Agent Skills
› brycewang-stanford/Awesome-Journal-Skills
› ipsn-writing-style
ipsn-writing-style
GitHub用于修订IPSN论文，确保首段呈现真实感知问题，强调硬件实测与基线对比、能量/延迟预算及诚实部署报告。遵循双盲原则和12页限制，提供写作骨架与句式改写指导，防止脱离实际的仿真或描述性陈述。
触发场景
修订传感器系统论文
检查论文是否符合IPSN风格
优化论文结构与证据链
安装
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill ipsn-writing-style -g -y
SKILL.md
Frontmatter
{
"name": "ipsn-writing-style",
"description": "Use when revising an IPSN-lineage paper for a real sensing problem on the first page, a load-bearing information-processing or platform contribution, ground-truth methodology, energy\/latency\/accuracy budgets, honest deployment reporting, double-blind wording, and disciplined use of the ≤12-page ACM two-column budget."
}
IPSN Writing Style
Use this when revising the main paper. IPSN papers are read by sensor-systems and information-processing reviewers, so they need a real sensing problem on the first page and evidence measured on real hardware against ground truth. The failure this skill prevents is a technically fine paper that reads like an offline ML result, or a hardware report with no measured trade-off.
Revision rules
- Lead with the sensing problem and its physical budget: the phenomenon being sensed, why current sensing/processing is inadequate, the contribution (an estimator/inference method and/or a platform), the real-hardware evidence, and the energy/latency/accuracy budget that makes it deployable.
- Make the information processing load-bearing (IP track). State the estimator, inference, or learning method precisely enough to re-implement, and tie it to a quantity that matters (error bound, joules, bits, latency) — not just "accuracy improved."
- Make the platform reusable and measured (SPOTS track). Give the MCU/SoC, sensor, radio, and toolchain, and justify design choices with measured power, timing, and robustness — a datasheet is not a contribution.
- Pair every claim with proportional, real-hardware evidence — a measured energy number, a ground-truth-referenced accuracy, a deployment yield — not adjectives.
- Report deployments honestly. Yield, synchronization error, packet loss, and energy as measured; an idealized deployment reads as a tell.
- Respect the ≤12-page ACM two-column budget as a design constraint; a paper that only fits by cutting the measurement setup or the limits is over-scoped.
- Maintain double-blind in self-citations, tool/testbed names, board photos, acknowledgements, funding, and dataset links.
Sensor-systems paper skeleton
| Section | Job it must do | Common failure |
|---|---|---|
| Intro | Sensing problem, inadequacy, contribution, real-hardware evidence preview, budget — first page | Leads with a model/technology trend, not a sensing problem |
| Background | The modality, the platform constraints, why this is hard physically | Motivation by assertion; no physical grounding |
| Method / System | The estimator/inference (IP) or the platform/tool (SPOTS), reproducibly | Method too thin to re-implement; platform under-specified |
| Evaluation | Each claim answered with measured, ground-truth evidence | Simulation or offline accuracy standing in for on-device/in-field results |
| Energy/latency budget | Joules, memory, timing as measured on the real platform | "Feasible on embedded devices" with no numbers |
| Limits / threats | Site-specificity, generalization, calibration drift, bounded | Generic paragraph untethered from this system |
| Related work | Delta-first against the sensing literature | Citation catalog with no contrast |
Sentence-level rewrites
| Draft pattern | IPSN-safe rewrite |
|---|---|
| "Our method significantly improves accuracy." | "reduces localization error to X m (95% CI ...) vs |
| "It is energy-efficient." | "consumes X µJ per inference on <MCU> at |
| "We deploy it in the real world." | "deployed <N> nodes for |
| "Runs in real time on embedded devices." | "end-to-end latency X ms on <SoC>; fits <RAM/flash> footprint" |
| "The model detects the event well." | "detection rate D at F false alarms/hour on hand-labeled field audio, per site" |
Energy/latency/ground-truth discipline
[Energy] joules or µJ per operation, measured (rail/shunt/instrument named), not estimated
[Latency] end-to-end timing on the real SoC/MCU at a stated clock; number of runs
[Footprint] RAM/flash used vs available; where the model/pipeline had to shrink
[Ground truth] the reference (labels, co-located instrument, surveyed positions) and its own error
[Limits] site-specificity / calibration drift / generalization -> stated next to the result
Vignette: compressing a deployment paper
A draft with a long system tour, six figures, and a thin evaluation: keep the system description at the level needed to reproduce, the two figures that carry the energy and yield story, a per-site false-alarm table, and a limits subsection; move the full protocol and secondary plots to the artifact with forward references. The test of a good cut: a reviewer should be able to answer "what was measured, on what hardware, against what ground truth, and how well did it hold up?" from the body alone.
Output format
[Writing diagnosis] clear / under-motivated / over-claimed / evidence-mismatched / over-scoped
[First-page fix] <new framing leading with the sensing problem and its budget>
[Claim audit] <claim -> measured evidence -> on real hardware / ground truth? yes/no>
[Budget check] <energy / latency / footprint reported? where>
[Anonymity edits] <tool/testbed names / board photos / self-citations / dataset links to fix>
版本历史
- 9f86f09 当前 2026-07-19 16:04
