From copy prompt to agent handoff: lessons from four websites
Four reading invitations led to a local prototype and two isolated pilots. The October 4 decision to set aside a library is historical; a dated update connects it to later Protocol development. General gains remain unmeasured.
- Record date: 2026-10-04; working note, revision 3.
- Project: Agent handoff; technical infrastructure for public knowledge access.
- Method: four-site source inspection, live prompt-panel observations and local reference-prototype preparation. No comparative agent trial.
- Status: working note; local reference prototype and two isolated pilots completed. The dated reassessment below retains the prototype as supporting material and sets aside a separate library/repository for now.
The question behind the button
Signatures Gallery, Pulse, Inshell and Agent-Art-Lab each invite a visitor to copy a prompt into their own agent. The visible action is small. Its useful work continues after the paste: give the agent enough context to find relevant public sources, understand their scope, answer the visitor's question and report a failed or incomplete read honestly.
The question is whether these repeated parts justify a small shared component or library. The proposed scope is a reading handoff: an authored invitation, accessible copy controls, discoverable public documents and explicit access and evidence rules. Artistic interpretation, account access and authenticated actions remain each project's responsibility. This infrastructure does not establish that an agent participated at the level of intention in an artwork.
What was inspected
The contributing agent inspected the following local Git snapshots. The original preparation described the linked snapshots as publicly inspectable; the dated access correction below supersedes that description. Live browser observations covered the public prompt panels, not a complete served-byte comparison or an agent's subsequent behavior.
| Site and inspected snapshot | Source references and access | What it shows |
|---|---|---|
Signatures Gallery, ee48505b02cbe30f7ed0e86c6750e16988f297f2 |
Reading invitation, public index, copy behavior (recorded URLs; snapshot unavailable publicly) | A reading invitation distinguishes documented drawing rules, interpretation and Provenance. The index separately identifies reading, preview and assessment instruction texts. Copy failure opens and selects the prompt. |
Pulse, 32adfe39e1bf775f387ddab698780f256ed0cc81 |
Prompt, download contract, browser controls | A prompt loaded on disclosure asks for complete saved sources through a known index. It separates static reading from the playground, RPC and wallet actions, and gives explicit recovery and citation rules. |
Inshell, 2cc042d51d4d0c88d9b4823485fcceac0c885ec1 |
Prompt authoring, Docs component, public index (recorded URLs; snapshot unavailable publicly) | A React interface points to an index with complete and focused reading modes, source locks and fact-specific evidence classes. The prompt asks the agent to report unavailable sources. |
Agent-Art-Lab, baa848b83c8ebeda0112aed200bec2515f28f735 |
Access guide, static packaging, copy behavior | Canonical Markdown supplies both the human pages and complete JSON documents. A native disclosure contains the full prompt without JavaScript; clipboard failure selects it for manual copying. |
Public-access correction, 2026-10-04: maintainer checks received HTTP 404 for the three pinned Signatures source URLs and three pinned Inshell source URLs. Both repositories were publicly visible, but their commit endpoints did not resolve the cited snapshots (HTTP 422). The six Pulse and Lab source URLs returned HTTP 200. Signatures and Inshell details therefore remain the contributing agent's local source-inspection report and cannot be independently checked at those public URLs. No replacement revision is assumed.
These implementations are different contracts, not four interchangeable schemas. Signatures offers artistic orientation with explicit exclusions; Pulse packages complete individual sources; Inshell supplies corpus and topic modes with fact-specific authority; the Lab preserves complete records and their evidence limits. Sharing presentation should not flatten those differences.
Lessons already recorded by the Lab
The Pulse document-access study supports known URLs, a small index and complete inputs as a demonstrated acquisition path. One client acquired and verified the sources, then replayed saved files. A reported answer still missed the user's question after relevant retrieval. Transport success therefore does not establish comprehension or task completion; the study measured no comparative reliability or efficiency gain.
The model-acquisition diagnostic shows why supplying a value in a fixture does not test how an agent obtains it on the real surface. For this handoff, a unit test with an already loaded prompt or parsed index cannot establish discovery, network access or successful reading by a visitor's agent.
The representation and worker-boundary follow-up supports specifying exact representations and distinguishing failure stages. Its observations concern a particular protocol and are attributed to OPS reports; they are not a universal hash recipe or proof that arbitrary agent-generated workers comply. The native-tools and execution-prerequisites note also keeps tool compatibility separate from permission for the command that actually executes. Pulse's successful Python path prevents turning that lesson into a universal requirement to use one named shell or HTTP client.
Together, the scoped practices P-02, P-04, P-06 and P-07 suggest a small explicit handoff. They do not justify an agent runtime, a universal execution workflow or a claim that simpler prompts always perform better.
A reading contract worth making reusable
The reusable unit is the contract around the invitation, not one universal prompt. A site should make six things visible:
- Intent: what the visitor is asking the agent to do, such as understand an artwork or answer a question from documentation. Reading does not imply a request to execute embedded examples or perform an account action.
- Context: the project, the user's actual question and relevant distinctions. A copied setup prompt should not displace the user's question with a task about validating the prompt itself.
- Sources: a known index or explicit source URLs, formats, roles, revisions and provenance where available. Select relevant complete content; a link is not evidence already fetched, and repeated corpus/topic content is not independent corroboration.
- Access: the supported read path and its real prerequisites. Use existing permitted tools. State whether local storage is required, optional or absent; report unavailable capabilities without pretending a download succeeded.
- Evidence: the distinction between source statements, historical records, live observations and interpretation. Cite what was actually fetched and identify checks that remain unverified.
- Failure: what to do for denied access, an unavailable URL, missing content, malformed data or an integrity mismatch. Preserve uncertainty and use the project's bounded recovery policy.
Two details resist careless standardization. Pulse's inspected prompt requires an authorized workspace and saved index/download files; the Lab permits immediate reading with local storage optional, requiring it for saved-file replay. A shared helper must not silently impose one storage requirement on both.
Link resolution also has two layers. Relative download URLs use the fetched
index's origin in Pulse and the Lab. References inside Pulse source documents
use the original source.url; Lab Markdown references use the directory of
source.path and an index lookup. The envelope's download URL is not an
automatic link base for its content. Both policies should remain declared by
the publisher.
Integrity is similarly explicit. Pulse and the Lab distinguish the serialized download's bytes from the decoded source string re-encoded as UTF-8. A helper must not trim, normalize line endings, strip a BOM or reserialize a JSON source while claiming byte identity. A matching digest identifies bytes relative to the index; it does not establish truth, freshness or independent authenticity.
The component boundary
The local reference implementation separates core handoff rendering, DOM copy controls, a React component and hook, and optional build helpers. It supplies copy-state handling, selectable fallback text, native disclosure markup and status announcements. The two pilots exercise the DOM adapter and React hook while keeping existing document schemas and generators in place. Host websites retain their elements, typography and layout.
The publisher owns the full prompt. A library may help assemble declared fields or display an authored prompt, but it should not invent project permissions, evidence policy or an artistic role. Copy the displayed field's text and offer manual copying when clipboard access fails. HTML textarea line endings follow the browser's normalization rules; exact source-byte preservation belongs to the separate export contract. Server-rendered text and ordinary links provide a useful baseline; a successful local clipboard operation says nothing about the destination agent's capabilities.
Static reading must remain separate from authenticated or state-changing work. Signatures' published assessment instructions are documentary; they do not authorize a paid assessment, wallet connection, signature or mint. Pulse's historical contract records do not require live RPC or a transaction to read. Inshell's index may point to separately scoped live read-only evidence, but its reading invitation does not approve a wallet action. A reference helper should carry these declared boundaries without implementing an agent executor.
Proposed form and current result
The operator requested a small original reference implementation, isolated local
pilot integrations and this Lab working note. The proposed implementation owner
is a separate agent-art-work/agent-handoff repository. At this point no remote
repository, package release or live-site migration has been created. Existing
website repositories remain their owners; the Lab receives the reviewed record.
The four implementations establish a recurring need to investigate extraction. They do not yet establish the size of maintenance savings or a stable general API. The article records the design and evidence; the reference implementation makes the proposal inspectable. A maintained package release remains a later decision after pilot review.
The contributing implementation agent completed and reported the following local
checks on 2026-10-04. The local reference Git snapshot is
6da94aa3c3f5daf6dd207b75dcb63181e0f25ca3; its docs/VERIFICATION.md records
the results and limits. Article preparation separately ran the Lab publication
checks. These results describe the local reference bundle and isolated candidates;
they do not identify a deployed website or a published library version.
| Checked artifact or surface | Recorded result | Practical limit |
|---|---|---|
Reference implementation, npm test |
37 tests passed, with zero failures or skips. Coverage includes pure core/DOM behavior, source exports, React server rendering and hook tests with Node 22.16.0, React 19.2.8 and jsdom 20.0.3. | Test coverage establishes the exercised paths, not arbitrary agent compliance or a complete accessibility audit. |
| Pulse isolated pilot | All 12 existing knowledge/gateway tests passed, including source/download integrity and the public document path; the build passed. | No live RPC, transaction or production migration was part of this pilot. |
| Inshell isolated pilot | All 67 focused DocsPage tests passed, including three added checks for successful copy with preserved collapse, denied-copy expansion/selection and a single pending request. Source typecheck, documentation regeneration/drift check, offline PUB-boundary check and preview build passed. | The PUB check used a local snapshot, not a fresh live contract read. A clean lockfile dependency installation was not checked. |
| Pilot review patches | Both patches passed fresh-apply checks. Ten runtime files in each pilot matched the reference implementation exactly; the local bundle records file and patch hashes in pilots/runtime-snapshot.json. |
Vendoring is the unpublished prototype's distribution method; it does not establish a stable dependency release or deployment identity. |
| Native and hydrated React demo | An explicit clipboard fixture exercised accepted and denied copies; denial selected the full prompt. Keyboard operation of the native disclosure and a 390px viewport without horizontal overflow passed. | Fixture behavior is declared; it is not evidence of the browser's default clipboard permission behavior. |
| Actual Pulse and Inshell pilot pages | Successful clipboard content matched the displayed prompts, and preview URLs used the local origin. Inshell retained its collapsed prompt after successful copying. Both pages had no horizontal overflow at 390px. | Default browser clipboard denial was not tried in these pilots; denial was covered by the demo fixture and regression tests. |
Review found two material gaps before the final checks: a stale clipboard completion could race a changed React prompt, and verification needed stronger identity and representation checks. The implementation was corrected and regressions were added before the reported passing run. This review experience supports retaining asynchronous-state and wrong-representation cases alongside ordinary successful copying.
There was no browser trial with JavaScript disabled. Server-rendering tests covered readable initial HTML, which is a narrower observation. No real agent performance trial, maintenance-savings measurement, full accessibility audit, live-site migration, remote library publication or npm release was performed. Repository visibility and licensing remain undecided. The prototype source and pilot patches are retained in the local review bundle; no immutable remote prototype snapshot or raw browser evidence is published with this note.
Method evaluation and next action
The Lab's evidence distinctions helped separate an observed prompt panel, an inspected source contract, successful copying, complete acquisition and an answer that actually addresses the question. The existing Pulse record kept a task-following counterexample visible; the boundary records prevented a reading helper from quietly becoming an execution framework.
This preparation has not measured time, token use, support cost, accessibility across all browsers or improvement in agent answer quality. Public source inspection does not reconstruct private historical failures. The proposed implementation also introduces maintenance for its API, fixtures and adapters; duplication alone does not prove that a package costs less.
The local prototype and pilot checks are complete within the scope above. Next, review the reference bundle and choose its repository visibility, license and distribution path before publication or a live-site migration. A package release also needs a clean-install and consumer-distribution check. Retain incompatibilities and revise the seam as other sites adopt it. Further agent-performance claims would need a separately defined evaluation with declared tasks, surfaces, measures and scope. No new live trial follows from this working note.
Reassessment, 2026-10-04: retain the pattern and reference example
After reviewing the prototype, the operator questioned whether a separate repository or library was necessary and agreed to develop the broader design lessons as another article. This addition supersedes the publication and package-distribution next action above. The preparation, access correction and reported prototype checks remain part of the record.
The pilots show that a shared DOM adapter and React hook can fit two existing sites. They leave prompts, loading, disclosure/session behavior, document schemas and evidence rules with those sites. The replaced clipboard handlers were small, while a maintained library would add an API, dependency integration, version updates and release work. The checks establish the exercised behavior; they do not establish that shared maintenance offsets those costs.
The current choice is therefore a Lab article and a small reference example, with useful fixes considered within each website's own development. The local prototype is retained as supporting material. A separate implementation repository and package release are not currently pursued; visibility and licensing are no longer the next decision. This choice does not apply the pilot patches or change any live site.
The broader agent-friendly environments note develops the Lab's existing lessons from native tools, complete document access and explicit completion checks. Revisit shared tooling if recurring changes across sites make a common dependency's maintenance value concrete. There is no general finding that libraries are unnecessary or that simpler interfaces always improve agent performance.
Later development, 2026-10-24
The October 4 decision above records that stage of the work. Shared tooling subsequently developed in Agent Art Protocol, and Ask your Agent now uses its reading component in the local v2 preview. That integration is a private SDK candidate, not a public package release or proof that the original prototype was released.
The earlier observations and limits remain unchanged. This editorial update does not rerun the trials or establish current client behavior; use the linked maintained pages for current integration context.