3080 brief

3080 Brief

30-second judgment, 80% in one picture

Create a new reader-first decision brief from a source document. Match the source format by default and keep the source unchanged.

Action-First Contract

For an explicit creation request, start tool work in the same turn. Do not end a turn with an acknowledgement, plan, status recap, capability description, or promise to continue. The first externally observable action must be the required dependency diagnostic when readiness is unknown; otherwise it must be the source fetch. Only a required approval or genuinely blocking clarification may interrupt execution. After either is resolved, resume immediately without asking whether to continue. Keep working until an output link/path exists or an exact blocker with its concrete next action is returned.

Output Contract

Open with TLDR, or a short source-language equivalent explicitly requested by the user. It contains exactly three core units:

1. 一句话 / one-sentence judgment: the first line carries exactly one highest-level judgment; 1–3 short lines may follow only as evidence, action, or boundary support. Together they form one opening unit that works within 30 seconds.
2. 一张图 / one-picture summary: one auditable visual covering at least 80% of value-weighted non-appendix claims; Feishu/Lark uses an editable whiteboard.
3. 一个表 / one key-question table: 3–5 source-grounded questions using 问题 / 结论 / 为什么 or Question / Conclusion / Why.

Title the new artifact 3080 Brief｜原文标题. Put a compact source link near the top without a standalone source heading.

Core Contract

Apply this priority when goals conflict: source truth → explicit user requirements → reader comprehension → 30-second judgment → one-picture 80% understanding → visual polish.

• Create a new output. Never edit, overwrite, reorder, comment on, clean up, rename, move, or change permissions on the source.
• Match output type to input type unless explicitly overridden: Feishu/Lark → new Feishu/Lark doc; Word/docx → new .docx; Markdown → new Markdown. Ask only when type/support is genuinely ambiguous.
• Follow the source primary language unless the user explicitly names another output language. Request/conversation/interface/locale language is not override evidence. Apply the selected language to title, TLDR, table, body, and visual.
• Exclude source appendix / 附录 / Appendix content unless the user explicitly includes it.
• Source-back every number, conclusion, risk, and recommendation. Preserve denominator, period, sample, scope, significance, confidence, and metric definition when present. Mark unsupported gaps 原文未提供 / not provided in the source or 推断 / inference.
• Rebuild the narrative around what readers must understand, trust, decide, or do. Put decision implication before implementation detail unless implementers are the explicit audience; do not impose a fixed body section order.
• Keep prompts, process rationale, style names, tool notes, placeholders, and internal method language out of the artifact. Machine-enforced restrictions and semantic colors live in config/3080-brief.json.
• Keep generated time, owner, version, location, permissions, and sharing defaults unless the user requests or operations require a change.
• Use lark-doc for Feishu docs, documents:documents for Word, lark-whiteboard for live Feishu preview, and beautiful-feishu-whiteboard for Feishu style selection.

Conditional Routing

Do not preload references. Read a resource only when its condition becomes true.

Condition	Read / run
Output type is ambiguous or conversion is requested	references/output-format-rules.md
Feishu output is selected, or dependency status is not PASS	Run scripts/check_dependencies.py --mode feishu --json; add --host-capability lark-doc and/or --host-capability lark-whiteboard only after confirming each workflow is actually callable; then read references/dependency-and-installation.md only for BLOCKED/SKIP/FAIL
Building evidence artifacts	references/source-inventory-template.md and references/claim-ledger.schema.json
Data, experiment, causal, metric, or risk evidence is material/unclear	references/evidence-and-risk-rules.md
Narrative is unclear, jargon-heavy, or fails readability	references/reader-optimization.md; add references/expression-anti-patterns.md only after a wording/expression signal
Planning the one-picture summary	references/visual-spec.schema.json and references/visual-pattern-library.md
Directional values/statuses appear	references/semantic-color-system.md
Building/debugging a Feishu SVG	references/whiteboard-patterns.md
Considering bitmap/image generation	references/image2-auxiliary-rules.md before sending any content
Preparing audit review	references/review-packet-template.md and references/review-loop.md
All three audit reviewers pass	references/blind-reader-replay.md
Creating/debugging Feishu XML or final Feishu output	references/feishu-doc-output.md

Runtime State Machine

1. Gate, Then Fetch

Determine source/output type, constraints, and explicit language override. For Feishu, run the dependency diagnostic first and declare lark-doc / lark-whiteboard host capabilities only when their workflows are actually callable. If anything is missing, display the complete approval_bundle, ask once to install or enable all listed items, execute the approved host actions, reload when required, and rerun the diagnostic. Keep Node.js installation without an exact command plus Feishu authentication/permissions as separate approvals. A document request is not installation approval. Non-Feishu work never prompts for Feishu dependencies.

Once ready, fetch the source immediately. A Skill specification, plan, or acknowledgement is not a read result. If no document content returns, report BLOCKED: lark-doc runtime capability unavailable plus the concrete host action; never ask a generic “continue?” question. Inventory embedded sheets, Base, images, charts, and whiteboards; inspect every object carrying P0/P1 evidence or chartable data. Build source_inventory.md and claim_ledger.json, including source/output language, excluded appendix, stable claim IDs, source locations, risks, actions, board/body mappings, and omission reasons.

2. Draft And Preflight

Ask at most three questions only when ambiguity can change the main conclusion, metric meaning/scope, risk boundary, reader, next action, or language decision. Otherwise continue with explicit missing-source/inference labels. Rebuild the narrative around the reader decision path; use short judgment headings, the required Pyramid opening, and the key-question table.

Create visual_spec.json before drawing and map each block to claim IDs. Preserve P0 claims and required weighted coverage. Use real quantitative encoding when the source supports it; otherwise show a truthful boundary. For Feishu, select an allowed beautiful-feishu-whiteboard style, use editable native-shape SVG, preserve semantic meaning beyond color, and validate the board.

Run and fix all deterministic gates before creating the output:

scripts/preflight_check.py DRAFT --source-inventory source_inventory.md --claim-ledger claim_ledger.json
scripts/check_coverage.py claim_ledger.json
scripts/validate_visual_spec.py visual_spec.json claim_ledger.json

Validate the rendered visual for its target format; for Feishu use scripts/validate_whiteboard.sh. Fix deterministic failures before review.

3. Create The Review Draft

Immediately after deterministic preflight passes, create the new output and store its URL/path; do not wait for independent reviewers. Write a complete review draft, not an empty placeholder. Feishu/Lark input must create a native Feishu/Lark document with an editable whiteboard; never silently substitute .docx unless the user explicitly requests Word. Keep the source unchanged. All later revisions update only this generated draft.

4. Review Without Stalling

When independent reviewer execution is available, build the three hash-locked role packets, run reviewers independently, aggregate matching artifact hashes, revise on any blocker, and rerun all three up to three rounds. After PASS, run Blind Reader Replay from Primary and escalate only under its configured conditions.

When independent reviewers are unavailable, do not stop or pretend they ran. Perform three sequential, role-separated self-checks using the same Reader, Source, and Visualization gates; fix every deterministic, grounding, or comprehension blocker; set review_status=LIMITED and blind_reader_status=UNAVAILABLE; then continue unless the user explicitly required independent review. This fallback may reduce validation strength but must not turn a document-creation request into a plan-only response.

5. Update, Verify, Deliver

Apply review fixes to the generated draft, never the source. For Feishu, inspect the live board for clipping, overlap, overflow, and stale rendering. Verify source unchanged, native output type, language, citation, TLDR units, coverage, semantic consistency, accessibility, and final artifact hashes with scripts/verify_reviewed_artifacts.py. A run is complete only when the new URL/path is accessible and distinct from the source. Progress text, Skill instructions, plans, acknowledgements, or promises to continue are never delivery; keep using tools or return the exact blocker.

Delivery

Respond briefly with:

• generated output link or absolute file path;
• source link or absolute file path;
• verification notes covering source unchanged, format/language decision, dependency status when applicable, clarification result, three-review status or blocker, blind-reader status or limitation, visual validation, and Feishu live-preview inspection when applicable.

Never report success without the generated output link or absolute file path. For a blocked Feishu run, name the missing executable capability and the concrete host enablement/authentication action; do not present loaded Skill instructions as output.

If lark-cli returns _notice.update, mention lark-cli update only after completing the task.