القائمة
Verifies claims and assesses source credibility. Use when fact-checking claims, verifying accuracy, validating sources, detecting hallucinations, triangulating evidence, or performing SIFT or RADAR assessment.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
| name | verifying-claims |
| description | Verifies claims and assesses source credibility. Use when fact-checking claims, verifying accuracy, validating sources, detecting hallucinations, triangulating evidence, or performing SIFT or RADAR assessment. |
Medium freedom: follow the phases sequentially (extract → mode → assess sources → weight → triangulate → resolve conflicts → calibrate), adapting verification depth to claim significance and risk.
Non-negotiable standards:
Extract discrete, verifiable assertions: statements presented as fact; names, dates, numbers, statistics, events; asserted causal relationships and predictions; factual claims embedded in interpretive passages.
Example: "The UK's ageing population, which reached 67.3 million in 2021, will strain healthcare resources" yields three claims — population 67.3 million in 2021 (factual), population ageing (factual, needs definition), will strain healthcare (interpretive/predictive).
Categorise each claim:
Filter uncheckable claims and note why: future predictions, personal experiences, vague claims ("many people believe…"), hypotheticals.
Prioritise by significance (affects key conclusions or decisions), specificity, risk, and feasibility.
Mode A — Retrospective (sources provided). Check claims against provided documents; the standard mode for detecting extrinsic hallucinations in AI-generated content. Compare each claim systematically against the sources and flag:
Check: are statistics and numbers present in the sources? Are quotes actually from the cited source? Are causal claims supported? Are dates, names, and details consistent? Does the output acknowledge limitations the sources mention? Is its confidence calibrated to source certainty?
AI-specific detection methods: references/hallucination-detection.md.
Mode B — Proactive (find sources). No sources provided, or provided sources insufficient. Apply SIFT:
Apply to each source before weighting its evidence:
Full rubric with red flags and scoring: references/source-evaluation-radar.md. Always prefer official/primary sources over third-party interpretation.
Context-dependent weighting adjustments: references/evidence-hierarchy.md.
Significant factual claims require at least three independent, high-quality sources.
| Level | Criteria | Language |
|---|---|---|
| Confirmed (90–100%) | Multiple high-quality independent sources agree; primary sources consistent; no credible contradiction | "Confirmed", "Verified" |
| Likely (70–89%) | Majority of quality sources agree; minor gaps remain | "Likely", "Strong evidence suggests" |
| Possible (50–69%) | Some credible support; limited independent verification | "Possible", "Some evidence suggests" |
| Unverified (30–49%) | Single-source claim or no authoritative corroboration found | "Unverified", "Cannot confirm" |
| Disputed | Authoritative sources explicitly disagree; expert opinion divided | "Disputed", "Experts disagree" |
Percentages are indicative ranges, not calculated scores — the qualitative criteria govern.
Audit trail — claims, categorisation rationale, sources with RADAR summaries, search strategies, findings, triangulation results, conflict resolution, confidence rationale, limitations. Standard: a reader can reproduce the verification.
Corrections when verification reveals errors: state the incorrect claim against the corrected information with sources; make the correction prominent; record the date, who identified it, and the process; scale to severity (brief note → detailed correction → flag for retraction).
Per-claim report block:
## Claim: [specific factual assertion]
**Categorisation:** Factual/Interpretive/Opinion/Uncheckable
**Mode:** Retrospective/Proactive
**Sources:** [list with tier + RADAR summary]
**Findings:** [evidence for/against with attribution + triangulation]
**Confidence:** Confirmed/Likely/Possible/Unverified/Disputed
**Rationale:** [why this confidence level]
**Limitations:** [uncertainties/scope boundaries]
More edge cases, time-constrained protocols, and common pitfalls: references/edge-cases-and-pitfalls.md.
references/source-evaluation-radar.md — RADAR rubric, red flags, scoring guidancereferences/hallucination-detection.md — AI-specific detection (semantic entropy, extrinsic checks, confidence vs correctness)references/evidence-hierarchy.md — tier detail and context-dependent weightingreferences/edge-cases-and-pitfalls.md — additional edge cases, time-constrained protocols, pitfallsDelegates a task to OpenAI Codex running as an interactive session in a herdr pane - uses the user's ChatGPT subscription, visible in herdr, steerable mid-session. Use when delegating work to codex, offloading a task to a GPT model (sol, terra, luna), or running a second opinion from a non-Claude frontier model. Not for Claude Code's own Agent tool or local-LLM delegation (delegating-to-local-llm).
Delegates a task to a local LLM running as a Pi coding-agent session in a herdr pane - the subagent is visible in herdr, can be steered mid-session, and costs no cloud tokens. Use when delegating work to a local model, spawning a Pi subagent, offloading a task to omlx (qwen, gemma), or running agent work offline. Not for Claude Code's own Agent tool or subagents inside Pi.
Loads the rageatc-code-oss operating rules directly from the plugin when the SessionStart hook has not delivered them. Use when the rageatc-code-oss failsafe line appears in session-start hook output, when the plugin's operating rules are absent from context, or when the injected rules look truncated.
Loads the rageatc-core-oss operating rules directly from the plugin when the SessionStart hook has not delivered them. Use when the rageatc-core-oss failsafe line appears in session-start hook output, when the plugin's operating rules are absent from context, or when the injected rules look truncated.
Raises an artefact's signal-to-noise ratio by removing text that earns nothing for its reader — filler, repetition, over-explanation, ornamental phrasing — and relocating conditional detail to references. Use to slim a skill, agent, or document, cut bloat, tighten prose, or shrink a file that has grown too long. Tuned for SKILL.md files.
Creates a design system for software with a UI. Use when a project has a user interface and architecture is confirmed — whether creating from scratch or extracting patterns from existing code.