zy839971925-zyy/Agent--deep-research-Workflow-Skills / integrations/deepseek-harness/skills/reasoning-workflow
reasoning-workflow
VerifiedDefault silent workflow for everyday questions, research, writing, design and engineering. Match effort to the request, resolve costly ambiguity, find specialist resources and verify the requested outcome. Explicitly selectable with “调用我的 workflow skills”.
references/formal-review-and-audit.md
View on GitHub> Upgrade 2.1: read the updated [host contracts](host-adapter.md) and [migration rules](../UPGRADE.md) for executable admission, review and storage semantics. The shorter root SKILL is the active entry. Older examples are descriptive and must satisfy the current schemas.
# Formal review and audit
**Trigger:** Systematic/scoping review, claim-by-claim audit, methodology audit, controlled-change audit, or independent quality review is requested/warranted.
**Reads:** existing claims/artifacts, evidence records, methodology, requirements, state/change records, verification results.
**Updates:** audit findings, severity/blocker status, missing evidence/coverage, corrective actions/reopen decisions.
**May invalidate:** claims, approvals, delivery readiness, change closure.
**Exit:** material findings are dispositioned or explicitly left open with impact and next action.
**Related:** [evidence and provenance](evidence-and-provenance.md), [traceability and integrity](traceability-and-integrity.md), [change governance and effectiveness](change-governance-and-effectiveness.md).
**Return:** [root workflow](../SKILL.md) → Challenge / Reconcile / Completion.
Use only when the user requests a reproducible review, systematic/scoping synthesis, formal evidence audit, or rigorous review of an existing research product. Ordinary tasks should not inherit this ceremony.
## Formal evidence review
Define the review question, intended use, eligibility criteria, information sources, search dates, search strategy, deduplication, screening stages, extraction fields, and synthesis method before claiming reproducible or comprehensive coverage. Use a domain-appropriate question framework; PICO is not universal.
Record protocol amendments and reasons. Distinguish title/abstract from full-text screening. Keep inaccessible full texts separate from exclusions and retain exclusion reasons. Track study/report identity so multiple publications do not become multiple independent studies. Check corrections, retractions, protocols/registries, and risk of bias when relevant.
Do not claim exhaustive coverage of an open web universe or independent dual screening unless actually performed. A flow diagram or large citation count does not establish methodological compliance.
## Audit an existing answer or report
Start by extracting the material claims and their evidence dependencies. Preserve sound work; do not redo everything by default. Prioritize claims that are high-impact, weakly supported, surprising, current, quantitative, causal, accusatory, or central to downstream decisions.
Test:
- whether cited evidence entails the exact claim;
- entity/version/date/jurisdiction/population/unit alignment;
- provenance independence and copied-source clusters;
- whether inference is mislabeled as observation or fact;
- omitted counterevidence or plausible alternatives;
- causal overreach, hindsight contamination, or unsupported psychological inference;
- stale/superseded material;
- negative/completeness claims whose searched universe is undefined;
- calculations, transformations, code, or experiments claimed but not actually performed.
Escalate to fresh research only for gaps that can materially change the audit result. The audit deliverable should distinguish wrong, unsupported, outdated, out-of-scope, inaccessible, and still-uncertain claims.
## Audit reasoning quality
For Question / Reasoning work, audit the reasoning architecture as well as citations. Check:
- **problem understanding:** did the analysis identify the real epistemic task and what would count as answering it?
- **question drift:** did it solve a nearby proxy problem instead of the original one?
- **model quality:** are observation, premise, assumption, hypothesis, inference, judgment, recommendation, and unknown kept distinct where material?
- **decomposition quality:** do branches overlap unnecessarily, miss answer-bearing dependencies, or mix incompatible analytic levels?
- **recomposition:** would perfect branch answers actually imply the parent answer?
- **relationship quality:** are causal, correlational, dependency, mediation, constraint, sequence, membership, and alternative relations distinguished rather than collapsed into generic “connections”?
- **hypothesis competition:** were plausible alternative explanations tested using discriminating evidence?
- **source fitness:** were sources actually positioned to know the exact claims they support?
- **uncertainty quality:** is uncertainty attributed to missing evidence, conflict, measurement, scope/definition, model/premise, causality, or future behavior rather than vague hedging?
- **challenge strength:** are reversal conditions, weak premises, counterexamples, and counterevidence proportionate to the claim?
- **stopping:** did inquiry stop because marginal answer gain became low, rather than because a source/search quota was reached?
A well-cited answer can still fail if its problem model or inference structure is wrong.
## Reviewer independence is not reviewer count
Agreement among several reviewers is weak evidence when they share the same source set, rubric, framing, model family, or blind spot. Treat reviewer count as redundancy, not automatic independence.
Prefer independence that changes a potential error mechanism: different underlying evidence, measurement, tool, methodology, model/context, deterministic reproduction, or qualified human review. A fresh reviewer with the same evidence can still be useful for error detection, but do not label it orthogonal unless the route can discover errors the solver's route may systematically miss.
When reviewers disagree, do not vote. Localize the disagreement to the exact claim, premise, scope, metric, method, value judgment, or evidence item and create the smallest targeted verification need that could resolve it.
## Audit the evaluator too
A rubric, judge, benchmark, or test harness is itself a measurement instrument. When evaluation drives a material conclusion, inspect:
- what construct the evaluator actually measures;
- whether task/environment/tool setup exposes the capability being judged;
- whether rubric coverage matches the real success condition;
- whether judge errors are correlated with candidate errors;
- whether verbosity, position, style, or formatting can distort the score;
- whether a deterministic check, direct observation, or domain-specific oracle can replace subjective judgment for part of the task.
Do not equate `score = capability` without validating that measurement chain.
## Synthesis discipline
Formal review quality is not citation volume. Preserve study/source dependence, incompatible estimands, methodological heterogeneity, and unresolved disagreement. Quantitative synthesis requires compatible inputs and justified methods; do not pool simply because numbers are available.
## Audit a controlled change or completed work
When auditing a work/change record, compare the documented baseline and approved delta to the actual current state. Check whether risk findings were dispositioned, required controls were implemented, procedures/training/ownership were updated, deviations were recorded, verification/validation evidence actually exists, and effectiveness criteria were checked at the intended time. Separate implementation fidelity from outcome effectiveness and from causal impact. Track audit anomalies to closure rather than treating the audit report itself as resolution.
[← Return to root workflow](../SKILL.md)
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation). | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK). | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |