# Research Evidence

> Verify claims, source-backed facts, tool limits, assumptions, and evidence boundaries for workflow decisions without inventing citations or provider capabilities.

- Skill: `framecoreworks/research-evidence` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add framecoreworks/research-evidence`
- Raw SKILL.md: https://api.skillmd.com/api/skills/framecoreworks/research-evidence/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Research & Search
- Author: FrameCoreWorks (https://skillmd.com/u/framecoreworks)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/framecoreworks/research-evidence

---


# Research Evidence

## Language Policy

Public instructions and installation guidance stay in English. After verified installation, resolve the working language from an explicit user preference, then the user's own conversation, then a reliably exposed host locale, otherwise English. Copied English setup prompts and source files are not language preferences; do not infer hidden account settings. Keep explicit deliverable-language requests and exact supplied artwork copy separate from conversation language. Do not translate or rewrite public source files to localize a session.

## When To Use

Use this skill when a workflow depends on public facts, source-backed claims,
tool or platform limits, technical capability checks, product evidence, market
context, competitor references, or uncertainty that could affect final output.
Use it before copy, prompt packs, tool routing, QA, or delivery when accuracy is
material.

## Inputs

- Brief, question, claim, assumption, or decision that needs verification.
- Available sources, user-provided references, repo files, screenshots, or
  public links.
- Freshness requirement, jurisdiction, market, platform, model, or tool scope.
- Required output format such as Evidence Note, claim ledger, or decision note.
- Current tool availability, including whether web browsing is actually
  available in the host environment.

## Outputs

- Evidence Note with verified facts, source boundaries, uncertainty, and next
  decisions.
- Claim ledger entries for approved, unsupported, conflicting, or blocked
  statements.
- Tool-limit notes for `tool-routing-cost`, prompt packs, or QA.
- Author-boundary notes that distinguish user-provided first-hand facts from
  unsupported personal or social-proof claims.
- Handoff notes for `brief-architect`, `reference-curator`, `copy-voice`,
  `instruction-packet-factory`, `qa-iteration`, or `delivery-documentation`.

## Process

1. State the exact claim, decision, or assumption being checked.
2. Identify source type: user-provided, local repo, official docs, public web,
   or unavailable.
3. Verify current or unstable facts from reliable sources when browsing is
   available.
4. Separate confirmed facts from inference, opinion, and missing evidence.
5. Record conflicts, stale sources, unavailable sources, and confidence limits.
6. For ready-to-use copy, state whether each material claim, quote, result,
   testimonial, first-hand observation, or personal statement is user-provided,
   verified, unresolved, or prohibited.
7. Convert evidence into a compact decision note for the next workflow role.

## Decision Rules

- Prefer primary sources, official docs, repo files, and user-provided source
  truth over summaries.
- If browsing is unavailable, do not claim live verification; mark the evidence
  path as unavailable or user-provided only.
- If a fact is time-sensitive, high-stakes, legal, financial, medical, pricing,
  model-capability, or policy-related, require a freshness check before final
  use.
- If evidence conflicts, route to `qa-iteration` with the conflict and proposed
  resolution options.

## Guardrails

- Do not fabricate citations, links, quotes, benchmarks, capabilities, or
  release status.
- Do not use API keys, paid provider calls, hidden web automation, uploads, or
  private cloud links.
- Do not store raw private project context in public artifacts.
- Do not turn evidence notes into long research dumps; preserve only what the
  next role needs.
- Do not infer an author's experience, beliefs, emotions, or history from a
  writing preference or a generic brief.

## Handoff

Review gate: `evidence_fit`.

Hand verified facts and unresolved risks to `brief-architect`, `copy-voice`,
`tool-routing-cost`, `instruction-packet-factory`, `qa-iteration`, or
`delivery-documentation`. When sources are missing, stop with a clear blocker
instead of allowing unsupported claims downstream.

## QA Checklist

- Checked item is explicit.
- Source type and freshness boundary are stated.
- Facts, inferences, and unknowns are separated.
- Unsupported claims are not promoted to final copy or prompts.
- Conflicts and missing evidence have a next action.
- Handoff includes only compact, usable evidence.