AI Artifact Provenance: Documents That Explain Themselves

    1

    Every memo, report or HTML page your agent writes carries its own context: the exact ask, assumptions, rejected options, unknowns and a per-version changelog.

    Free

    2 usesSecurity scanned

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    AI Artifact Provenance: Documents That Explain Themselves

    Example session with this skill installed

    Write a one-page decision memo comparing Postgres and DynamoDB for our orders service. Keep it under 500 words, it's for my VP. We're staying on AWS.

    • Read your context and instructions
    • Compiled the ai artifact provenance

    Summary

    I wrote orders-db-memo.md: a one-page decision memo that recommends Postgres (Aurora) over DynamoDB, 430 words. At the end is an ai-artifact-provenance block, so your VP, or their assistant, can see what you asked, what I assumed, and what I didn't verify, without having been in this session.

    The memo (excerpt)

    Recommendation: Aurora PostgreSQL for the orders service.
    Orders need multi-row transactions (order + line items + inventory hold) and ad-hoc reporting queries. DynamoDB would mean designing single-table access patterns up front and building reporting through exports.

    The provenance block (end of the file)

    {
      "ai_artifact_provenance": "0.1",
      "title": "Orders service: Postgres vs DynamoDB",
      "ask": "Write a one-page decision memo comparing Postgres and DynamoDB for our orders service. Keep it under 500 words, it's for my VP. We're staying on AWS.",
      "purpose": "Pick the primary datastore for the orders service",
      "audience": "VP Engineering",
      "status": "proposal",
      "owner": "<you>",
      "generator": { "tool": "Claude Code", "model": "claude-opus-5-5" },
      "created": "2026-10-16",
      "updated": "2026-10-16",
      "inputs": [ { "name": "services/orders/schema.sql", "as_of": "2026-10-16" } ],
      "constraints": [
        { "quote": "Keep it under 500 words", "source": "user" },
        { "quote": "it's for my VP", "source": "user" },
        { "quote": "We're staying on AWS.", "source": "user" }
      ],
      "assumptions": [
        { "text": "Peak load stays under 2,000 orders/minute for the next 18 months", "confirmed": false },
        { "text": "The team has more SQL than DynamoDB experience", "confirmed": false }
      ],
      "decisions": [
        {
          "id": "D1",
          "decision": "Recommend Aurora PostgreSQL",
          "why": "Multi-row transactions and ad-hoc reporting are core to orders",
          "quote": "We're staying on AWS.",
          "rejected": [
            { "option": "DynamoDB", "why": "Access patterns must be fixed up front; reporting needs an export pipeline" },
            { "option": "Self-managed Postgres on EC2", "why": "Ops burden with no benefit at this scale" }
          ],
          "superseded_by": null
        }
      ],
      "unknowns": [
        { "text": "Aurora cost at projected volume", "how_to_check": "AWS pricing calculator with last quarter's order counts" }
      ],
      "changelog": [
        { "version": 1, "date": "2026-10-16", "trigger": "Write a one-page decision memo comparing Postgres and DynamoDB...", "changed": "First draft." }
      ],
      "how_to_question": "Give this file to your own assistant and ask it to read the ai-artifact-provenance block."
    }
    

    Next steps

    • Confirm or correct the two assumptions; I'll flip confirmed and add a changelog entry.
    • Ask for revisions as usual. Each republish appends to the changelog and never rewrites it.
    • Want the context visible on the page? Ask for the provenance panel (HTML) and a collapsed "How this document was made" bar appears at the top.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Hand an AI-written memo to your manager and they get the answer, but not the question. They can't see what was asked, what the AI assumed, what it ruled out, what nobody checked, or why version 4 differs from version 3.

    This skill fixes that. Before every publish, your agent embeds a small JSON provenance block in the document:

    • ask: the original request, quoted exactly
    • constraints: the limits you set, quoted
    • assumptions: what the agent chose without being told, flagged until you confirm them
    • decisions: each choice with your quote, the options it rejected, and why
    • unknowns: what wasn't verified, and how to check it
    • changelog: one append-only entry per publish, naming the prompt behind it

    The block is invisible by default, so the page looks the same. It stays in the file through email, upload and re-hosting. On request, the agent adds a collapsed "How this document was made" panel at the top of the page (inline, no CDN, works under a strict Content Security Policy).

    It also works in reverse. Give your agent a document that has a block, and it can answer "why was X rejected?" or list what to challenge.

    Works with: Claude Code, Codex, Gemini CLI and Cursor: any agent that loads SKILL.md. Includes: SKILL.md, the panel renderer, the full spec, the JSON schema, and worked HTML and Markdown examples. No dependencies. If the open-source aap CLI (Node 18+) is available, the agent uses it to validate a document before publishing. Without it, the skill works exactly the same. Open source (MIT): the spec, CLI, Claude Code plugin and git hooks are at github.com/anshuljhawar/ai-artifact-provenance.

    It relies on the agent following instructions: only Claude Code's publish hook and the git pre-commit hook (both in the open-source repo) block a document that's missing a block. Unconfirmed assumptions stay flagged until you confirm them. It isn't C2PA or watermarking: it explains why a document looks the way it does, not which model made it.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    2 uses

    Times used to date

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 1 month ago

    • Free with an Agensi account

    Requires

    github
    Node.js

    Permissions

    Terminal / Shell
    Write Files
    Read Files
    Listed1 month ago

    What's inside

    Frequently Asked Questions

    It solves the problem of "context loss" by embedding the logic, rejected alternatives, and original constraints directly into the file, ensuring the rationale travels with the document.

    Any agent that loads the SKILL.md standard: Claude Code, Codex CLI, Gemini CLI, Cursor and others. The output is plain HTML or Markdown, so readers need nothing installed.

    No. The skill works with nothing installed. If the open-source aap CLI (Node 18+) is available, the agent also uses it to validate a block before publishing and to turn the visible panel on or off.

    Yes, it features a 'panel mode' that injects a visible, interactive 'How this document was made' section into HTML artifacts for human readers, while keeping the machine-readable JSON in the background.

    The skill maintains an append-only changelog within the artifact, tracking every version, trigger, and modification to provide a complete audit trail of how the document evolved.

    New in Technical Documentation