Verify claims in technical writing — procedure

Procedure for verifying technical writing claims, controlling overclaims, and producing evidence-aligned clean final text.

Purpose

Use this procedure in AI workflows when:

  • A draft was written or rewritten by an LLM and may contain added/strengthened claims.
  • The draft includes world-claims (standards/specs, versions, “latest”, metrics, comparisons, security/performance assertions).
  • You need an auditable output (clear separation of what is proven vs not proven).
  • Your runtime cannot browse (or browsing is disallowed) and the draft relies on external facts without pasted evidence.

Enforcement (fail-closed):

  • Inputs must include DRAFT and EVIDENCE MODE (ARTIFACTS_ONLY or EXTERNAL_VERIFICATION_ALLOWED).
  • You must not add new technical/factual claims. You may only remove, qualify, or rephrase claims to match available evidence.
  • Treat any link-only “source” as NOT VERIFIED unless the relevant evidence is quoted in the input or external verification is explicitly allowed and available.
  • Output must follow this exact section order:
    • A) TERMINOLOGY CONTROL
    • B) CLAIM LEDGER (VERIFIED / NOT VERIFIED / DISPUTED)
    • C) OVERCLAIM SCAN
    • D) FINAL CLEAN TEXT (or INSUFFICIENT_EVIDENCE)
    • E) CONFIDENCE (0–100)
  • Out of scope (architecture review / code-quality review / performance tuning / repo-specific best practices): output only
    INSUFFICIENT_EVIDENCE: out_of_scope_use_quality_gate
  • If the draft’s core intent cannot be preserved without NOT VERIFIED claims, output only:
    INSUFFICIENT_EVIDENCE: <what is missing>

Choose a mode

  • Option 1 (Artifacts-only): verify using only pasted excerpts/quotes/standard IDs (no browsing).
  • Option 2 (External verification allowed): verify using authoritative sources if browsing/search is available in the runtime.
  • Option 3 (Out of scope): use the Engineering Quality Gate for architecture/code-quality reviews.

Setup

1) Install the system prompt template:

  • evidence-gated-technical-writing-gate.system.txt 2) Paste the user runner template and fill it in:
  • evidence-gated-technical-writing-gate.user.txt 3) Set EVIDENCE MODE:
  • ARTIFACTS_ONLY (no browsing; pasted evidence only)
  • EXTERNAL_VERIFICATION_ALLOWED (browsing may be used if available) 4) Paste the DRAFT and SUPPORTING EVIDENCE:
  • Prefer pasted quotes/excerpts + standard IDs + doc sections.
  • Avoid link-only evidence unless EXTERNAL_VERIFICATION_ALLOWED is selected and browsing is available. 5) Require the exact output sections A→E (in order).

Verify (smoke test)

1) Provide a draft with at least one non-trivial claim, and provide no supporting evidence.

  • Expected: output only INSUFFICIENT_EVIDENCE: <what is missing> 2) Provide an out-of-scope request (e.g., “review my architecture”).
  • Expected: output only INSUFFICIENT_EVIDENCE: out_of_scope_use_quality_gate 3) Provide a draft + pasted authoritative evidence for each core claim.
  • Expected: D) FINAL CLEAN TEXT contains only VERIFIED claims, and E) CONFIDENCE (0–100) is present.

Options

Option 1 — Artifacts-only

Example

  • Question: “Verify claims in this draft and rewrite it so only VERIFIED claims remain.”
  • You must provide: the full draft + pasted evidence excerpts/quotes/standard IDs for each core claim. If missing evidence prevents preserving core intent, output INSUFFICIENT_EVIDENCE: <what is missing>.

Option 2 — External verification allowed

Example

  • Question: “Verify these claims using authoritative sources and rewrite the draft to remove overclaims.”
  • You must provide: the draft + scope constraints (product/version/date/time window) + confirmation that browsing/search is available. If browsing is unavailable, link-only sources remain NOT VERIFIED.

Option 3 — Out of scope: use the Engineering Quality Gate

Example

  • Question: “Review this repo for layering violations and best-practice issues.”
  • You must provide: goal + materials (files/tree) + constraints + authoritative sources (or explicit permission to browse).

Common mistakes

  • Using this procedure for architecture/code-quality review (use the Engineering Quality Gate instead).
  • Providing link-only “sources” while browsing is unavailable and expecting claims to be VERIFIED.
  • Leaving NOT VERIFIED or DISPUTED claims inside D) FINAL CLEAN TEXT (final text must contain only VERIFIED claims).
  • Not enumerating all non-trivial claims in B) CLAIM LEDGER.
  • Breaking the required output section order A→E.