Agent Skills
› NeverSight/learn-skills.dev
› reference-project-synthesizer
reference-project-synthesizer
GitHub将结构化参考源卡片转化为项目决策。通过读取源卡片和项目记忆，分析其对主张、实验、风险及实施的影响，生成持久化结论并链接至对应记忆对象，辅助ML项目的规范化管理。
Trigger Scenarios
需要评估参考文献对当前ML项目主张或实验设计的影响
基于已生成的源卡片更新项目记忆中的风险、基准或实施计划
Install
npx skills add NeverSight/learn-skills.dev --skill reference-project-synthesizer -g -y
SKILL.md
Frontmatter
{
"name": "reference-project-synthesizer",
"description": "Connect structured reference source cards to the active ML project. Use when papers, collaborator docs, Markdown notes, specs, scripts, BibTeX files, or source bundles should inform claims, risks, baselines, benchmarks, experiments, algorithm design, implementation, writing contracts, citations, collaborator actions, project initialization, or memory writeback.",
"allowed-tools": "Read, Write, Edit, Bash, Glob",
"argument-hint": "[project-root] [source-card-or-source-id] [--link-claims] [--benchmark] [--baseline] [--writing] [--risk] [--feedback] [--spec] [--implementation] [--seed-project]"
}
Reference Project Synthesizer
Convert source cards into project decisions. This skill answers: What does this source mean for our project?
Use after reference-reading-summarizer has produced a source card. Do not repeatedly re-read raw sources unless the card is insufficient or low confidence.
Skill Directory Layout
<installed-skill-dir>/
├── SKILL.md
├── references/
│ └── project-link-routing.md
└── templates/
└── project-use-note.md
Core Contract
- Read project memory before linking a source to claims, experiments, risks, implementation, collaborator actions, or writing.
- Use source cards as compressed inputs. Raw source files and raw reading trajectories stay out of project memory.
- Stronger reasoning is required when the source changes a claim, baseline, benchmark, novelty boundary, implementation plan, collaborator commitment, or project seed.
- Every project implication should link to a source card, confidence level, and target memory object.
- Do not promote a cheap skim into a project decision without marking uncertainty or requesting deeper reading.
Inputs
Prefer:
reference/cards/<source-id>.mdreference/.agent/source-index.mdreference/.agent/processing-status.md- backward-compatible
reference/.agent/reference-index.mdandreading-status.md memory/current-status.mdmemory/claim-board.mdmemory/evidence-board.mdmemory/risk-board.mdmemory/action-board.mdpaper/.agent/writing-contract.mdcode/.agent/benchmark-plan.mdor equivalent, when present
Model Routing
Read references/project-link-routing.md.
- Tier 2 normal main model: routine project-use notes, citation placement, writing-pattern routing, structured feedback actions.
- Tier 3 strong/deep model: closest-work risk, must-have baseline, benchmark protocol, core algorithm/theory impact, project seed, conflicted collaborator feedback, final positioning.
- Do not use cheap sidecar output as the final authority for project decisions.
Workflow
- Locate the project root and existing memory.
- Read the relevant source card and its confidence/provenance.
- Identify requested synthesis mode:
- claim support or challenge
- baseline implication
- benchmark/task/protocol implication
- method/theory implication
- implementation implication
- writing pattern
- citation placement
- collaborator feedback/action
- project spec/constraint
- project seed
- reviewer-risk or novelty boundary
- Create or update
reference/project-use/<source-id>.mdfromtemplates/project-use-note.md. - Write only durable conclusions into project memory — link-mode determines target:
- claim support/challenge →
memory/claim-board.md+memory/evidence-board.md - identified risk or blocker →
memory/risk-board.md+memory/action-board.md - baseline or benchmark → code-side
docs/experiments/ormemory/evidence-board.md - implementation hint → code-side design doc or
memory/decision-log.md - writing exemplar or style pattern →
paper/.agent/writing-memory/viapaper-writing-memory-manager - reviewer-risk or novelty framing →
memory/risk-board.mdorpaper/.agent/paper-evidence-board.md - provenance link →
memory/provenance-board.md
- claim support/challenge →
- If the card is too weak, route back to
reference-reading-summarizerwith the required reading mode.
Output Rules
Each project-use note should include:
- one-line project relevance
- supported or challenged claims
- baseline, benchmark, implementation, or spec implications
- algorithm/theory implications
- writing/citation implications
- collaborator actions if relevant
- risks and actions
- confidence and required follow-up
End with the closure audit:
Goal: What project question did this source help answer?
Answer: What did we learn?
Path: Which card, sections/files, and memory objects support this?
Correctness: Why should we trust it, and what needs deeper reading?
Next: What changes in the project?
Version History
- e0220ca Current 2026-07-05 21:36
