More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Idea Discovery Pipeline

    1

    Turn a broad research direction into one validated, pilot-tested idea: map the landscape, generate candidates, kill them with cheap experiments, verify novelty, freeze a proposal.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Idea Discovery Pipeline

    Idea Discovery Pipeline

    Example session with this skill installed

    I'm a PhD student with one A100 (shared, ~20 GPU-hours a week) and a workshop deadline in 10 weeks. My direction is reducing hallucinations in LLM answers over long documents. Run the pipeline and tell me which idea I should commit to.

    • Read your context and instructions
    • Compiled the idea discovery pipeline

    Landscape first: the area splits into retrieval-side fixes (better chunking, reranking), generation-side fixes (constrained decoding, self-consistency), and verification-side fixes (post-hoc claim checking). The recurring limitation across recent papers: verification methods are evaluated on short answers, and nobody reports cost per verified claim. Open questions I could name: does claim-level verification hold up at 8k-token answers, and what does it cost?

    Nine candidates generated; the paper filter killed four (one needed a private dataset, two were generation-side variants already crowded, one was a career, not a paper). Three pilots, each capped at 4 GPU-hours: (1) sentence-level claim extraction + retrieval check on 50 long answers — caught 71% of planted errors at 0.9 GPU-hours per 10 answers; positive signal. (2) Self-consistency voting across 5 samples — caught 44%, cost 5× the tokens; killed by pilot. (3) Fine-tuned small verifier — training alone blew the cap before any measurement; not pilotable at your budget, demoted.

    Novelty check on the survivor: closest work verifies short QA answers; one preprint from last month checks claims in summaries, not long-form answers, and reports no cost numbers. Verdict: novel with caveats — your differentiation (long answers, cost-per-claim as a headline metric) has to carry the paper, and the proposal says so.

    Risk register, top risk: extraction errors masquerading as hallucinations — cheapest test is a hand-labeled set of 200 claims before any scale-up. Frozen proposal: three claims, each paired with its experiment; roadmap of four runs inside your 200 total GPU-hours; kill criterion written down: if verification cost exceeds 15% of generation cost at equal catch rate, pivot to the parked reranking candidate instead.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    The most expensive mistake in research is spending months on an idea that's already published, already tried, or untestable. This skill is a five-phase pipeline that front-loads the killing so the idea you commit to is one that earned it.

    Phase 1 maps the landscape — sub-directions, leading approaches, recurring limitations, and the open questions you must be able to name before generating anything. Phase 2 generates 8–12 concrete candidates (one-sentence hypothesis, method sketch, what counts as evidence), filters them on paper, then runs the cheapest discriminating pilot for the top few — each pilot capped in advance and designed so a negative result is final. Phase 3 verifies novelty the way a hostile reviewer would: closest existing work named, differentiation written in one paragraph, last six months checked for concurrent work, with explicit verdicts (novel / novel with caveats / dead). Phase 4 steelmans the survivor, then attacks it into a risk register where every risk gets its cheapest resolving test. Phase 5 freezes the proposal: problem anchor, method thesis, claims each paired with the experiment that proves them, a budgeted roadmap, and kill criteria written while you're still honest.

    You also get a candidate scoring rubric for the paper filter, a pilot design checklist, a fill-in final report template, and edge-case playbooks for the painful moments — everything dying in the novelty check, two survivors and budget for one, concurrent work appearing mid-project. The output is a single canonical report, dead ends included, so the next direction starts from evidence instead of memory.

    This skill is a process, not a database: it tells you how to search and what verdicts to draw, but the literature searching and the pilots are run by you or your agent with your own access and compute. It is calibrated for empirical ML/AI research; purely theoretical work (math, theory of computation) needs the pilot phases reinterpreted as proof-sketch attempts. It also cannot make a mined-out direction fertile — if everything dies in the novelty phase, the honest output is a narrower direction, not a resurrected idea.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 24 days ago

    Permissions

    No special permissions declared or detected

    Listed24 days ago
    Updated21 days ago

    What's inside

    Frequently Asked Questions

    For a typical direction, two to five days of part-time work: a day of reading, a day of generation and pilot design, then pilots bounded by the caps you set. The phases are sized so the whole pipeline costs less than two weeks of pursuing the wrong idea.

    You can enter at Phase 3, but the skill will push back once: favorites that skip the landscape and pilot phases are exactly how people end up scooped or stuck. Run at least a one-hour pilot and the closest-work paragraph before committing.

    That's a valid outcome and the report template has a place for it. The playbook is to narrow the direction or change a constraint (cheaper setting, different data, different metric) and re-run Phase 1 — not to revive a dead idea with new terminology.

    New in Research & Analysis