research-workflow
Dynamic research workflow management with self-reflection and backtracking
ソース情報
- リポジトリ
- openags/auto-researcher
- ソースの最終更新活動
- 2026年3月20日 18:36
- 検出された SKILL.md の言語
- 英語
- スター
- 9
- フォーク
- 2
インストール方法
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
ソースファイルを確認
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
SKILL.md を表示中
SKILL.md
ソースの指示 · 読み取り専用プレビュー- name
- research-workflow
- description
- Dynamic research workflow management with self-reflection and backtracking
- roles
- ["ags"]
- tools
- ["dispatch_agent","check_progress","ask_user"]
- triggers
- ["research","workflow","pipeline","run project","start research","always"]
- version
- 1.0.0
## Research Workflow Management
When managing a research project, follow this adaptive workflow:
### Stage Progression (typical order, but flexible)
1. **Literature Review** → Understand the field
- Dispatch: `dispatch_agent(role="literature", task="...")`
- Expected output: Review notes in `literature/notes/`, BibTeX in references
- Proceed when: Review covers key related work with cited papers
2. **Research Proposal** → Define the research question
- Dispatch: `dispatch_agent(role="proposer", task="...")`
- Expected output: Proposal document in `proposal/ideas/`
- Proceed when: Clear hypotheses, methodology, and expected outcomes
3. **Experiments** → Validate the hypothesis
- Dispatch: `dispatch_agent(role="experimenter", task="...")`
- Expected output: Code in `experiments/code/`, results in `experiments/results/`
- Proceed when: Code runs successfully and produces meaningful results
- **Common backtrack**: If results don't support hypothesis → re-examine proposal
4. **Manuscript** → Write the paper
- Dispatch: `dispatch_agent(role="writer", task="...")`
- Expected output: LaTeX in `manuscript/main.tex`
- Proceed when: All sections drafted with citations
5. **Peer Review** → Quality check
- Dispatch: `dispatch_agent(role="reviewer", task="...")`
- Expected output: Structured review with scores
- **Common backtrack**: If scores < 6/10 → address specific feedback
### Self-Reflection Protocol
After each agent completes, reflect on:
- **Quality**: Is the output good enough for the next stage?
- **Consistency**: Does it align with previous stages?
- **Completeness**: Are there gaps that need filling?
If issues are found, you have three options:
1. **Fix**: Dispatch the same agent with more specific instructions
2. **Backtrack**: Go to an earlier stage to address root causes
3. **Consult**: Use `ask_user` to get human guidance
GitHubで見る