agentsclimarketplace

Source grounding

Skill HetCreep/CoalMine/skills/source-grounding

Verify version-sensitive facts against live authoritative sources before asserting them in code or answers. Triggers on: "/source-grounding", "source-grounding", "sourcing". Standing rule — always active via CLAUDE.md. Invoke for deep verification work (API signatures, CVEs, model IDs, auth flows, deprecated patterns, security advisories).From its SKILL.md

Install
npx -y skills add HetCreep/CoalMine --skill source-grounding

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

One thing to look at

  • 15 stars15 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

5.4 KB, as published. Nobody here has run it

Source Grounding

<!-- SHARED:LANGUAGE_HEADER -->

Standing rule — active every response. No invocation needed for routine use.

Consent gates (G1–G4, declared)

#what it gateswhen it firesAgent laneHook lane
G1model tierbefore work startsask_question, 3 tiers, wait for picksuppressed — auto-Light, no tier question
G2how to proceed on an unfetchable sourcesource can't be fetched, user presentask_questionask_question
G3entanglement hand-offafter the findings, cross-domainask_question, onceask_question, once
G4self error-reportskill misbehavesoffer, never auto-submitoffer, never auto-submit

Hook cells assume an interactive session (G2–G4 need a user to answer); non-interactive Hook fires D1 instead of G2, and offers nothing for G3/G4.

What to verify (not memory)

  • CRITICAL (always fetch or flag — P2): API/SDK call signatures · library versions & deprecations · CVEs/security advisories · auth/crypto specs · LLM model IDs & params
  • MEDIUM (verify when unsure): package names · config keys · CLI flags · protocol specs
  • LOW/stable: math, algorithms, language syntax → memory fine (P3)

How

  1. Identify the version-sensitive claim.
  2. Name the authoritative source (official docs, advisory DB, package registry, spec, source code).
  3. Fetch (WebSearch/WebFetch/docs MCP) — or flag ⚠️ unverified: check [source] (D1).
  4. Cite at CRITICAL/MEDIUM. Don't over-verify stable facts (P3).

Per-claim-type authoritative source map: read references/sources.md when choosing where to verify.

Source hierarchy (1 = strongest)

  1. Source code / spec / RFC
  2. Official/vendor docs — authoritative secondary (honor .coalmine.json trustedDomains if set: treat those domains as additional authoritative / tier-2 sources)
  3. Multiple reputable third-party sources
  4. Single blog — corroborate first (P4)
  5. Training memory — weakest for volatile facts

Why each rank sits where it does: references/sources.md.

Non-interactive runs: log unfetchable claims as ⚠️ UNVERIFIED and continue (D1). Interactive: when sources cannot be fetched, confirm how to proceed via ask_question (G2).

Prohibitions (P1–P6, declared)

#never …
P1default to English just because this file is English
P2skip fetching or flagging a CRITICAL version-sensitive claim
P3over-verify a stable/LOW fact
P4cite a single blog source without corroborating first
P5auto-submit the self error-report
P6include unapproved code or paths in the self error-report

The shared footer's never fix without a chosen option does not apply here — this skill defines no Fix mode section, so that clause resolves vacuously; not counted above.

Degrade paths (D1–D4, declared)

#branchfires when:
D1log as ⚠️ UNVERIFIED, continue, never blocknon-interactive, source unfetchable
D2degrade to model tier + reasoning depth, never fake parallelismno capability lever for the target tier on this host
D3fall back to a numbered text menuhost has no question tool
D4fixed at Light, no tier question, no sub-agentsHook Context (auto-triggered)

This ledger deliberately diverges from skill-authoring.md §3b's column-or-separate-ledger rule for lane-applicability — gold-standard keeps a lane column, but here a column asserting a lane value that contradicted its own row's condition text measured worse than no column at all (SKILL-VARIANCE-WALK.md §Run 43: a bimodal Hook-Q4 split, the pre-registered key matching neither camp). D2 is restated at four sites in the shared partials — the general clause, the Standard row's "(else single-agent)", the Heavy row's "if supported", and the Heavy-specific "escalate by model + reasoning only" — all one row. D3 is stated once, in the shared Escalation footer's question-tool list ("…none → numbered text menu"). Neither is a new branch. D4's own branch text is restated verbatim in the footer's Hook Context line — same row, not a new one. The Freshness cap (scope already audited this session → cap at Light) is a tier-selection modifier on G1, not a degrade branch. The footer's Fix-mode-dependent offer clause is not a fifth branch — this skill defines no Fix mode section, so it never fires.

Output — 2 locations, declared

A location is a place this skill writes something a reader can see; the absence of an annotation is not one.

  • Verified: ✅ [claim] — source: [link/file]
  • Unverified: ⚠️ unverified — check [exact source]

Stable fact: no annotation is written — not a location, not counted above.

AUTHORITATIVE vs DIVERSE

  • AUTHORITATIVE (one ground truth): API/version/config/spec → go to the actual source code or official docs.
  • DIVERSE (triangulate ≥ 3): "what's best" / landscape / patterns → multiple repos + docs + community; note conflicts.
<!-- SHARED:ORCHESTRATION --> <!-- SHARED:ESCALATION_FOOTER -->

What ships with it: 2 files

2.7 KB alongside SKILL.md

references/

Gives 0 of the 12 instructions most quality gates skills give

Counted across 1,524 of the 2,830 authors here whose files we hold, read 2026-09-06

  • Read full output and check exit codein 45 of 1524, across 40 files
  • Verify output confirms the claimin 44 of 1524, across 39 files
  • Identify the command that proves the claimin 43 of 1524, across 39 files
  • Execute the full verification commandin 36 of 1524, across 30 files
  • Produce a verification reportin 34 of 1524, across 18 files
  • Review git diff changesin 30 of 1524, across 16 files
  • Fix build failures immediatelyin 29 of 1524, across 9 files
  • Group findings by severityin 28 of 1524
  • State claim only with evidencein 27 of 1524, across 22 files
  • Verify regression tests with red-green cyclein 26 of 1524, across 22 files
  • Run the full test suitein 26 of 1524, across 25 files
  • Run test suite with coveragein 25 of 1524, across 10 files

Said here and by no other author read

  • identify version-sensitive claims
  • name the authoritative source
  • fetch authoritative sources for critical claims
  • cite sources for critical and medium claims
  • triangulate diverse claims with three sources
  • log unverified claims in non-interactive runs

Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once.

Keep looking

Skills are one crate of 611,919. 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.