An AI-agent skill that turns a structured applicant profile into a 16-20 page, chart-rich scholarship & study-abroad strategy PDF Report — with 10 ranked funding combinations, quantified cost gaps, past-recipient benchmarks and an honest verdict on whether your goal is even realistic.
A normal web search returns a wall of links and scattered facts. It never stacks awards into a funding path, never quantifies the gap you still owe, and never tells you the one thing that matters: is this goal even possible?
Every run emits a single polished PDF: cover, table of contents, executive summary, profile analysis, landscape, past-recipient benchmarks, comparison radar, the 10 funding combinations, feasibility, roadmap, a 12-month timeline and a final verdict.
The skill is a versioned, reproducible workflow built on a five-stage loop-engineering framework. Each loop is independently testable — and an ablation shows loops L2–L4 produce ~69% of the decision-grade artifacts.
Works with 70+ agents (opencode, Claude Code, Cursor, Codex…) via skills.sh, or run the generator directly with Python.
# One command — works with 70+ agents $ npx skills add piyushomanwar16/scholarship-strategy-report-skill # Then just ask your agent for a scholarship strategy report "Build me a funded-study strategy for an MSc in Canada"
# Via agentskill.sh $ npx @agentskill.sh/cli@latest install \ @piyushomanwar16/scholarship-strategy-report-skill # Or import the repo at https://clawhub.ai/import # (OpenClaw / ClawHub)
# Local / manual — clone, then run the generator $ pip install -r skills/scholarship-strategy-report/requirements.txt $ python skills/scholarship-strategy-report/generate_report.py \ skills/scholarship-strategy-report/example_profile.json demo.pdf # Omit the JSON to render the built-in EXAMPLE report
From the accompanying research paper: a controlled A/B test of 10 matched experiments (9 real profiles across Austria, Germany, UK, USA, Canada, Australia, France, Netherlands, Switzerland + one sparse profile). Both arms got the identical prompt — so every gap is attributable to the synthesis loop, not to better inputs.
% of 7 decision-critical artifact classes produced
correctly classifying goal as impossible / conditional / selective / consistent
These four charts are drawn from the bundled example_profile.json (a Canada data-science case). The Python generator produces the same figures in the PDF via matplotlib.
comparative radar · 6 dimensions
you vs applicant average
% conversion for well-matched profiles
relative funding value (higher = better)
The full 16- 20 page Report PDF built from the example profile — every section, table and chart described above. Generated by generate_report.py.
The controlled empirical comparison with conventional web search: methodology, per-experiment case studies, ablation and the L5 verification pass.
research_paper.pdf →