AI Manifest: Embedded Workflow Instructions for AI Agents

Introduction Large Language Model (LLM)-based AI agents increasingly interact with web services via browser-automation protocols such as the Model Context Protocol (MCP), Playwright, Puppeteer, and Selenium WebDriver. Current agents typically parse entire DOM trees or screenshots on every page to infer UI structure, producing three well-known problems:

Substantial token consumption due to repeated analysis of large DOMs on every session.
High failure rates on complex multi-step transactional UIs such as enterprise resource planning (ERP) systems, academic manuscript submission portals, and government e-services.
Absence of a standardized mechanism for a website operator to declare an AI-agent-friendly ("AI-Ready") operational surface.

Related prior work includes robots.txt, llms.txt, agents.txt, and ai-plugin.json. These address crawling permissions, LLM-friendly documentation, agent capability declarations, and API-level integration respectively. None provides step-by-step UI workflow instructions for multi-page transactional flows. AI Manifest fills this gap by specifying a JSON format that enumerates ordered UI operations keyed to CSS selectors. An AI agent detects, parses, and verifies the manifest before executing the listed steps, and avoids further DOM-based inference for those steps.

Conventions and Definitions The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here.

AI Manifest: A structured data object, expressed in JSON, that a website operator embeds into a web page or serves at a well-known URI, containing at minimum a task identifier, an ordered steps array, and for each step an action field and a CSS selector.
AI Agent: A software process, typically driven by an LLM, that accesses a web page via a browser-automation tool and executes actions on behalf of a human user.
Central Trust Registry: A network service that stores SHA-256 hash values of manifests pre-registered by publishers, and responds to real-time trust lookups from AI agents with a status of white-listed, black-listed, or unknown.
Canonical Form: The representation of an AI Manifest obtained by lexicographically sorting all JSON object keys at every nesting level and serializing the result using the JSON encoding defined in with UTF-8.

Protocol Overview

Embedding Methods A website MAY provide an AI Manifest via one or more of the following methods:

Method A: Well-Known URI The server SHOULD make the manifest retrievable at the following well-known URI : In addition, the HTML document SHOULD declare the entry point via an HTML meta element: ]]>

Method B: Hidden DOM Element The server MAY embed the manifest into the HTML response as a hidden element with the display:none style and aria-hidden="true" attribute: ]]>

Method C: HTTP Response Header The server MAY declare the manifest location and hash in a response header: ]]> Method C is RECOMMENDED in conjunction with Method A, so that an AI agent can discover the manifest URL and validate the hash in a single request-response round trip before fetching the body.

Method D: HTML Link Element The server MAY declare the manifest via an HTML link element in the document head: ]]> Method D allows AI agents that parse the DOM but do not inspect HTTP headers to discover the manifest location. The rel value ai-manifest is registered in .

Method E: AI-Side Curation A user MAY place a manifest at a vendor-conventional path in their own file system, for example ~/.claude/afrm/{domain}.json or ~/.codex/afrm/{domain}.json. The AI runtime loads the file when it visits a matching domain. Because the manifest originates from the user's authored context, it enters the agent's trust context without the prompt-injection screening applied to externally fetched data. When both an AI-side curated manifest and a site-side published manifest exist for the same origin, the curated manifest SHOULD take precedence as the more explicit trust signal. Method E is a transitional mechanism intended for domains that have not yet published a site-side AFRM. Once the site publishes its own manifest, the AI-side copy SHOULD be retired or used only to override individual fields.

Manifest Schema (AFRM Three-Part Model) An AI Friction-Recovery Manifest (AFRM) is a JSON object with the following top-level fields:

version (string, REQUIRED): Specification version. This document defines version "1.0".
publisher (string, REQUIRED): Canonical domain name of the publishing website.
manifestId (string, REQUIRED): Stable identifier scoped to the publisher, used for registry lookup.
registry_url (string, REQUIRED): HTTPS URL of the trust registry that the publisher has pre-registered this manifest with.
frameworkHints (object, OPTIONAL): Metadata about the UI framework in use (e.g., framework name, version, rendering model). Allows the AI agent to apply framework-specific heuristics before encountering friction.
knownTraps (array, REQUIRED): Primary content of the AFRM. An ordered array of trap descriptors, each describing a UI pattern that predictably blocks autonomous AI agents. Each element MUST include trapId (string), category (string from the IANA trap registry, see ), selector (string), and escapeAction (string). Optional fields include description and condition.
shortcuts (array, OPTIONAL): An array of direct navigation or action shortcuts that allow the AI agent to bypass multi-step UI flows when a direct API or deep-link equivalent exists.

Agent Detection Algorithm Upon loading a page, an AI agent implementing this specification SHOULD perform the following detection sequence before any full-DOM inference pass:

Check the agent's local file system for a user-curated manifest at the vendor-conventional path (e.g., ~/.claude/afrm/{domain}.json) matching the current origin (Method E). If found, load it into the trust context and continue; the site-side steps below may still be consulted for supplementary data.
Inspect the HTTP response headers for X-AI-Manifest (Method C).
If absent, retrieve /.well-known/ai-manifest.json (Method A) or resolve the URI declared by the meta element.
If still absent, check the document head for a <link rel="ai-manifest"> element and fetch the referenced URL (Method D).
If still absent, search the DOM for an element with id="ai-manifest" and read its data-manifest attribute (Method B).
If none is found, fall back to conventional DOM-based inference.

Canonical Hash and Trust Verification Prior to execution, the AI agent MUST compute a SHA-256 hash over the canonical form (see ) of the manifest and send a trust lookup request to the URI in the registry_url field. The request MUST use HTTPS and MUST carry the tuple {publisher, manifestId, hash} as a JSON body. The registry response is a JSON object containing a status field with one of the following values:

"white" — the manifest is trusted; the agent MAY proceed to execution.
"black" — the manifest is explicitly distrusted; the agent MUST abort and SHOULD alert the human user.
"unknown" — the manifest is not registered; the agent SHOULD warn the user and MAY fall back to DOM-based inference.

Implementations MAY cache a non-expired registry response keyed by the manifest hash, to avoid repeated network round trips for an identical manifest.

Execution When trust is confirmed, the agent executes the steps array in declared order, mapping each step's action and selector to a browser-automation primitive (e.g. "click", "fill", "select", "upload"). For the duration of a manifest-driven execution the agent SHOULD NOT perform additional LLM-based inference over the page DOM.

Central Trust Registry A Central Trust Registry accepts manifest registrations from publishers and answers real-time hash lookups from AI agents. A conforming registry SHOULD:

Store the SHA-256 hash of the canonical form of each registered manifest, together with the publisher and manifestId fields.
Perform static analysis on the submitted steps array and reject or black-list manifests whose selectors or actions match a published pattern of prompt-injection risk (for example, selectors targeting iframe elements for cross-origin form submission, or actions outside the registered action set).
Expose a community-driven mechanism for reporting and black-listing malicious manifests.

This document does not mandate a specific registry operator. Multiple interoperable registries MAY exist, and each manifest declares which registry is authoritative for it via registry_url.

IANA Considerations

Well-Known URI Registration This document requests IANA to register the following entry in the "Well-Known URIs" registry established by :

URI Suffix:: ai-manifest.json
Change Controller:: Independent Submission Stream editor
Reference:: This document
Status:: provisional
Related Information:: None

AI Manifest Actions Registry (initial) This document requests IANA to create a new registry named "AI Manifest Actions", with the following initial registrations. Registration policy: Specification Required .

click: Invoke a click event on the selected element.
fill: Type a value into a text input element.
select: Choose an option from a drop-down list element.
upload: Attach a file to a file input element.
wait: Pause for a condition or duration.
navigate: Change the current URL.
assert: Verify that a condition holds before proceeding.

AI Manifest Trap Categories Registry (initial) This document requests IANA to create a new registry named "AI Manifest Trap Categories", with the following initial registrations. Registration policy: Specification Required .

shadow-dom-trap: A UI component rendered inside a Shadow DOM tree whose selectors are not accessible via standard document.querySelector calls.
virtual-scroll-trap: A list or grid that renders only a visible viewport window, requiring programmatic scrolling before off-screen items can be targeted.
iframe-context-trap: An action target located inside a cross-origin or same-origin <iframe> that requires a context-switch before interaction.
native-dialog-trap: A browser-native modal dialog (invoked via window.alert(), window.confirm(), or window.prompt()) that freezes the renderer event loop until dismissed. An AI agent must call the appropriate WebDriver dismiss or accept command rather than attempting a DOM interaction, which will block indefinitely.
delayed-render-trap: A UI element that does not appear in the DOM until after an asynchronous operation completes; the agent must wait for the element's visibility predicate before acting.

Link Relation Type Registration This document requests IANA to register the following entry in the "Link Relations" registry established by :

Relation Name:: ai-manifest
Description:: Refers to an AI Friction-Recovery Manifest (AFRM) that describes known UI traps, framework hints, and interaction shortcuts for the current origin, to assist autonomous AI agents.
Reference:: This document

Security Considerations

Prompt Injection Risk A malicious website could embed an AI Manifest whose steps array leads an AI agent to perform actions harmful to the user (for example, submitting a form to a third party with user-supplied credentials). The Central Trust Registry mechanism () is the primary mitigation. Agents MUST NOT execute a manifest whose registry lookup returns "black" and SHOULD warn the user before executing an "unknown" manifest.

Interaction with LLM Agent Prompt Injection Defenses Modern Large Language Model (LLM) agents (including but not limited to Claude, ChatGPT, and Gemini) implement prompt-injection defense as an immutable system-level policy. Such policy treats web-page-embedded content that contains executable instructions, including AI Manifests, as untrusted external data and requires explicit user approval prior to execution. Empirical study by the author (2026-04-28) on a representative enterprise multi-step UI (one trial per condition) compared three configurations:

Tier A — Native page without manifest: 97.7 seconds total wall-clock to task completion;
Tier B — Embedded manifest with the LLM agent processing the manifest directly: 154.7 seconds, due to user-approval gate latency. The agent in this trial classified the manifest content as a prompt-injection pattern and reverted to manual page navigation, defeating the manifest's intent;
Tier C — Embedded manifest with a deterministic runtime (SDK) consuming the manifest outside the LLM reasoning loop: 5.0 seconds, with no user-approval interruption.

Implications:

Manifest embedding alone (Method B without a runtime) SHOULD NOT be relied upon as a complete automation surface; in practice it may produce negative value due to repeated user-approval prompts.
Implementations SHOULD pair the manifest with a deterministic runtime that performs discovery, registry verification, static analysis, and step execution outside the LLM context, returning only a structured success/failure summary to the LLM.
Externally-referenced manifests (Method A, /.well-known/) avoid prompt-injection-defense triggering when the runtime fetches the manifest out-of-band, although they require additional HTTP round-trips.

This finding does not alter the protocol on the wire but informs deployment patterns and the role of the runtime component vis-a-vis the LLM agent.

Integrity of the Manifest The SHA-256 hash is computed over the canonical form of the manifest so that semantically equivalent encodings produce identical digests. Implementations MUST NOT rely on a hash computed over non-canonical bytes.

Transport Security All communication with the registry MUST use HTTPS with server authentication per . Registry operators SHOULD sign their responses with a public key published out of band so that an AI agent can verify the integrity of a cached response.

Privacy Considerations Registry lookups necessarily expose to the registry operator the fact that a particular AI agent has visited a particular publisher's manifest. Registry operators SHOULD minimize the retention of client identifiers associated with lookup requests. Agents MAY employ private, time-limited caching of registry responses to reduce the frequency of such lookups.

Implementation Status Note to RFC Editor: This section is intended to be removed prior to publication as an RFC. A reference implementation, including an example publisher server, a reference registry, two AI agent variants (DOM-analysis baseline and manifest-aware), and an automated benchmark harness, is available at under the MIT License. In the reference benchmark — a two-step ERP order-entry transaction repeated 30 times with input tokens counted via the tiktoken cl100k_base encoding — the manifest-aware agent consumed an average of 341 input tokens per task with a 100% task success rate (30 of 30 runs), while the DOM-analysis baseline consumed an average of 1887.6 input tokens with a 20% success rate (6 of 30 runs). Raw results accompany the reference implementation. A subsequent three-tier comparison study (2026-04-28, one trial per condition) used a representative enterprise approval UI ("IRIS-style vacation request") and the same LLM (Claude) across the three configurations described in . Wall-clock results: Tier A (native, no manifest) 97.7 s; Tier B (manifest embedded, processed by the LLM directly) 154.7 s; Tier C (manifest embedded plus deterministic SDK runtime) 5.0 s. Tier C achieved approximately a 19-fold speedup over Tier A and a 31-fold speedup over Tier B. The surprising Tier-B-slower-than-Tier-A result is attributed to the LLM agent rejecting the embedded manifest as a possible prompt-injection pattern and reverting to manual navigation, consistent with LLM safety policy. A controlled experiment (Phase 1, n=30 trials, 2026-04-28) examined whether an AI agent spontaneously detects and loads a site-side AFRM during routine task execution without explicit user instruction. The primary hypothesis (H1: manifest_detected fires at least once during 30 trials) was rejected: the event was never observed. This result confirms that a deterministic SDK runtime, not unaided LLM inference, is required to achieve reliable manifest consumption. All eight pre-registered secondary hypotheses (H2-H8, covering token reduction, task-completion rate, and step-accuracy metrics conditioned on successful detection) were accepted with large effect sizes (Cohen's d = 5.00 for token reduction in the SDK condition). Raw data and analysis scripts accompany the reference implementation. A cross-AI interoperability study (Phase 2, 2026-05-11) tested whether three major LLM agents (Claude, ChatGPT, Gemini) can consume an AFRM presented directly in a conversation context. Results: Claude correctly extracted and applied the AFRM content (trap descriptors, selectors, escape actions) when the manifest was provided in-context; ChatGPT demonstrated label-level recognition (identified the document as an AFRM) but did not apply individual trap entries procedurally; Gemini returned generic speculation without manifest-specific extraction. These results suggest that AFRM consumption at the level of individual knownTraps entries is not yet uniform across LLM agents, further motivating the deterministic SDK runtime architecture specified in this document.

Intellectual Property Rights Disclosure The technology described in this document is the subject of Korean Patent Application No. 10-2026-0071716, filed on 2026-04-21 by the author. The applicant commits to offer any essential claims under Fair, Reasonable, and Non-Discriminatory (FRAND) terms to implementers of this specification, as declared in the project repository.

Acknowledgments The author thanks the Anthropic Claude Code, Model Context Protocol, and OpenAI function-calling communities for the empirical observations that motivated this work.