科研 Skill
智能体资源节约守则
agent-resource-discipline
Applies token- and context-saving habits for tool choice, PDF handling, persistent memory, and privacy checks. Use at the start of sessions with many files or PDFs, or multi-session work.
SKILL.md
Agent Resource Discipline
How this skill is organised (progressive disclosure)
This skill follows the three-level progressive disclosure pattern
codified by Anthropic's skill-creator (see "Adjacent prior art +
lineage" below):
- Level 1 (always in context once the skill is loaded): this
SKILL.md, ~250 lines. Contains the universal rules + the decision-procedure for which references to consult. - Level 2 (loaded on demand by name): the
references/*.mdfiles -- one per discipline. Loaded only when a session actually exercises that discipline. - Level 3 (planned future work): enforcement hooks (SessionStart / PreToolUse / Stop) that mechanise the protocols here so they fire reliably without depending on agent discipline alone. Specification in section "Planned future work: enforcement hooks" below.
This is the same pattern the skill itself preaches: load only what you are about to use; defer the rest.
When to load this skill
Load this skill at the start of any session that will involve any of:
- reading or grepping more than ~5 files;
- handling PDFs (literature survey, related work, supplementary material);
- editing or creating files in multiple project sub-directories;
- web fetching (publisher pages, arXiv, GitHub, doc sites);
- working across multiple agent sessions on the same project (where cross-session memory matters).
In practice that's most non-trivial sessions in this ecosystem. The
universal one-liners in a project-level AGENTS.md (if this
ecosystem provides one, typically its Section 6) cover the basics so
cheap-and-fast rules fire even without this skill loaded; this skill
expands them with the full how-to.
Why this matters
Agent tokens / quota / context-window are the scarcest resources in any non-trivial session. Beyond raw cost, the attention budget -- the agent's ability to pick the right detail out of its context -- degrades faster than the nominal context window suggests. Empirical work (Chroma's "context rot" study, cited by Anthropic in their Sep 2025 Effective Context Engineering for AI Agents post) shows that recall quality drops well before the window fills. Heavily-loaded contexts also introduce recency bias and goal drift. This skill therefore optimises for both raw token cost AND for keeping the working set small enough that the agent's attention stays sharp.
Default agent behaviour wastes resources in predictable ways:
- Tool mis-selection (
bash grepinstead of the dedicatedGreptool,bash catinstead ofRead, ...) costs tokens AND loses features (paging, structured results). - Bulk reads (
Readwith no offset/limit on a 2000-line file when 50 lines would do) burn context window for no gain AND accelerate context rot. - Re-derivation (re-reading a PDF that was already summarised in a survey note last session) wastes tokens AND risks contradicting the prior summary.
- Forgetting (not reading PLAN.md / collection log / notes index at session start) causes the agent to either re-do work or to make decisions inconsistent with prior sessions.
- Re-fetching (calling WebFetch on the same URL twice in one session) burns external quota and adds latency.
- Goal drift in long sessions -- the original PLAN.md fades from recent attention as conversation length grows. Manus calls this the "lost-in-the-middle" failure mode and addresses it via recitation (re-reading the plan into recent context).
These are all preventable with explicit rules. This skill codifies them as a research-flavoured operationalisation of the broader file-as-memory + just-in-time retrieval patterns now standard in the agent-engineering literature.
The six disciplines
This skill loads a small SKILL.md (you are reading it) and provides six per-topic reference files, each loaded on demand. Each codifies one resource-management discipline:
| Discipline | Reference file | When to load |
|---|---|---|
| Tool selection | references/tool-selection.md | First time in this session you need a non-trivial file/search/edit operation. |
| Targeted reads | (covered in tool-selection.md) | (same) |
| PDF lifecycle | references/pdf-lifecycle.md | Whenever a session involves PDF intake or re-reading. |
| Persistent memory | references/persistent-memory.md | Start of any session on a project with PLAN.md / collection log / notes index. |
| Context-window budget | references/context-window-budget.md | When loading multiple skills, multiple reference files, or multiple PDFs simultaneously. |
| Web-fetch discipline | references/web-fetch-discipline.md | Whenever WebFetch is called in this session. |
| Public-surface hygiene | references/public-surface-hygiene.md | Before any commit, issue, PR or comment in a public repository, or any other text published outside the project. |
Load only the references relevant to the current session. Do NOT load all six at once -- that defeats the purpose.
Critical rules (apply unconditionally; do not require loading a reference file)
Rules 1 to 5 may also live in a project-level AGENTS.md
Section 6, so they fire even if this skill is not loaded. Restated here for
in-skill reference:
- Use dedicated tools, not Bash equivalents.
- File search:
Glob(notfind/ls -R). - Content search:
Grep(notbash grep/bash rg). - File read:
Read(notcat/head/tail). - File edit:
Edit(notsed/awk). - File create:
Write(notcat <<EOF/echo >). - User communication: response text (never
echo/printf).
- File search:
- Batch independent tool calls into a single message. A message
with three independent
Reads costs less and finishes faster than three sequential messages. - Read targeted, not bulk. For files >300 lines, use
Grepfirst to locate the relevant section ORReadwith explicitoffset+limit. The default 2000-lineReadis for skimming, not routine consumption. - Re-use prior work before generating new work. Before re-reading
a PDF, check
notes/survey_<citekey>.md. Before re-deriving a fact, check the audit log / notes / PLAN.md. - Indices are the persistent memory. Read
PLAN.mdstatus +_collection_log.md+notes/README.mdat session start; update them at session end if work was done. - Recitation in long sessions. For sessions exceeding ~50 tool
calls, re-read
PLAN.md(or the relevant section thereof) every ~30-50 calls to combat goal drift. The Manus team identified this as the simplest defence against the "lost-in-the-middle" failure mode in long agent runs. Recitation is cheap; goal drift is expensive. - Do not edit
AGENTS.mdor system-prompt-equivalent files mid-session. If the agent client uses prompt caching (Claude Code does, OpenCode does for Claude models), editing the cached prefix invalidates the cache and silently 10x's the per-token cost of all subsequent calls in the session. Restart the session if you genuinely need to change agent-facing rules. - Keep errors in the conversation; do not silently retry. When a
tool call fails (dead URL, rate limit, file not found), let the
error sit in the conversation so the model adapts. Silent retry
loops both burn quota and hide useful failure signal. For
structural failures (a citation's PDF really is unobtainable, an
arXiv ID is wrong), log to the appropriate audit entry
(
_collection_log.md"Items not found / left for user",PLAN.md"Open Questions") so the failure becomes part of the persistent record. - Images are the most expensive thing an agent can look at.
Render small, view few, never re-view. One slide at 150 DPI costs
about 1,800 tokens. Cap renders at about 1568 px on the long edge
(larger is downscaled before the model sees it), view only the pages
that changed, and never re-render an unchanged image. This is not a
reason to stop looking. Measurements and detail:
references/context-window-budget.md"Image budget".
Note on prompt caching
OpenCode (and Claude Code, and Cursor) on Claude models supports
prompt caching of stable prefixes (system prompt + tools +
typically the most recently loaded skill content). Cached tokens are
~10x cheaper than uncached. Implication: re-loading a small skill via
Read mid-session is cheaper than carrying its content forward in
conversation, because the cached version pays cached-rate on every
subsequent turn. This is part of why the progressive-disclosure model
above works: levels 2 + 3 can be loaded fresh when needed without
worrying that they'll dominate cost.
Common rationalizations + rebuttals
The agent will, in real sessions, invent plausible-sounding reasons to skip the disciplines above. The pattern is sufficiently consistent that we name + rebut the common ones explicitly. When the agent catches itself thinking one of these, it should treat that thought as a signal to STOP and re-evaluate.
| Rationalization | Why the agent thinks it | Rebuttal |
|---|---|---|
| "I already read this file last turn; I'll trust my memory." | Avoids the cost of re-Read-ing. | The file might have been edited (by you or the user). Read is cheap; recall is not always reliable. |
"It's just one extra bash cat, no big deal." | The override feels small in isolation. | This is the rationalization that turns a 200-token session into a 20k-token session. One bash-cat is fine; the habit isn't. |
| "Let me re-read the PDF to make sure the survey note is right." | Healthy scepticism + low confidence in your own past summaries. | If you have specific reason to doubt the note, target-grep the .txt for the suspect fact. If not, trust the note; that's what it's for. Re-reading the whole PDF "to be safe" is the most expensive single action in this ecosystem. |
| "I'll load all the section references now so I have them ready." | Tidy-up instinct; wants to "set up" before working. | Loading speculatively is the failure mode the context-window-budget exists to prevent. Load when you actually use. |
| "I'll fetch the publisher page to confirm the year." | Wants external verification; doesn't trust local data. | The user verified the bib entry; that's what verification IS. Trust the bib unless you have specific reason to doubt it. |
"I'll skip updating notes/README.md; it's just an index." | The deposit feels like overhead at the end of a session. | The deposit funds the next session's withdrawal. Skipping it is the most expensive bug in this ecosystem. |
| "I'll process all 14 PDFs now while I have momentum." | Wants to batch-finish a sub-task. | Process one at a time; close each before opening the next. The context-window cost of 14 simultaneous .txt files is much larger than the round-trip cost of 14 separate Reads. |
| "Let me just retry that fetch, it might work this time." | Hope-based rather than evidence-based. | Twice per session is the cap. After that, log to "Items not found" and move on. |
| "I'll silently fix this typo in the bib." | Helpful instinct; wants to clean up. | Silent fixes break the audit trail. Add a "Corrections to apply" entry; let the user batch-apply. |
"pdftotext gave me the title; it's the first line of output." | Extraction succeeded, so the output feels authoritative. | Extraction order is content-stream order, not reading order. On posters/slides the banner is a separate layer and lands anywhere. Extraction proves a string is present, never that it is the title. Render (pdftoppm -png -r 40) and look. See pdf-lifecycle.md Rung 3. |
| "The metadata record must be stale; my extraction says otherwise." | Trusts the artifact over the database, which is usually right. | Not for titles. Records are typed by a human looking at the thing; extraction is a lossy projection of it. Render before you conclude the record is wrong — and before you propagate your version anywhere else. |
| "It's a small task; the protocol overhead would dominate." | Wants to skip first-action / last-action for speed. | A genuinely small task (one file edit, one question answered) is fine. Anything multi-file or multi-step earns the protocol's overhead back several times over. |
If you (the agent) find yourself thinking ANY of the left-column phrases mid-session, stop and re-read this table.
First-action protocol (every non-trivial session)
At the start of any session that touches a project with the standard layout (paper-skeleton or similar):
- Load (in parallel, single message):
AGENTS.md,PLAN.md(status fields + open questions),references/_collection_log.md(verification status),notes/README.md(which surveys exist + their status). Total: 4 small reads. - Decide which skills the session actually needs (research-paper- writing? literature-survey? human-facing-doc-authoring? this skill? often only 1-2 are relevant -- not all of them).
- Decide which references this session needs from each loaded
skill (e.g. just
references/introduction.mdfrom research-paper-writing, not the whole references/ tree). - Then start the user's actual task.
Step 1 is cheap (4 small reads) and prevents the most common waste mode: doing work the previous session already did, or doing work inconsistent with what the previous session decided.
Last-action protocol (every session that produced work)
Before declaring the session done:
- Update the indices that record this session's output:
- new survey notes -> add row to
notes/README.md. - new bibliography entries / verifications -> append to
references/_collection_log.md. - status change -> update the relevant
PLAN.mdstatus field. - new section drafted -> mark in
PLAN.mdoutline + maybe addnotes/section_<N>.md.
- new survey notes -> add row to
- Surface contradictions explicitly. If something this session discovered contradicts prior notes / plan / bib entries, do not silently proceed; add a "Corrections to apply" entry to the relevant log.
- Report to the user what was done + what indices were updated.
Steps 1+2 are the "deposit" that funds the next session's cheap "withdrawal" via the first-action protocol.
Output contract
When this skill is loaded, every action the agent takes should be auditable against the rules above. If the agent finds itself about to:
- run a Bash command that has a dedicated-tool equivalent -> stop and use the dedicated tool.
- do a bulk
Readof a >300-line file -> stop and eitherGrepfirst or useoffset+limit. - re-read a PDF that has a survey note -> stop and read the note first.
- start work without reading
PLAN.md/_collection_log.md/notes/README.md-> stop and read them (in parallel). - finish work without updating those same indices -> stop and update.
The goal is no avoidable waste, not "minimise tokens at the cost of correctness". When the rules conflict with correctness, correctness wins -- and the conflict gets logged as a "Corrections to apply" entry so the rule can be refined.
Tool-availability assumptions
This skill assumes the agent has tools approximately equivalent to
OpenCode's Read, Grep, Glob, Edit, Write, Bash, and
WebFetch. For agents with more limited toolsets:
- Shell-only agents (some Claude Code tool configs): use
pdftotext,rg,fd,sed/awkcarefully (quote everything; prefer here-docs overechochains; cap output withhead/tailEXPLICITLY rather than relying on the agent's truncation). - Agents without WebFetch: load
references/web-fetch-discipline.mdfor the protocol of caching fetches into the repo via shell commands (curl->references/_cache/<hash>.html). - Agents without parallel tool calls: serialise; the parallelism rule simply does not apply, but the targeted-read and re-use-prior-work rules still do.
Adjacent prior art + lineage
This skill is a research-flavoured aggregation of patterns that have crystallised across the agent-engineering literature since mid-2025. Citations are given so users (and future maintainers) know what we borrowed, what we adapted, and where the genuinely novel pieces are.
Foundational sources (cited in the rules above):
- Manus team blog post -- Context Engineering for AI Agents:
Lessons from Building Manus
(
https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus). Source for the file-as-memory pattern (rule 5 of their post), recitation against goal drift (rule 4 -> our Critical Rule 6), KV-cache stability (rule 1 -> our Critical Rule 7), and keep-errors-in-context (rule 5 -> our Critical Rule 8). - Anthropic engineering: Effective Context Engineering for AI
Agents (Sep 2025,
https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents). Source for the "attention budget" framing in "Why this matters", the "context rot" empirical finding (Chroma research,https://research.trychroma.com/context-rot), and the hybrid pre-load + just-in-time retrieval pattern that underlies our first-action protocol. - Anthropic engineering: Equipping agents for the real world with
Agent Skills (Oct 2025,
https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills). Source for the three-level progressive-disclosure pattern this skill follows; informs the SKILL.md length budget guidance inreferences/context-window-budget.md. - Anthropic engineering: Writing effective tools for agents -- with
agents (Sep 2025,
https://www.anthropic.com/engineering/writing-tools-for-agents). Tool-author-side counterpart of our discipline 1; informs why dedicatedRead/Grep/Globare higher-signal than Bash equivalents. - MemGPT paper (Packer et al. 2023, arXiv:2310.08560,
https://arxiv.org/abs/2310.08560). Academic ancestor of all subsequent file-as-memory patterns; introduces the OS-memory-hierarchy framing.
Closest comparable agent-rule skills:
OthmanAdi/planning-with-files(https://github.com/OthmanAdi/planning-with-files) -- 3-file workflow (task_plan.md/findings.md/progress.md) enforced via SessionStart / PreToolUse / Stop hooks. The hook-enforcement pattern is the model for our planned future work below; their evidence-based 96.7% pass-rate (vs 6.7% baseline) is what motivates shipping hooks rather than relying on agent discipline alone.thedotmack/claude-mem(https://github.com/thedotmack/claude-mem) -- automated observation capture into SQLite + Chroma vector DB with lifecycle hooks. We deliberately diverge: our markdown-on-disk approach is git-diffable, blameable, and human-readable, which matters more for research reproducibility than the higher recall of an embedding store. Both approaches are valid; pick by use case.addyosmani/agent-skills(https://github.com/addyosmani/agent-skills) -- 22 production- engineering skills with anti-rationalization tables in every skill; we adopted that pattern explicitly in "Common rationalizations + rebuttals" above.anthropics/skills/skill-creator(https://github.com/anthropics/skills/blob/main/skills/skill-creator/SKILL.md) -- the authoritative how-to-write-a-skill skill; informs our three-level progressive-disclosure structure and the "pushy description" anti-undertriggering convention applied to this skill's own description.Future-House/paper-qa(PaperQA2) (https://github.com/Future-House/paper-qa) -- programmatic RAG over PDFs (parse + cache + embed + retrieve). Solves the same problem as our PDF lifecycle (avoid re-reading PDFs) with a fundamentally different mechanism (vector embeddings vs human- curated survey notes). Ourreferences/pdf-lifecycle.mddiscusses the trade-off explicitly; both approaches are valid.
Where we are genuinely novel (per the prior-art audit, no publicly-available agent-rule skill found that codifies these):
- PDF lifecycle as a per-session protocol for a research project,
with
notes/survey_<citekey>.mdas the agent-readable cache. - Web-fetch discipline grounded in a research-paper context (bib
fields and survey notes consulted before WebFetch; arXiv abstract
preferred over PDF page; cache to
references/_cache/). - The aggregation itself: no other skill we found packages all five disciplines together with a coherent first-action / last-action protocol grounded in a research-paper-skeleton's specific file layout.
Adjacent ecosystems (mentioned for context, not directly borrowed):
- Cline memory bank
(
https://docs.cline.bot/prompting/cline-memory-bank) -- one-task- one-goal + auto-compact +.clineignorepatterns. We have no.agentignoreanalogue yet; possible future addition. - MemGPT / Letta -- programmatic memory backend; conceptually parallel but different abstraction layer.
- LangChain / LlamaIndex memory -- programmatic backends; out of scope for an agent-operation skill.
Planned future work: enforcement hooks
The disciplines above currently rely on agent self-discipline. The single biggest reliability improvement available is to mechanise the first-action / last-action protocols as hooks invoked by the agent client, so the protocol fires regardless of whether the agent remembered to follow it.
OthmanAdi/planning-with-files reports a 96.7% pass-rate on
Anthropic's skill-creator eval after introducing
PreToolUse + PostToolUse + Stop hooks vs 6.7% without. Our
expectation is similar gains for the first/last-action protocols,
because the same failure mode -- agent skipping the bookkeeping when
context fills up -- is what the hooks defend against.
Hooks are deferred for now (kept the repo simple at this stage). When implemented, the design should be:
Hook spec (for future implementation)
Layout (probably bin/hooks/ shipped in this repo, with users
copying or symlinking into their per-project .opencode/hooks/):
bin/hooks/
├── session-start.sh reads AGENTS.md + PLAN.md (status section)
│ + _collection_log.md (Last updated +
│ Corrections-to-apply) + notes/README.md
│ (status section). Echoes a compact
│ summary into the agent's startup context.
├── pre-tool-use.sh on tool calls that risk losing work
│ (Edit, Write, Bash with `git commit`),
│ checks PLAN.md mtime > AGENTS.md mtime,
│ warns if AGENTS.md was edited mid-session
│ (Critical Rule 7 violation).
├── post-tool-use.sh on Write / Edit calls under notes/ or
│ references/, prompts the agent to update
│ notes/README.md or _collection_log.md if
│ the modified file is a new survey note or
│ a new bib entry.
└── stop.sh before declaring session done, checks
that PLAN.md / _collection_log.md /
notes/README.md have been touched if
survey notes / bib entries / experiment
dirs changed during the session. If not,
refuses to stop and prompts for the
last-action update.
Compatibility: the hooks should be CLI-agnostic shell scripts (no agent-client-specific assumptions in the hook bodies); each agent client (OpenCode, Claude Code, Cursor) wires them via its own hook mechanism, but the script bodies are portable.
Implementation order when picked up:
session-start.sh-- highest payoff, lowest implementation risk. Single read-only pass over 4 files; echoes a status summary.stop.sh-- second-highest payoff. Walksgit statusfor changes undernotes//references//experiments/and verifies the matching index was updated. Refuses to stop if not.post-tool-use.sh-- soft prompt; less critical.pre-tool-use.sh-- defensive against Critical Rule 7 violation; only relevant for users iterating heavily on AGENTS.md.
Open questions to resolve at implementation time:
- Where exactly to install for OpenCode (
.opencode/hooks/per repo?~/.config/opencode/hooks/user-global?). Per-repo is more correct but requires users to copy explicitly. - How to make the prompts actionable rather than annoying (the user should never feel the hook is in the way of legitimate work).
- Whether to log hook fires to
~/.scicomp-research-skills.hooks.logfor debugging (probably yes, very small). - Compatibility testing matrix: OpenCode + Claude Code + Cursor at least.
When to revisit: when at least 3 real research-paper sessions have shown the agent skipping the first-action or last-action protocol despite the rules being loaded. Until then, the cost of maintaining hooks across multiple agent clients exceeds the benefit.
See also
references/tool-selection.md-- dedicated-tools-vs-bash + targeted read rules + parallelism rules.references/pdf-lifecycle.md-- one-shot pdftotext extraction; survey-note-first lookup; section-targeted reads of.txtfiles.references/persistent-memory.md-- first-action / last-action protocols for cross-session memory; the indices as memory.references/context-window-budget.md-- max-N skills / references / PDFs simultaneously; when to summarise + close.references/web-fetch-discipline.md-- cache-first; bib-fields before publisher page; arXiv abstract over PDF.references/public-surface-hygiene.md-- privacy on public surfaces: generic provenance, fictional examples, thebin/privacy-scan.shscanner and its personal terms file, peer-review material, cleanup of already-published content.- The universal one-liners in a project-level
AGENTS.mdSection 6 (if present) are a strict subset of the rules above; that section is what fires for agents that have not loaded this skill.
Prerequisites and Limitations
- This skill was adapted from the upstream
scicomp-research-skillsframework. References tobin/privacy-scan.sh,.githooks/, a CI privacy-lint workflow, or a rootAGENTS.md/~/.scicomp-research-skills/path describe that framework's own repository; they are not bundled with this skill and will not exist in a typical project. Treat them as a design pattern to adapt (e.g. write an equivalent pre-commit check) rather than as paths that already exist. - The sibling skills named in this file (e.g.
agent-resource-discipline/references/pdf-lifecycle.md) are this skill's own bundled references; cross-references to other skills by name (e.g. "the research-paper-writing skill") assume that skill is installed in the same agent/marketplace -- confirm it is available before relying on it.
Created 2026-05-13 by A. Attia. Distilled from observed waste modes
across multiple agent sessions on a research paper + this skills
repository's own bootstrapping. Revised 2026-05-13 (post-prior-art
audit): adopted progressive-disclosure framing + pushy description
(Anthropic skill-creator); added "Why this matters" framing around
attention-budget + context-rot (Anthropic context-engineering post +
Chroma); added Critical Rules 6-8 (recitation, no-mid-session-prompt-
edits, keep-errors-in-conversation) from Manus; added prompt-caching
note; added "Common rationalizations + rebuttals" table
(addyosmani/agent-skills pattern); added "Adjacent prior art +
lineage" section citing Manus / Anthropic / planning-with-files /
claude-mem / addyosmani / paper-qa / MemGPT; added "Planned future
work: enforcement hooks" with full spec (deferred implementation,
specification kept here so future work has the design ready). Revised
2026-09-27: added two rows to "Common rationalizations + rebuttals"
covering PDF text-extraction order vs visual order, and the prior that
an external metadata record beats your own extraction on questions of
title; full treatment in references/pdf-lifecycle.md Rung 3.
Revised 2026-10-14: added the sixth discipline, public-surface hygiene
(references/public-surface-hygiene.md), backing root AGENTS.md
Section 6.7; added critical rule 9, image cost discipline, from a
slide-review session measured in 2026-10 (detail in
references/context-window-budget.md "Image budget").
示例提示词
这次任务要读很多PDF和文件,帮我先规划好省token的工作流程。
所属插件
安装该插件会一并获得同领域的 54 个 Skill / MCP。
同插件的其他科研 Skill
- 英文论文双语精读器把英文论文转成中英对照精读档并生成离线双语阅读器,标注要点并输出读书笔记。精读单篇英文论文、要中英对照阅读时使用。
- 研究方法比较论证Compares quantitative, qualitative, and mixed research methods against what each can and cannot answer to build a methods table. Use when justifying a method choice in a thesis.
- 复杂系统分析方法Analyzes complex systems with power-law fitting, Hurst exponent, fractal dimension, agent-based modeling, and entropy measures. Use for nonlinear or complex-system data.
- 研究环境容器化部署Builds Docker and Singularity containers for reproducible computation and HPC deployment of research environments. Use when packaging an analysis pipeline for reuse or HPC.
- DVC数据版本管理Sets up DVC for dataset and pipeline versioning, remote storage, and experiment comparison in reproducible ML research workflows. Use when tracking large datasets or ML experiment runs.
- 期刊级统计图表绘制Creates publication-quality multi-panel figures with matplotlib, seaborn, or ggplot2 using colorblind-safe palettes and journal export settings. Use when preparing figures for a paper.
- 综述模块拆解规划Breaks a refined literature-review question into 3-5 non-overlapping modules to build a reading map, outline, or search plan. Use when structuring a literature review.
- 学位论文答辩教练模拟学位口试委员提问并提供承认回应证据三段式答辩框架与模拟问答。准备论文口试或研讨会问答环节时使用。
- 实验设计与样本量规划Designs rigorous experiments covering sample-size calculation, randomization strategy, pre-registration, and A/B test duration estimation. Use when planning an experiment or A/B test.
- 管理实验设计架构为管理与商学情境实验规划组间组内设计、对抗平衡与操弄检核,并产出设计书与情境材料模板。做商学实验设计时使用。
- 科研经费申请撰写Structures research grant proposals including NSF/NIH Specific Aims, budget justification, ERC narrative sections, and biosketch formatting. Use when drafting a funding application.
- 人类可读文档撰写规范Guides writing human-facing docs like README, PLAN, survey notes, and reviewer-response drafts with a clear two-tier structure. Use for text a human collaborator will read.
- 知识图谱与SPARQL检索Builds knowledge graphs and runs SPARQL queries over RDF and Wikidata for entity linking in research and library data. Use when modeling research entities as a linked graph.
- LaTeX论文写作与排版Writes, compiles, and submits LaTeX papers covering IMRaD structure, key packages, bibliography management, and arXiv preparation. Use when authoring or debugging a LaTeX manuscript.
- 文献比较矩阵构建器把一批PDF整理成含撤稿查核与APA7书目的文献资料库,并生成横向比较矩阵表。整理综述前要比较文献时使用。
- 跨库学术文献检索Searches OpenAlex, Semantic Scholar, and arXiv for academic literature with deduplication, citation snowballing, and BibTeX export. Use when collecting papers for a review or topic.
- 文献调研语料库维护Finds, verifies, and maintains a literature corpus from arXiv, OpenAlex, Crossref, and related databases with BibTeX entries and per-paper notes. Use when sourcing references for a paper.
- 文献综述研究缺口定位Converts a literature-review synthesis into a concrete, defensible research gap and future-research direction. Use when writing a conclusion or thesis contribution section.
- 文献矛盾结论梳理Analyzes conflicting findings across papers and explains disagreement by sample, method, or context differences for a stronger synthesis. Use when review sources disagree.
- 效应量汇总元分析Pools effect sizes across studies with fixed and random-effects models, heterogeneity tests, forest and funnel plots, and subgroup analysis. Use to synthesize quantitative study results.
- 缺失数据处理Diagnoses missing-data mechanisms such as MCAR, MAR, or MNAR and applies multiple imputation or deletion with sensitivity analysis. Use when a dataset has missing values.
- 混合方法研究设计Designs mixed-methods studies integrating qualitative coding, triangulation, content analysis, and QCA with inter-rater reliability checks. Use when combining qual and quant data.
- 科研机器学习流程Applies machine learning to research questions with scikit-learn pipelines, AutoML, SHAP explainability, and nested cross-validation. Use when building a predictive model for a study.
- 开放数据存储与FAIRDeposits and documents research data via Zenodo or OSF with FAIR assessment, data management plans, and DataCite metadata. Use when publishing or sharing a research dataset.
常见问题
智能体资源节约守则能做什么?
Applies token- and context-saving habits for tool choice, PDF handling, persistent memory, and privacy checks. Use at the start of sessions with many files or PDFs, or multi-session work。
如何安装智能体资源节约守则?
在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「智能体资源节约守则」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。
智能体资源节约守则需要哪些权限?
该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。
智能体资源节约守则经过官方审核吗?
是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。