relative Markdown path. Prefer primary-source figures when available.Notes/ when safe: prompt snippets, command/env flags, schema fields, API/tool names, mode names, version boundaries, and failure cases. Prefer short excerpts or paraphrases over long raw prompt dumps, and mark version-sensitive/platform-specific details instead of silently dropping them.For high-value materials, include the next concrete action, such as:
huangruiteng/CS-Notes
feda44a
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.