{
  "schema": "agent-art-lab.document/v1",
  "id": "projects-agent-handoff-studies-2026-10-04-copy-prompt-reading-handoff",
  "title": "From copy prompt to agent handoff: lessons from four websites",
  "page": "/projects/agent-handoff/studies/2026-10-04-copy-prompt-reading-handoff.html",
  "revision": "sha256:69a9ecfb3390aa244efd65e30bb4d071a2cc73725276d81f72a6c7dc72f3559c",
  "source": {
    "path": "projects/agent-handoff/studies/2026-10-04-copy-prompt-reading-handoff.md",
    "url": "https://github.com/agent-art-work/agentart.work/blob/main/projects/agent-handoff/studies/2026-10-04-copy-prompt-reading-handoff.md",
    "sha256": "69a9ecfb3390aa244efd65e30bb4d071a2cc73725276d81f72a6c7dc72f3559c",
    "byteLength": 20233
  },
  "study": {
    "project": "Agent handoff",
    "date": "2026-10-04",
    "status": "Working note",
    "evidence": "Source inspection and reported browser/prototype checks; some snapshots unavailable"
  },
  "contentFormat": "markdown",
  "content": "# From copy prompt to agent handoff: lessons from four websites\n\n- Record date: 2026-10-04; working note, revision 3.\n- Project: Agent handoff; technical infrastructure for public knowledge access.\n- Method: four-site source inspection, live prompt-panel observations and local\n  reference-prototype preparation. No comparative agent trial.\n- Status: working note; local reference prototype and two isolated pilots\n  completed. The dated reassessment below retains the prototype as supporting\n  material and sets aside a separate library/repository for now.\n\n## The question behind the button\n\n[Signatures Gallery](https://signatures.gallery/about),\n[Pulse](https://pulse.inshell.art/), [Inshell](https://inshell.art/docs) and\n[Agent-Art-Lab](https://agentart.work/) each invite a visitor to copy a prompt\ninto their own agent. The visible action is small. Its useful work continues\nafter the paste: give the agent enough context to find relevant public sources,\nunderstand their scope, answer the visitor's question and report a failed or\nincomplete read honestly.\n\nThe question is whether these repeated parts justify a small shared component\nor library. The proposed scope is a reading handoff: an authored invitation,\naccessible copy controls, discoverable public documents and explicit access and\nevidence rules. Artistic interpretation, account access and authenticated actions\nremain each project's responsibility. This infrastructure does not establish\nthat an agent participated at the level of intention in an artwork.\n\n## What was inspected\n\nThe contributing agent inspected the following local Git snapshots. The\noriginal preparation described the linked snapshots as publicly inspectable;\nthe dated access correction below supersedes that description. Live browser\nobservations covered the public prompt panels, not a complete served-byte\ncomparison or an agent's subsequent behavior.\n\n| Site and inspected snapshot | Source references and access | What it shows |\n| --- | --- | --- |\n| Signatures Gallery, `ee48505b02cbe30f7ed0e86c6750e16988f297f2` | [Reading invitation][signatures-prompt], [public index][signatures-index], [copy behavior][signatures-copy] (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. |\n| Pulse, `32adfe39e1bf775f387ddab698780f256ed0cc81` | [Prompt][pulse-prompt], [download contract][pulse-format], [browser controls][pulse-client] | 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. |\n| Inshell, `2cc042d51d4d0c88d9b4823485fcceac0c885ec1` | [Prompt authoring][inshell-prompt], [Docs component][inshell-client], [public index][inshell-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. |\n| Agent-Art-Lab, `baa848b83c8ebeda0112aed200bec2515f28f735` | [Access guide][lab-guide], [static packaging][lab-documents], [copy behavior][lab-copy] | 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. |\n\n**Public-access correction, 2026-10-04:** maintainer checks received HTTP 404\nfor the three pinned Signatures source URLs and three pinned Inshell source\nURLs. Both repositories were publicly visible, but their commit endpoints did\nnot resolve the cited snapshots (HTTP 422). The six Pulse and Lab source URLs\nreturned HTTP 200. Signatures and Inshell details therefore remain the\ncontributing agent's local source-inspection report and cannot be independently\nchecked at those public URLs. No replacement revision is assumed.\n\nThese implementations are different contracts, not four interchangeable\nschemas. Signatures offers artistic orientation with explicit exclusions; Pulse\npackages complete individual sources; Inshell supplies corpus and topic modes\nwith fact-specific authority; the Lab preserves complete records and their\nevidence limits. Sharing presentation should not flatten those differences.\n\n## Lessons already recorded by the Lab\n\nThe [Pulse document-access study](../../pulse/studies/2026-09-28-document-access.md)\nsupports known URLs, a small index and complete inputs as a demonstrated\nacquisition path. One client acquired and verified the sources, then replayed\nsaved files. A reported answer still missed the user's question after relevant\nretrieval. Transport success therefore does not establish comprehension or task\ncompletion; the study measured no comparative reliability or efficiency gain.\n\nThe [model-acquisition diagnostic](../../thought/studies/2026-09-20-model-acquisition.md)\nshows why supplying a value in a fixture does not test how an agent obtains it\non the real surface. For this handoff, a unit test with an already loaded prompt\nor parsed index cannot establish discovery, network access or successful reading\nby a visitor's agent.\n\nThe [representation and worker-boundary follow-up](../../thought/studies/2026-09-24-boundaries-and-canary-follow-up.md)\nsupports specifying exact representations and distinguishing failure stages.\nIts observations concern a particular protocol and are attributed to OPS reports;\nthey are not a universal hash recipe or proof that arbitrary agent-generated\nworkers comply. The [native-tools and execution-prerequisites note](../../thought/studies/2026-10-01-native-explicit-execution.md)\nalso keeps tool compatibility separate from permission for the command that\nactually executes. Pulse's successful Python path prevents turning that lesson\ninto a universal requirement to use one named shell or HTTP client.\n\nTogether, the scoped practices [P-02, P-04, P-06 and P-07](../../../findings/REGISTER.md)\nsuggest a small explicit handoff. They do not justify an agent runtime, a\nuniversal execution workflow or a claim that simpler prompts always perform\nbetter.\n\n## A reading contract worth making reusable\n\nThe reusable unit is the contract around the invitation, not one universal\nprompt. A site should make six things visible:\n\n1. **Intent:** what the visitor is asking the agent to do, such as understand an\n   artwork or answer a question from documentation. Reading does not imply a\n   request to execute embedded examples or perform an account action.\n2. **Context:** the project, the user's actual question and relevant distinctions.\n   A copied setup prompt should not displace the user's question with a task\n   about validating the prompt itself.\n3. **Sources:** a known index or explicit source URLs, formats, roles, revisions\n   and provenance where available. Select relevant complete content; a link is\n   not evidence already fetched, and repeated corpus/topic content is not\n   independent corroboration.\n4. **Access:** the supported read path and its real prerequisites. Use existing\n   permitted tools. State whether local storage is required, optional or absent;\n   report unavailable capabilities without pretending a download succeeded.\n5. **Evidence:** the distinction between source statements, historical records,\n   live observations and interpretation. Cite what was actually fetched and\n   identify checks that remain unverified.\n6. **Failure:** what to do for denied access, an unavailable URL, missing content,\n   malformed data or an integrity mismatch. Preserve uncertainty and use the\n   project's bounded recovery policy.\n\nTwo details resist careless standardization. Pulse's inspected prompt requires\nan authorized workspace and saved index/download files; the Lab permits immediate\nreading with local storage optional, requiring it for saved-file replay. A shared\nhelper must not silently impose one storage requirement on both.\n\nLink resolution also has two layers. Relative download URLs use the fetched\nindex's origin in Pulse and the Lab. References inside Pulse source documents\nuse the original `source.url`; Lab Markdown references use the directory of\n`source.path` and an index lookup. The envelope's download URL is not an\nautomatic link base for its content. Both policies should remain declared by\nthe publisher.\n\nIntegrity is similarly explicit. Pulse and the Lab distinguish the serialized\ndownload's bytes from the decoded source string re-encoded as UTF-8. A helper\nmust not trim, normalize line endings, strip a BOM or reserialize a JSON source\nwhile claiming byte identity. A matching digest identifies bytes relative to\nthe index; it does not establish truth, freshness or independent authenticity.\n\n## The component boundary\n\nThe local reference implementation separates core handoff rendering, DOM copy\ncontrols, a React component and hook, and optional build helpers. It supplies\ncopy-state handling, selectable fallback text, native disclosure markup and\nstatus announcements. The two pilots exercise the DOM adapter and React hook\nwhile keeping existing document schemas and generators in place. Host websites\nretain their elements, typography and layout.\n\nThe publisher owns the full prompt. A library may help assemble declared fields\nor display an authored prompt, but it should not invent project permissions,\nevidence policy or an artistic role. Copy the displayed field's text and offer\nmanual copying when clipboard access fails. HTML textarea line endings follow\nthe browser's normalization rules; exact source-byte preservation belongs to the\nseparate export contract. Server-rendered text and ordinary links\nprovide a useful baseline; a successful local clipboard operation says nothing\nabout the destination agent's capabilities.\n\nStatic reading must remain separate from authenticated or state-changing work.\nSignatures' published assessment instructions are documentary; they do not\nauthorize a paid assessment, wallet connection, signature or mint. Pulse's\nhistorical contract records do not require live RPC or a transaction to read.\nInshell's index may point to separately scoped live read-only evidence, but its\nreading invitation does not approve a wallet action. A reference helper should\ncarry these declared boundaries without implementing an agent executor.\n\n## Proposed form and current result\n\nThe operator requested a small original reference implementation, isolated local\npilot integrations and this Lab working note. The proposed implementation owner\nis a separate `agent-art-work/agent-handoff` repository. At this point no remote\nrepository, package release or live-site migration has been created. Existing\nwebsite repositories remain their owners; the Lab receives the reviewed record.\n\nThe four implementations establish a recurring need to investigate extraction.\nThey do not yet establish the size of maintenance savings or a stable general\nAPI. The article records the design and evidence; the reference implementation\nmakes the proposal inspectable. A maintained package release remains a later\ndecision after pilot review.\n\nThe contributing implementation agent completed and reported the following local\nchecks on 2026-10-04. The local reference Git snapshot is\n`6da94aa3c3f5daf6dd207b75dcb63181e0f25ca3`; its `docs/VERIFICATION.md` records\nthe results and limits. Article preparation separately ran the Lab publication\nchecks. These results describe the local reference bundle and isolated candidates;\nthey do not identify a deployed website or a published library version.\n\n| Checked artifact or surface | Recorded result | Practical limit |\n| --- | --- | --- |\n| 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. |\n| 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. |\n| 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. |\n| 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. |\n| 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. |\n| 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. |\n\nReview found two material gaps before the final checks: a stale clipboard\ncompletion could race a changed React prompt, and verification needed stronger\nidentity and representation checks. The implementation was corrected and\nregressions were added before the reported passing run. This review experience\nsupports retaining asynchronous-state and wrong-representation cases alongside\nordinary successful copying.\n\nThere was no browser trial with JavaScript disabled. Server-rendering tests\ncovered readable initial HTML, which is a narrower observation. No real agent\nperformance trial, maintenance-savings measurement, full accessibility audit,\nlive-site migration, remote library publication or npm release was performed.\nRepository visibility and licensing remain undecided. The prototype source and\npilot patches are retained in the local review bundle; no immutable remote\nprototype snapshot or raw browser evidence is published with this note.\n\n## Method evaluation and next action\n\nThe Lab's evidence distinctions helped separate an observed prompt panel, an\ninspected source contract, successful copying, complete acquisition and an\nanswer that actually addresses the question. The existing Pulse record kept a\ntask-following counterexample visible; the boundary records prevented a reading\nhelper from quietly becoming an execution framework.\n\nThis preparation has not measured time, token use, support cost, accessibility\nacross all browsers or improvement in agent answer quality. Public source\ninspection does not reconstruct private historical failures. The proposed\nimplementation also introduces maintenance for its API, fixtures and adapters;\nduplication alone does not prove that a package costs less.\n\nThe local prototype and pilot checks are complete within the scope above. Next,\nreview the reference bundle and choose its repository visibility, license and\ndistribution path before publication or a live-site migration. A package release\nalso needs a clean-install and consumer-distribution check. Retain incompatibilities\nand revise the seam as other sites adopt it. Further agent-performance claims\nwould need a separately defined evaluation with declared tasks, surfaces,\nmeasures and scope. No new live trial follows from this working note.\n\n## Reassessment, 2026-10-04: retain the pattern and reference example\n\nAfter reviewing the prototype, the operator questioned whether a separate\nrepository or library was necessary and agreed to develop the broader design\nlessons as another article. This addition supersedes the publication and\npackage-distribution next action above. The preparation, access correction and\nreported prototype checks remain part of the record.\n\nThe pilots show that a shared DOM adapter and React hook can fit two existing\nsites. They leave prompts, loading, disclosure/session behavior, document\nschemas and evidence rules with those sites. The replaced clipboard handlers\nwere small, while a maintained library would add an API, dependency integration,\nversion updates and release work. The checks establish the exercised behavior;\nthey do not establish that shared maintenance offsets those costs.\n\nThe current choice is therefore a Lab article and a small reference example,\nwith useful fixes considered within each website's own development. The local\nprototype is retained as supporting material. A separate implementation\nrepository and package release are not currently pursued; visibility and\nlicensing are no longer the next decision. This choice does not apply the pilot\npatches or change any live site.\n\nThe broader [agent-friendly environments note](2026-10-04-agent-friendly-environments.md)\ndevelops the Lab's existing lessons from native tools, complete document access\nand explicit completion checks. Revisit shared tooling if recurring changes\nacross sites make a common dependency's maintenance value concrete. There is\nno general finding that libraries are unnecessary or that simpler interfaces\nalways improve agent performance.\n\n[signatures-prompt]: https://github.com/agent-art-work/Agent-Art-signatures.gallery/blob/ee48505b02cbe30f7ed0e86c6750e16988f297f2/src/openMint/aboutReadingPrompt.ts\n[signatures-index]: https://github.com/agent-art-work/Agent-Art-signatures.gallery/blob/ee48505b02cbe30f7ed0e86c6750e16988f297f2/src/openMint/agentDocuments.ts\n[signatures-copy]: https://github.com/agent-art-work/Agent-Art-signatures.gallery/blob/ee48505b02cbe30f7ed0e86c6750e16988f297f2/src/openMint/promptCopy.ts\n[pulse-prompt]: https://github.com/inshell-art/pulse/blob/32adfe39e1bf775f387ddab698780f256ed0cc81/evm/playground/docs/agent-prompt.txt\n[pulse-format]: https://github.com/inshell-art/pulse/blob/32adfe39e1bf775f387ddab698780f256ed0cc81/evm/playground/docs/document-format.md\n[pulse-client]: https://github.com/inshell-art/pulse/blob/32adfe39e1bf775f387ddab698780f256ed0cc81/evm/playground/agent.js\n[inshell-prompt]: https://github.com/inshell-art/inshell.art/blob/2cc042d51d4d0c88d9b4823485fcceac0c885ec1/apps/h%6fme/src/content/docs.ts\n[inshell-client]: https://github.com/inshell-art/inshell.art/blob/2cc042d51d4d0c88d9b4823485fcceac0c885ec1/apps/h%6fme/src/components/DocsPage.tsx\n[inshell-index]: https://github.com/inshell-art/inshell.art/blob/2cc042d51d4d0c88d9b4823485fcceac0c885ec1/apps/h%6fme/public/docs/agent-index.json\n[lab-guide]: https://github.com/agent-art-work/Agent-Art-Lab/blob/baa848b83c8ebeda0112aed200bec2515f28f735/docs/AGENT_ACCESS.md\n[lab-documents]: https://github.com/agent-art-work/Agent-Art-Lab/blob/baa848b83c8ebeda0112aed200bec2515f28f735/scripts/agent-documents.mjs\n[lab-copy]: https://github.com/agent-art-work/Agent-Art-Lab/blob/baa848b83c8ebeda0112aed200bec2515f28f735/site/assets/copy-prompt.js\n\n## Later development, 2026-10-24\n\nThe October 4 decision above records that stage of the work. Shared tooling\nsubsequently developed in [Agent Art Protocol](../../../docs/PROTOCOL.md), and\n[Ask your Agent](../../../docs/ASK_YOUR_AGENT.md) now uses its reading component\nin the local v2 preview. That integration is a private SDK candidate, not a\npublic package release or proof that the original prototype was released.\n\nThe earlier observations and limits remain unchanged. This editorial update\ndoes not rerun the trials or establish current client behavior; use the linked\nmaintained pages for current integration context.\n",
  "assets": []
}
