agentsclimarketplace

Deep research

Skill aburan28/crypto-autoresearcher/.claude/skills/deep-research

Deeply analyze the whole research portfolio -- ledger, knowledge corpus, and literature -- and produce a ranked, justified shortlist of what experiments to run next. Dispatches idea-generator for fresh divergent candidates and the coordinator for portfolio-wide prioritization, with an optional red-team stress-test of the shortlist. Read-only: writes no ledger record and changes no status. Use when the user asks what to work on next, wants a research roadmap or priority ranking, a goal's next_action needs deeper justification than the mechanical fallback order, or before opening a new campaign against an open question.From its SKILL.md

Install
npx -y skills add aburan28/crypto-autoresearcher --skill deep-research

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

13.7 KB, as published. Nobody here has run it

Deep research

Answer one question with rigor: of everything this program could do next, what is actually worth doing, and why? This is not another status readout — /research-status already gives the read-only ledger scan. This skill spends real reasoning effort turning that scan, the knowledge corpus, and external literature into a ranked, falsifiable, cost-aware shortlist, and defends it against the two failure modes AGENTS.md and docs/inventor-protocol.md name as symmetric: overclaiming a result, and prematurely closing a promising lead (AGENTS.md rule 9; inventor-protocol §4).

Read-only and side-effect-free. This skill writes no ledger record, changes no hypothesis or goal status, and commits nothing. It is safe to run from any worktree, on any branch, at any time — no git hygiene, ID pre-allocation, or snapshot/ledger archive is needed, because the output is a report handed to the user (or fed into the Coordinator inside /coordinate-research-goal), never a durable artifact by itself. Acting on a shortlist entry still goes through /propose-ideas, /design-experiment, /run-experiment, /curate-knowledge, or a Coordinator ledger archive, exactly as it would without this skill.

Scope

Take a scope from the user's request, in priority order:

  1. An explicit RQ-*, GOAL-*, or H-* — analyze within that boundary.
  2. "the active goal" / current branch context — resolve the same way /launch-research-harness step 3 resolves goal selection.
  3. No scope given — analyze the whole portfolio. This is the expensive, comprehensive mode and is the default for a bare /deep-research.

State the resolved scope back to the user before dispatching any subagent.

1. Ground the scan in current committed state

Many worktrees write to this ledger concurrently (CLAUDE.md "Concurrency: many agents, many worktrees"). A recommendation built on a stale local branch can duplicate work another session already landed, or already has an open PR for.

  • git fetch origin, then read ledger and knowledge state as of origin/main (or git merge-base HEAD origin/main) rather than trusting an out-of-date local branch tip.
  • gh pr list --state open and skim titles/bodies for record IDs that overlap the scope — an open PR may already be doing the work a naive scan would recommend.
  • If the local branch carries uncommitted or unpushed records relevant to the scope (this branch's own in-flight work), fold them in and say so explicitly; they are real even though origin/main cannot see them yet.

2. Comprehensive context gathering (this session, read-only)

Do this directly — no subagent needed, exactly as /research-status does its scan. Read broadly before ranking anything:

  • Governing docs: AGENTS.md (rule 9 — good-faith pursuit, no suppressing a plausible lead without recorded evidence/budget/boundary/ successor; rule 4 — scoping; rule 6 — citation discipline; rule 5 — no fabrication), docs/target-result-profile.md (Parts A-C, the C1-C18 checklist), docs/inventor-protocol.md (all 8 sections), docs/claims-and-verification.md, docs/task-lifecycle.md.
  • Ledger, the whole tree: ledger/questions/, ledger/proposals/ (not yet converted to hypotheses), ledger/hypotheses/ grouped by status, ledger/evidence/, ledger/decisions/ (recent next_actions), ledger/goals/ — including sharded goals under ledger/goals/GOAL-*/{goal.yaml,checkpoints/*.yaml} — and ledger/handoffs/. Grep the whole tree recursively (grep -rl <term> ledger/ knowledge/), never just the top-level ledger/*.yaml files: those are frozen legacy records covering a fraction of the area codes, and a top-level-only scan reports a clean "free" for an area that is in fact heavily worked.
  • Knowledge corpus, every subtype: knowledge/open-problems/ (read the "why it matters" / forward-guidance text on every KN-OPEN-* record in scope — this is the program's own list of promising, unresolved leads), knowledge/findings/ (proven boundaries — what NOT to re-propose), knowledge/techniques/, knowledge/literature/.
  • Reconciliation flags — grep goal records for verdict: UNRESOLVED, unresolved_carried_forward, and what_would_settle_it. These are pre-scoped, uncontended, already-justified units of work with acceptance criteria already written down by whoever raised them; treat each as a free-standing candidate before generating anything new, and verify its stated premise rather than inheriting it — a session that raised the flag by analogy can be wrong about the state it assumed.
  • Promotion-gate and validation debt: hypotheses carrying an asymptotic_claim stuck below supported — check which of the four promotion gates (agents/coordinator.md; docs/target-result-profile.md Part D) are still open: archived proof decomposition, every heuristic validated or scheduled, a concrete-cost table, independent xhigh review. Also check hypotheses whose heuristic_assumptions[] carry an empty validation_experiment_ids, and proof-oriented hypotheses whose proof_search_map is null or has an unaudited field (docs/inventor-protocol.md §8). Discharging a named gate on existing, already-approved work is frequently higher-value and lower-risk than generating a new idea from nothing.
  • Knowledge retrieval (crypto-kb MCP): call search_knowledge and find_related for every candidate direction under consideration, per AGENTS.md "Knowledge retrieval policy" — this skill's entire purpose triggers that policy's "proposing an experiment likely to duplicate earlier work" and "asserting an avenue already tested" clauses. If the server has no index built (:memory: / empty), say so plainly rather than silently skipping novelty checks.

Build, in your own working notes (not committed anywhere), a frontier map per live target problem in scope: current best internal/external result, dominated_by and sota_delta as last honestly recorded (never assume null without checking — docs/inventor-protocol.md §5), and which open problems or findings bound it.

3. Divergent pass — fresh candidates (optional, dispatch idea-generator)

Skip this only when the scope already has more ready, ranked work than the user's stated budget can cover (several approved-but-unrun hypotheses, several open reconciliation flags) and the user wants ranking, not new ideas. Otherwise dispatch the idea-generator subagent, giving it:

  • the resolved scope, the frontier map from step 2, and which established families/objects are already exhausted per the corpus, so it applies docs/inventor-protocol.md §1's off-limits naming rather than regressing to a variant in new notation;
  • an explicit instruction to run the lossy-projection test (§2) on every candidate before returning it — free, no compute, the cheapest available filter for "is this actually new";
  • an explicit instruction that "this target looks saturated" is not by itself a reason to generate nothing (§4's premature-closure standard) — if it reaches that conclusion for part of the scope, it must meet the named-obstruction-plus-argument-plus-forward-guidance bar, not report a screened-and-rejected count as a closure.

Take its full output, including the mandatory §5 honest-accounting block (dominated_by, sota_delta, enumerated closures, open directions) even when it proposes nothing usable — a session that finds nothing still owes that block.

4. Convergent pass — portfolio ranking (dispatch coordinator)

Dispatch the coordinator subagent — prioritization is its authority (AGENTS.md "Roles"), and this step is where that authority earns its keep, even though nothing here is written as an official decision. Give it:

  • everything gathered in steps 1-3: the frontier map, reconciliation flags, promotion-gate debts, and the idea-generator's fresh candidates (if run);
  • explicit instruction that this is advisory synthesis, not a coordinator_decision: no ledger write, no status change, no experiment approval — the deliverable is a ranked report only;
  • the ranking rubric below, to apply explicitly per candidate rather than asserting a holistic order.

Ranking rubric — state each axis per candidate, do not just assert a final order:

  1. Cost tier. Free/cheap pre-compute audits first — the lossy-projection test, proof_search_map's four audits, a reconciliation flag's named check, a controls-before-belief null-object run — before anything that needs a real experiment budget.
  2. Expected information gain vs. cost — the idea-generator's own convention, applied portfolio-wide: what does a positive vs. negative outcome each teach, and does the minimal test actually discriminate between explanations.
  3. Gate/dependency unlock — does this discharge a named promotion gate, validate a heuristic something else depends on, or unblock a hypothesis stuck at specified/approved for lack of exactly this?
  4. Target-profile ambition (C1-C2) — exponent-moving mechanisms over logarithmic-cofactor or constant-factor polishing; a cofactor-level candidate is not excluded, but must justify its priority as a building block rather than being scored as if it were exponent-first.
  5. Pareto/SOTA honesty — dominated_by checked against every frontier row, not defaulted to null; sota_delta stated quantitatively.
  6. Premature-closure check — for any candidate the scan wants to drop, or any open problem the corpus already calls saturated, hold it to the §4 closure standard before excluding it; a fatigue report is not a reason to deprioritize.
  7. Contention — flag candidates already covered by an open PR found in step 1; prefer surfacing the uncontended alternative over recommending duplicate work.

5. Adversarial stress-test (optional, dispatch red-team)

For a portfolio-wide run, or whenever the top-ranked candidate would be expensive, run one red-team pass against the top 3-5 shortlist entries only. This is a deliberate extension of Red Team's normal trigger: it usually reviews a Coordinator-committed snapshot of a completed result; here there is no snapshot because nothing has run yet, so it is reviewing the prioritization reasoning itself. Say this explicitly in its handoff, and keep its authority unchanged — it may not approve, rank, or write any ledger record; its output is a challenge memo the shortlist is weighed against, not a gate that blocks anything.

Ask it specifically: which top candidate has a hidden assumption or omitted cost; which "obviously next" pick is actually regression to an off-limits established family under new notation; whether any dominated_by: null in the shortlist was actually checked against every frontier row; whether any proposed closure meets the §4 standard rather than being a fatigue report.

Fold surviving objections back into the shortlist — reorder or annotate; do not silently drop a challenged entry without recording why it survived.

6. Report

Present, in this order:

  1. Resolved scope, and the commit/PR state it was checked against.
  2. Ranked shortlist, tiered (free/cheap pre-compute first, then highest information-gain, then fresh speculative candidates last). Each entry carries: target ID (existing RQ-*/H-*/GOAL-*/KN-OPEN-*, or NEW), a one-line description, the rubric axes from step 4 that drove its rank, its dominated_by/sota_delta line, and the exact next skill/command to run it (/propose-ideas RQ-..., /design-experiment IDEA-..., /run-experiment EXP-..., /curate-knowledge KN-..., or "feed to /coordinate-research-goal as next_action for GOAL-...").
  3. Red-team objections, if run, and how each was resolved.
  4. One explicit top pick — "if only one thing happens next, it is X, because Y" — mirroring the idea-generator's own single-pick convention.
  5. What this report is not: no ledger record was written and no status changed; every recommendation still needs its normal lifecycle skill and, where applicable, Coordinator approval and a ledger archive commit before it is official.

Rules

  • Never fabricate a citation, a prior result, a cost estimate, or a dominated_by/sota_delta value; an unchecked null is a fabrication under AGENTS.md rule 5, in this report exactly as in a filed idea.
  • Never treat "this looks saturated" as a valid reason to drop a candidate without meeting the docs/inventor-protocol.md §4 closure standard — rule 9 makes that the same failure as overclaiming.
  • Never let this skill's own output be cited as evidence, an approval, or a status change. It is a pointer to work, not the work.
  • Every candidate cites the exact record IDs or corpus paths it rests on; a shortlist entry with no citable basis is a guess, and must be labeled one.
  • If the portfolio scan or the knowledge-retrieval step finds nothing usable in scope, say so plainly and name what would need to happen first (e.g. no open RQ-* in scope, or crypto-kb has no index built) rather than inventing a plausible-sounding shortlist.

Keep looking

Skills are one crate of 608,475. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.