AI Answers from Provided Files Only: Step-by-Step Workflow

Step-by-step guide for answers based only on files you provide, with file references for each factual claim.

Purpose

This guide defines a files-only factual workflow. It tells the assistant to use only the files, logs, screenshots, excerpts, or other artifacts provided in the current task; cite file locations for factual claims; and stop when a required claim cannot be supported by the provided material.

When to use this

Use this section to decide whether this workflow is the right fit before you configure prompts, policies, or reference material.

  • Use case
    The evidence is already provided
    Use this when the user supplies the files, logs, screenshots, excerpts, repository snapshot, ZIP, or source material that the answer must inspect.
  • Use case
    External sources are out of scope
    Use this when the answer must not rely on web browsing, public documentation, model memory, prior knowledge, or uncited assumptions.
  • Use case
    Traceability is required
    Use this when factual claims need file-location support such as filename, line range, page, section, screenshot region, or another locator.
  • Use case
    The task may require a full scan
    Use this when the provided material is large or multi-file and the assistant must report coverage, gaps, and unsupported areas.

Step-by-step implementation procedure

Follow the workflow in order. Each step gives one action and one verification check before continuing.

  1. Step 1 · Configure the artifacts-only rule

    Add the artifacts-only system prompt to the instruction layer of the tool you use.
    Use a reusable instruction surface when the workflow will be repeated; otherwise place the rule at the top of the working prompt.
    Check: The rule explicitly blocks external sources, prior knowledge, guessing, and unsupported factual claims.

  2. Step 2 · Provide the allowed source material

    Attach or paste the files, logs, screenshots, excerpts, repository snapshot, ZIP, or other artifacts the answer is allowed to use.
    Keep the evidence boundary limited to the provided artifacts.
    Check: Every source the answer may use is present in the current task context.

  3. Step 3 · Add runtime prompt components

    Use deep-read for all artifact-bound tasks. Add deep-scan when the material is large or multi-file.
    Place the current question, selected artifacts, target output, and one-time constraints in the runtime prompt.
    Check: The task states what must be answered and how evidence should be cited.

  4. Step 4 · Require file-location references

    Require each factual claim to cite the supporting file and locator.
    Use filenames, line ranges, pages, sections, screenshot regions, or equivalent locators.
    Check: Every factual claim can be traced back to a provided artifact.

  5. Step 5 · Fail closed when evidence is missing

    If a core claim cannot be supported by the provided artifacts, stop instead of filling the gap.
    Return the fail-closed result or ask for the missing artifact.
    Check: Unsupported claims are not rewritten as conclusions.

  6. Step 6 · Add optional final verification

    Use chain-of-verification or confidence scoring only when the workflow needs a final acceptance gate.
    Run the final verification layer after the answer is drafted and before it is accepted.
    Check: The final answer reports only supported claims or clearly marks unsupported areas.

Verification checklist

Use this checklist before accepting the output, publishing it, or using it as evidence for a downstream workflow.

  • Source boundary
    Only provided artifacts were used
    The answer does not rely on web browsing, prior model knowledge, public documentation, or uncited assumptions.
  • Traceability
    Each factual claim has a locator
    Every factual claim cites the relevant file, line range, page, section, screenshot region, or equivalent locator.
  • Coverage
    Large inputs were scanned deliberately
    Repos, ZIPs, logs, or multi-file inputs include coverage notes, explicit gaps, and evidence pointers.
  • Unsupported claims
    Missing evidence was handled explicitly
    Unsupported claims are marked unsupported or fail closed; they are not completed from inference.
  • Boundary integrity
    Public-source prompts were not mixed into this workflow
    Web browsing, citation, authoritative-source, or literature-review prompts are not used as required assets for this files-only workflow.

Next steps