科研 Skill

研究计划任务执行

executing-research-plan

Executes a ready research plan task by task, gating acquisition before ingest and dispatching analysis. Use after writing-research-plan to run the planned work.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Executing a Research Plan

Run a pre-registered plan task by task. Dispatch the right subagent per task, review each result (spec + quality), and only mark done when both reviews pass. Deviations are logged, not hidden.

Announce at start: "Using executing-research-plan to run <slug>-plan.md."

If unmet: explain to the user which condition is missing, ask for a short reason, write it to knowledge/_meta/gate-overrides.log, and start the task.

A task is "done" only when the type-specific review passes (see "Review by methodology" below). Never silently rewrite a hypothesis mid-run — deviations go into the deviation log.

Checklist

  1. Load the plan — read input/ideas/<slug>-plan.md end-to-end
  2. Verify the plan is ready (methodology-aware): methodology: hermeneutic → frontmatter status: ready is enough (no frozen hypothesis); methodology: quantitative/mixed (or a task block marked pre-registered: true) → status: pre-registered + user confirmation of the hypothesis
  3. Create task-list items — one per task in the plan (use the plan's own wording)
  4. Personal-data check (before any task that reads input/data/ or input/notes/) — interviews, questionnaires, field notes or excavation diaries naming living people, correspondence, unpublished data shared by colleagues. Ask once: does this material contain personal or confidential data, and do the consent forms / data agreements allow processing by the AI provider this session uses? If not, or unknown: stop that task and say so — the options are anonymising first (remove names, places, identifiers before the file is read), or running the task with a local / institution-hosted model (e.g. KISSKI/GWDG, a local Ollama or LM Studio model). Record the answer in the plan's task block (personal_data: none | anonymised | local-model-only) so the next session does not ask again. Published scholarship and the user's own non-personal notes need no check
  5. Acquisition gate (before any ingest task) — guarantee the originals are on disk first; run the Acquisition gate algorithm below. literature-review only searches — without this gate every source-ingester hard-stops on a missing PDF.
  6. Route each task to the right subagent:
    • Ingest task → source-ingester (see agents/source-ingester.md)
    • Analysis task → analyst (runs Python/R in output/data-analysis/)
    • Synthesis task → main conversation, not subagent (high context integration) — follow Writing a synthesis
    • Draft task → invoke drafting-manuscript skill (which may dispatch drafter)
  7. Two-stage review per task:
    • Spec reviewer (fresh subagent): does the output match the task spec exactly?
    • Quality reviewer (fresh subagent): is the code/page/result methodologically sound?
    • Any "no" → route back to the executing subagent with reviewer feedback
  8. Log deviations — if findings force change, append to knowledge/_meta/log.md and mark subsequent analyses as exploratory
  9. Run wiki-lint after every ingest and synthesis task
  10. After all tasks done: check plan's Verification section; only green on all boxes means complete
  11. Transition: if plan output target is book/article → drafting-manuscript; if grant → grant-finder; if done → finishing-a-research-project

Acquisition gate (before any ingest task)

literature-review searches but downloads nothing; source-ingester hard-stops on a missing original. This gate sits between them: it guarantees every ingest task's PDF is on disk (or explicitly deferred) before the first source-ingester is dispatched — so you get one clean acquisition step instead of a wall of per-source hard-stops.

  1. Collect required originals — from every ingest task in the plan, derive the expected library path <library>/pdf/<bibkey>.pdf (the naming convention acquire-sources writes and ingest-source reads).
  2. Scan the library — <library>/pdf/*.pdf (resolve with scripts/library.py; never in subfolders). If every required original is present → skip the gate, go to routing.
  3. Attempt acquisition — run acquire-sources on the A+B set (dispatch the source-acquirer subagent for ≥ ~8 missing items). It auto-downloads the Open-Access copies into the library and (re)generates input/bibliography/acquisition-todo.md.
  4. If acquisition-todo.md is non-empty (originals still missing), enter the interactive resume loop — do NOT silently skip:
    • a. Offer the list. Present the missing sources from acquisition-todo.md plus the manual-download instruction (university VPN / library proxy; save each into <library>/pdf/ under the exact <bibkey>.pdf filename). Mark the dependent ingest tasks as blocked and pause — do not dispatch those ingests.
    • b. On the user's resume signal, re-scan <library>/pdf/ for the newly added PDFs and reconcile — re-run acquire-sources in reconcile mode: resolved rows drop out of (are checked off) acquisition-todo.md.
    • c. Ask the branch: "Search for alternatives (open-access substitutes via literature-review) for what's still missing, or continue with what's present?"
      • Alternatives → loop back to literature-review / literature-scout to find OA alternatives for the gaps, grade them into the guide table, then re-run acquire-sources on the new candidates → back to step 4.
      • Continue → proceed to routing with the acquired subset.
    • d. Loop a–c until the user says to move on to ingest.
  5. Proceed to routing — dispatch source-ingester only for tasks whose original is now on disk; any still-missing stay blocked/deferred per the user's choice. Non-blocked downstream work (analysis on existing data, etc.) is not held up by a blocked ingest.

"Alternativen" = a different open-access source on the same topic, NOT a preprint/prior-version/review substitute for the same paper. Substituting one edition for another remains ingest-source's consent-gated decision and is out of scope for this gate.

Process Flow

digraph executing {
    "Load plan" [shape=box];
    "Pre-registered & signed?" [shape=diamond];
    "Back to writing-research-plan" [shape=box];
    "Create task list from plan" [shape=box];
    "Acquisition gate: scan the library" [shape=box];
    "All originals present?" [shape=diamond];
    "Run acquire-sources" [shape=box];
    "Worklist empty?" [shape=diamond];
    "Offer manual list, block ingests, pause" [shape=box];
    "User resumes" [shape=box];
    "Reconcile (re-scan + re-run)" [shape=box];
    "Alternatives or continue?" [shape=diamond];
    "literature-review for OA alternatives" [shape=box];
    "Next task" [shape=box];
    "Route by task type" [shape=diamond];
    "Dispatch source-ingester" [shape=box];
    "Dispatch analyst" [shape=box];
    "Handle synthesis inline" [shape=box];
    "Invoke drafting-manuscript" [shape=box];
    "Spec review" [shape=box];
    "Quality review" [shape=box];
    "Both pass?" [shape=diamond];
    "Route back with feedback" [shape=box];
    "Mark task done" [shape=box];
    "More tasks?" [shape=diamond];
    "Verify against plan checklist" [shape=box];
    "All green?" [shape=diamond];
    "Transition" [shape=doublecircle];

    "Load plan" -> "Pre-registered & signed?";
    "Pre-registered & signed?" -> "Back to writing-research-plan" [label="no"];
    "Pre-registered & signed?" -> "Create task list from plan" [label="yes"];
    "Create task list from plan" -> "Acquisition gate: scan the library";
    "Acquisition gate: scan the library" -> "All originals present?";
    "All originals present?" -> "Next task" [label="yes"];
    "All originals present?" -> "Run acquire-sources" [label="no"];
    "Run acquire-sources" -> "Worklist empty?";
    "Worklist empty?" -> "Next task" [label="yes"];
    "Worklist empty?" -> "Offer manual list, block ingests, pause" [label="no"];
    "Offer manual list, block ingests, pause" -> "User resumes";
    "User resumes" -> "Reconcile (re-scan + re-run)";
    "Reconcile (re-scan + re-run)" -> "Alternatives or continue?";
    "Alternatives or continue?" -> "literature-review for OA alternatives" [label="alternatives"];
    "literature-review for OA alternatives" -> "Run acquire-sources";
    "Alternatives or continue?" -> "Next task" [label="continue"];
    "Next task" -> "Route by task type";
    "Route by task type" -> "Dispatch source-ingester" [label="ingest"];
    "Route by task type" -> "Dispatch analyst" [label="analysis"];
    "Route by task type" -> "Handle synthesis inline" [label="synthesis"];
    "Route by task type" -> "Invoke drafting-manuscript" [label="draft"];
    "Dispatch source-ingester" -> "Spec review";
    "Dispatch analyst" -> "Spec review";
    "Handle synthesis inline" -> "Spec review";
    "Invoke drafting-manuscript" -> "Spec review";
    "Spec review" -> "Quality review";
    "Quality review" -> "Both pass?";
    "Both pass?" -> "Route back with feedback" [label="no"];
    "Route back with feedback" -> "Next task";
    "Both pass?" -> "Mark task done" [label="yes"];
    "Mark task done" -> "More tasks?";
    "More tasks?" -> "Next task" [label="yes"];
    "More tasks?" -> "Verify against plan checklist" [label="no"];
    "Verify against plan checklist" -> "All green?";
    "All green?" -> "Next task" [label="no, fix"];
    "All green?" -> "Transition" [label="yes"];
}

Task Type Routing

Plan keywordSubagent / SkillOutput landing zone
"Acquire / fetch PDFs / besorge Quellen"acquire-sources skill (→ source-acquirer for ≥8)<library>/pdf/*.pdf, input/bibliography/acquisition-todo.md
"Ingest X"source-ingesterknowledge/sources/, knowledge/entities/, output/bibtex/
"Analyze / compute / run"analystoutput/data-analysis/, output/data-analysis/results/
"Synthesize / integrate / write synthesis"inline (main conversation)knowledge/synthesis/
"Draft chapter / section / article"drafting-manuscript skilloutput/**/*.md

Writing a synthesis

A synthesis is where the project says what its sources add up to, and it is where model-written text is weakest. The typical failure is not a false statement but an empty one: "a complex picture emerges", "the evidence must be assessed in a differentiated way", "cult centralisation played an important role", "further research is needed". Each sentence is true of almost any topic, so it says nothing about this one. Such sentences are the model's safe default when it integrates many sources at once: the averaged position, hedged on all sides. The rules below make the specific conclusion the easier one to write.

  1. Re-read before concluding. Write the conclusions from the source pages (and their focus blocks) re-read now — not from the ingest summaries still in the conversation, which are compressed already. A synthesis written at the end of a long ingest run is a summary of summaries.

  2. The author's conclusion first. Before writing the conclusions, ask the user what they think the material shows, in their own words (rough is fine) — the same discipline as drafting-manuscript STOP 0, for the same reason: approving a polished machine conclusion is not drawing one. Do not offer a candidate in that message. If the user asks you to propose, mark the conclusion agent-proposed in the page until they restate it.

  3. Each conclusion is a claim block, not a paragraph of summary:

    ### C1 — <short label>
    - **Claim:** <one sentence that could turn out wrong>
    - **Kind:** finding (what the sources report) | reading (an interpretation in
      the literature — whose) | own thesis (the project's inference — marked as such)
    - **Carried by:** [[source-a]] (p. 61), [[source-b]] (p. 12–14) — n independent sources
    - **Concrete anchor:** <the named site / passage / stratum / dating / dataset
      the claim is about>
    - **Against:** <the strongest counter-position, cited — or "none found in the wiki">
    - **Would be wrong if:** <the finding or argument that would overturn it>
    

    A conclusion with no concrete anchor, or nothing that could make it wrong, is not finished: narrow it until it has both, or move it to What stays open.

  4. Specificity test. For every sentence in the conclusions, swap the topic words for a neighbouring topic (another site, another text, another period). If the sentence is still true, it says nothing here — cut it or make it specific. Typical offenders, in both languages: komplexes / vielschichtiges Bild, differenziert zu betrachten, spielt eine wichtige / zentrale Rolle, im Spannungsfeld von, weitere Forschung ist nötig, a nuanced picture emerges, plays a crucial role, highlights the importance of, multifaceted, it remains to be seen.

  5. Disagreement is a finding, not a conclusion. "Scholars disagree" says something about the literature. Say who holds what, on which evidence, since when (see the research chronology) — and what the project concludes from that: one side, a third reading, or an explicit bracket with the reason.

  6. Open questions are specific too. What stays open names the question and what would settle it (a source not yet read, an excavation report, a test) — not "further research is needed".

  7. Keep the hedges the sources give, and add none of your own. A conclusion is as strong as its sources carry; the hedge goes on the claim where the evidence is thin, not across the whole page as insurance.

Critical Thinking — Cross-Cutting Checklist

Walk through this on every method choice and before every subagent dispatch:

  1. Claim first, then evidence — what exactly is being claimed? State the claim precisely, then ask what evidence would support it.
  2. Evidence type — primary data, secondary synthesis, expert opinion, theoretical argument. Drives the method choice.
  3. Framework matched to discipline:
    • GRADE — causal / quantitative claims
    • Cochrane ROB — experimental designs
    • Source-critical (Quellenkritik) — textual-historical sources
    • Stratigraphic reliability — archaeological claims
    • Reproducibility audit — computational / DH
  4. Name confounders explicitly — what gets overlooked, what could explain the finding otherwise
  5. Check logical fallacies — ad hominem, affirming the consequent, equivocation, cherry-picking, post hoc
  6. Falsifiability — what would refute the claim?

If the answer to any of these is "I don't know," that belongs in the quality review or gets resolved before dispatch.

Review by Methodology

The review tiering depends on the methodology field in the project frontmatter (CLAUDE.md) or in the task-block frontmatter.

Quantitative Tasks — Two-Stage Review

For tasks under methodology: quantitative or task blocks with pre-registered: true: dispatch two fresh subagents (minimal context: just task spec + output):

Spec reviewer prompt template:

You are a spec compliance reviewer. Task spec: . Output: . Does the output match the spec exactly? Report mismatches. Do not assess quality.

Quality reviewer prompt template:

You are a methodological quality reviewer. Output: . Methodological context: . Is this sound? Check: (1) are assumptions stated, (2) are limitations named, (3) does the script contain control checks with known answers, and do they pass, (4) does the load-bearing result agree across the independent recomputation, (5) is each procedure's choice justified by the plan rather than by its result, (6) would a peer reviewer accept this? Report issues.

Both "yes" → task done. Any "no" → feedback back to the executing subagent.

Hermeneutic Tasks — Synthesis Review

For tasks under methodology: hermeneutic (no pre-registration): single-stage review in the main context, focused on plausibility and source fidelity rather than spec compliance.

Synthesis reviewer prompt template (in main context, no subagent needed):

Output: . Method sketch: . (1) Is the argument actually supported by the cited sources (spot-check 3 central citations)? (2) Are alternative readings named where they suggest themselves? (3) Is the hermeneutic position transparent (which pre-understanding horizon, which method is applied)? (4) Is every conclusion specific? Apply the specificity test from "Writing a synthesis" to each sentence of the conclusions and quote the ones that would hold for a neighbouring topic just as well. (5) Is every conclusion a complete claim block — concrete anchor, carrying sources with pages, kind (finding / reading / own thesis), a counter-position, and what would make it wrong?

Return findings as correction notes, not as a pass/fail gate. The user decides final acceptance.

Mixed Projects

The plan marks per task which review mode applies. Pre-registered tasks run through two-stage; hermeneutic tasks run through synthesis review.

Deviations

If a task surfaces a result that contradicts the pre-registered hypothesis, or if the method must change:

  1. Stop. Do not rewrite the hypothesis.
  2. Append to knowledge/_meta/log.md:
    - YYYY-MM-DD · deviation · plan=<slug> · <what changed and why>
    
  3. Mark any analysis downstream of the deviation as exploratory in its frontmatter (status: exploratory).
  4. Surface to user — "The plan said X, the data shows Y. Continuing as exploratory. OK?"

Red Flags

ThoughtReality
"Just start ingesting — the PDFs are probably there"No — run the acquisition gate first; otherwise every source-ingester hard-stops on a missing original, one source at a time.
"Some originals are still missing, I'll just skip them quietly"No — surface the worklist, block those ingests, and let the user fetch them or pick alternatives. Silent skipping hides gaps.
"The hypothesis doesn't fit, I'll tweak it slightly"No — log entry, result becomes exploratory.
"Review is overkill for such a small task"Small tasks are exactly where errors slip past. Always both reviews.
"I'll just do it myself instead of dispatching a subagent"Subagent isolation protects context and forces prompt clarity.
"The synthesis is stable now"Only the user sets stable.
"A complex picture emerges — that's an honest conclusion"It is true of every topic. Name who holds what, on which evidence, and what the project concludes.
"I just ingested all the sources, I can write the synthesis from that"You have their summaries in context, already compressed. Re-read the source pages before concluding.
"I'll hedge the conclusions to be safe"Hedge the claim whose evidence is thin, with the reason. A blanket hedge makes every conclusion unfalsifiable.

Key Principles

  • Acquire before ingest — the acquisition gate guarantees originals are on disk (or explicitly deferred) before any source-ingester runs
  • The plan is law — anything outside the plan is scope creep or deviation
  • One subagent per task — fresh context, no context collapse
  • Two-stage review — spec THEN quality, not mixed
  • Deviation > revision — divergent findings get flagged, not absorbed
  • Verification is a list, not a feeling — the plan's checkboxes are the completion test

示例提示词

帮我按已生成的研究计划逐项执行任务,先获取原始文献再入库。

所属插件

人文学科面向历史、哲学、语言学、考古与艺术研究:TEI/OCR 与数字档案、论证分析、语料库与语音学、放射性碳测年、乐谱与音频分析,以及人文论证写作与审稿。

安装该插件会一并获得同领域的 62 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

研究计划任务执行能做什么?

Executes a ready research plan task by task, gating acquisition before ingest and dispatching analysis. Use after writing-research-plan to run the planned work。

如何安装研究计划任务执行?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「研究计划任务执行」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。

研究计划任务执行需要哪些权限?

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。

研究计划任务执行经过官方审核吗?

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。