| name | arxiv-intelligence-reporting |
| description | Summarize arXiv paper intelligence digests for ML, AI, foundation/foundation models, LLMs, post-training, agentic workflows, evals, harness engineering, loop engineering, recursive self-improvement, and self-improving systems. Use when asked to run or interpret the loopcraft arXiv workflow, prepare a daily paper digest, or identify the most interesting recent arXiv submissions. |
arXiv Intelligence Reporting
Workflow
- Work in the loopcraft source tree from the runtime context (the repo with
Makefile, config/, and src/) unless the user gives another workspace.
- Fetch a fresh digest by running the direct CLI:
PYTHONPATH=src python -m loopcraft.research_intel.arxiv.cli run --config config/arxiv_intel.yaml.
You are executing inside a headless loop run, so do not re-enter the control
plane for this loop (no loopctl run / make run of arxiv-intel) — that
recurses. (The control plane also enforces this programmatically: it sets
LOOPCRAFT_ACTIVE_LOOP for the run and refuses re-entry into the same
loop while it is set.) If network access or arXiv API errors occur, report that first.
- Read
state/research/arxiv/latest.md and latest.json from the loopcraft memory ledger (or the emitted direct-CLI paths). Prefer JSON for exact fields and Markdown for human-readable ordering.
- Summarize the 5-10 most interesting papers, focusing on:
- Foundation/frontier models and LLM systems
- Post-training, preference optimization, RLHF, and evals
- Agentic workflows, tool use, planning, memory, and orchestration
- Harnesses, loop engineering, verifiers, and production agent infrastructure
- Recursive self-improvement, self-improving systems, and self-evolving agents
- Include arXiv abstract links and PDF links for every cited paper.
- Mention code, model, dataset, or project links from arXiv comments when present.
- Be clear when a conclusion is inferred from abstract metadata rather than the full paper.
Reporting Standards
- Lead with errors, empty fetches, or suspiciously low paper counts.
- Mention generated date and fetched/ranked counts when available.
- Group findings by theme rather than only by score.
- Prefer papers with concrete systems, benchmarks, training recipes, eval methods, agent loops, or self-improvement mechanisms.
- Avoid presenting broad surveys as the top signal unless they contain unusually useful taxonomy or synthesis.
- End with loopcraft implications: what to read, what to prototype, what to add to eval/harness practice, and what to watch.
Useful Commands
In-loop, always use the direct CLI below; never re-run this loop through the
control plane (that recurses). Operators trigger the loop from a shell via its
documented make/loopctl target — this skill should not.
PYTHONPATH=src python -m loopcraft.research_intel.arxiv.cli run --config config/arxiv_intel.yaml
PYTHONPATH=src python -m loopcraft.research_intel.arxiv.cli run --config config/arxiv_intel.yaml --include-seen
PYTHONPATH=src python -m pytest -q
Expected Output Shape
**Daily arXiv Intelligence**
Fetched N papers; ranked M. Errors: none.
**Most Interesting Papers**
- Title: concise synthesis. arXiv: [...](...). PDF: [...](...). Code/data/model: [...](...). Why it matters: ...
**Themes**
- Theme: papers and shared implication.
**Loopcraft Implications**
- Actionable observation for loop engineering, harnesses, evals, agents, tool use, or self-improving systems.
This shape only describes the human-readable digest. Updating the durable
files (papers.jsonl, seen.json, latest.md/latest.json, and the
run-scoped history/ archives) is enforced by the CLI/store, declared as the
loop's outputs in loops/arxiv-intel.yaml, and checked against
skills/arxiv-intelligence-reporting/verify.md. See that verify file for the
authoritative goal conditions.