Fable Brain

Export Chapter Computer Use and Search

6 min read

Export Chapter Computer Use and Search

Lines 1027-1692 of the export mandate SKILL.md-first computer use, a three-directory file scheme, 20-line artifact thresholds, a visual-output routing checklist, effort-scaled search, and copyright hard limits of one sub-15-word quote per source.

What it is

This note digests the third chapter of the System Prompt Export 2026-07, spanning L1027-1692: computer_use (skills, file creation, file handling, outputs, sharing, artifact criteria, package management), the request_evaluation_checklist and Visualizer sections, search_instructions with CRITICAL_COPYRIGHT_COMPLIANCE, harmful_content_safety, and using_image_search_tool. It governs how the claude.ai harness produces files, visuals, and searched answers.

How it works

Claim list with line refs:

  • Skills are mandatory reading: “Reading the relevant SKILL.md is a required first step before writing any code” (System Prompt Export 2026-07, L1033), because skills encode environment constraints missing from training data; several may apply at once (L1031-1033). The check is unconditional, and task-to-skill mapping is spelled out: pptx, xlsx, docx, pdf (not pypdf), frontend-design (L1192).
  • The computer is a Linux box (Ubuntu 24) with bash, str_replace, create_file, and view; the filesystem resets between tasks (L1067-1069).
  • File locations: uploads at /mnt/user-data/uploads, scratch at the container scratch directory which users cannot see, and deliverables copied to /mnt/user-data/outputs, the only path users see (L1076-1079). Files under 100 lines are written in one call; longer builds go outline first (L1094-1095). Sharing requires present_files with no long postamble (L1102).
  • Artifact thresholds: any code over 20 lines, standalone documents over 20 lines or 1500 characters, long-form creative writing, and reusable content become artifacts; lists, tables, and enumerated content never do, regardless of length (L1121-1136). Verbatim: “A standalone text-heavy document >20 lines or >1500 characters” (System Prompt Export 2026-07, L1128). Only .md, .html, .jsx, .mermaid, .svg, and .pdf render specially (L1140).
  • React artifacts get a pinned library list (lucide-react 0.383.0, recharts, d3, three r128 with named geometry caveats, shadcn/ui, and more) and Tailwind core classes only (L1150-1160). HTML artifacts may import external scripts solely from cdnjs.cloudflare.com (L1147).
  • Browser storage is banned: “NEVER use localStorage, sessionStorage, or ANY browser storage APIs in artifacts” (System Prompt Export 2026-07, L1163); state lives in React state or in-memory variables, with an explain-and-offer-alternative exception on explicit request (L1162-1164).
  • Package management: npm works normally with globals under the container user npm-global directory; pip always needs --break-system-packages (L1172-1173).
  • Visual output routing walks a fixed ladder: prose if no visual is needed (L1203), a category-matching connected MCP tool next, with subdividing categories to dodge a tool called out as rationalization (L1206-1208), file tools if a file was requested (L1214-1215), and the inline Visualizer last (L1217-1218). Routing is never narrated (L1220). Specification-style noun phrases like “comparison table” count as visual requests without any verb (L1238).
  • Search scales with complexity: 1 call for single facts, 3-5 medium, 5-10 deeper, and 20-plus redirects to the Research feature (L1303). Queries stay 1-6 words with no operators (L1314-1318).
  • The unrecognized entity rule makes search non-negotiable for any game, film, show, book, album, product, or event Claude cannot place, including opinion questions (L1299). Knowledge-cutoff disclaimers are banned as annoying (L1301).
  • Copyright hard limits: “15+ words from any single source is a SEVERE VIOLATION” (System Prompt Export 2026-07, L1280), one quote per source maximum after which the source is closed (L1374-1377), and lyrics, poems, and haikus are never reproduced in any form (L1354). Displacive summaries over 30 words and structure-mirroring are also banned (L1356-1357).
  • File-creation triggers separate standalone artifacts from conversational answers: a blog post, story, or article is always a file, while strategies, summaries, and outlines stay inline; tone and length do not change the bucket (L1059). Anything over 10 lines of code becomes a file (L1057), and docx is reserved for explicit Word requests because it costs far more time and tokens (L1061).
  • Worked routing decisions at L1181-1187 show the boundary cases: attached-file summaries stay in conversation, knowledge questions take no tools, and web-search answers stay conversational without report-style headers (L1186).
  • Harmful-content search rules ban citing extremist sources like the 88 Precepts and locating harmful archives (L1584-1585). Image search wants 3-6 word queries, 3-4 images, interleaved with prose, never ending a response (L1652-1658), with a long blocked-category list including copyrighted characters and celebrity photos (L1636-1646).

Best practice

  • Adopt SKILL.md-first as a general agent pattern: read the relevant skill before any file-producing step, even on familiar formats (L1031-1033). EVIDENCE-BASED
  • Deliver files only through the outputs directory plus present_files; anything left in scratch is invisible to users (L1076-1113). EVIDENCE-BASED
  • Apply the 20-line and 1500-character thresholds when deciding chat versus artifact in claude.ai-style UIs (L1121-1136). EVIDENCE-BASED
  • Enforce the one-quote, sub-15-word rule in every derived document, including this vault, since it applies to artifacts too (L1279-1284). EVIDENCE-BASED
  • Reuse the effort-scaling table (1, 3-5, 5-10, redirect at 20+) as a budget heuristic for your own research agents. PRACTITIONER

Pitfalls

  • Writing deliverables to the scratch directory and assuming the user sees them; only /mnt/user-data/outputs renders (L1079).
  • Putting lists or tables in artifacts; the export forbids it at any length (L1133).
  • Using localStorage in artifacts, which fails inside claude.ai (L1162-1163).
  • Treating these copyright limits as legal advice; the export itself bars fair-use determinations (L1355).
  • Assuming this chapter binds Claude Code; it describes the claude.ai container, a different surface.

Sources

  • System Prompt Export 2026-07, L1027-1692 (captured 2026-07-07).
  • https://docs.claude.com (mandated product-docs search target per export L29; retrieved 2026-07-07).
  • https://cdnjs.cloudflare.com (sole permitted external script host per export L1147; retrieved 2026-07-07).
  • Claim-pack verification of thresholds and copyright limits, see Research Pack 2026-07-07 (2026-07-07).

Next actions

  • Verify the React library pin list against a live artifact session before relying on versions.
  • Diff the copyright block against the next capture; its numbered limits are likely to move.
  • Probe whether the Visualizer ladder holds when both an MCP diagram tool and a file request coexist.

Graph View

Backlinks