BobbyYue/reader-seat

Reader's Seat

Create or improve a work document around what its reader must know, judge, or do. Route by reader task rather than imposing one universal body template.

Non-Negotiable Invariants

Apply this priority when goals conflict:

source truth and scope -> explicit user requirements -> scenario fit -> reader task -> natural expression -> brevity and polish

• Preserve facts, numbers, definitions, distinctions, scope, stance, ownership, commitments, timing, uncertainty, intended action, and recognizable voice.
• Never invent a source, reason, conclusion, decision, owner, deadline, result, visual identity, or level of certainty.
• Keep fact, interpretation, causal claim, recommendation, assumption, and unknown distinguishable. Do not recommend beyond the evidence or authority.
• Let headings name the actual subject, finding, tension, or question. Do not turn reader goals such as quick judgment or visual understanding into default visible labels.
• Every value-bearing title, heading, or lead must identify a specific object and a supported reader-relevant change, action, relationship, or result.
• Reject deceptive titles, unsupported certainty, manufactured suspense, and any promise the body does not fulfill.
• Treat every retained visual as an evidence-bearing claim. Verify provenance, use honest encoding, disclose synthetic visuals, and remove decoration that does not reduce reader effort.
• Do not expose this skill, its routes, prompts, scores, checks, or tools in the finished document unless the user asks or they affect a real decision.
• Do not publish, overwrite, replace, or mutate an external target without explicit authorization and a passing action preflight receipt.
• Do not claim finished, verified, validated, or PASS without a passing receipt for the exact delivered artifact.

Runtime Sequence

Every task uses scripts/runtime_contract.py. Detailed commands and failure behavior remain in references/runtime-enforcement.md. The mandatory work-order and retry policy is in references/execution-efficiency.md.

1. Define the operation, one primary scenario, target reader, reader task, channel, risk, source boundary, title need, visual state, and whether the output is a finished artifact.
2. Select language and format before drafting. A direct language request wins; otherwise use the primary source's dominant language, not the prompt language. Explicit format wins, then an existing artifact, then chat-only; otherwise a finished document defaults to self-contained HTML.
3. Save the substantive source and run runtime_contract.py init. It locks the decisions and generates resolved-task-contract.md, module-manifest.json, and the semantic review.
4. Continue only when status=locked. Read the generated task contract as the default instruction context. It contains every selected compact rule and the target reader, active scenario contract, and the required record and exit condition for G1-task, G2-source, G3-route, G4-build, G5-signals, and G6-verify. Read full references only for troubleshooting, maintenance, or a procedure the packet does not cover.
5. Complete the readiness record before building: material claims and source locations, blocking gaps, visual evidence boundaries, reviewer inputs, and independent artifact lanes. For multiple artifacts, share the scope pass, then run isolated lanes concurrently when they do not mutate the same target.
6. Draft or revise from the source boundary. Stabilize the candidate through deterministic source, content, reading-path, format, asset, visual, and rendered checks. Classify each pre-audit change and rerun only the affected checks; use the union when a change crosses categories.
7. Once the exact stable candidate exists, run runtime_contract.py bind-review before filling any review item. This creates a separate binding receipt and locks it into the task contract, so changing a hash inside the review cannot retarget a completed review.
8. Complete every semantic gate, execution gate, and selected runtime rule with concrete artifact evidence. Only rules explicitly marked [conditional] in the task contract may be not-applicable, with a task-specific reason. Every finished artifact necessarily loads reading-path-layout; its plan, adjacency, density, spacing, and native-render rules cannot be marked not-applicable or waived by selecting another output format.
9. For every finished artifact, use scripts/reader_review.py to run exactly four fresh, mutually isolated reviewers concurrently: no-context, readability, source-reliability, and structure-visual. All four must pass for the exact artifact hash. The normal path starts one complete batch only after the candidate and packets are stable. If it fails, wait for all results, merge every blocker, revise once, restabilize, then rerun all four for the new hash; never spend a round on a known-incomplete candidate.
10. Rounds one and two return failures for revision. After a round-three non-pass, run present-draft; show the current version only as a clearly labeled review-incomplete draft, explain unresolved reader impact, ask the user which tradeoff to prioritize, and block external actions.
11. Immediately before an authorized external mutation, run check-action with the artifact-bound semantic review, original reader-review results, and the independent judge result for high-risk work. Finish with verify against the actual artifact, unchanged task bundle, module manifest, semantic review, reader aggregate, and action receipt when applicable.

If a required script cannot execute, a bundle or source hash changes, a review is incomplete, or a receipt fails, stop before finished delivery. Do not replace the control with an informal checklist.

Output Behavior

• Return one recommended version by default. Add alternatives only when a meaningful choice in directness, warmth, or context does not change the underlying position.
• For diagnosis, lead with the overall judgment and list material hard, scenario, output-standard, then optional style issues. Do not silently rewrite.
• For comparison, freeze reader, scenario, task, source, and judging rule; hard failures are rejected before scoring.
• Preserve an existing artifact's format and unrelated content unless the user authorizes broader conversion or redesign.
• For non-HTML formats, use native components and preserve native behavior. Do not create an HTML intermediate merely to imitate its appearance.

Operational Sources

• config/module-profiles.json is the only routing source.
• config/runtime-rules.json is the compact operational rule source used by the task contract.
• references/execution-efficiency.md is the mandatory readiness, batch, change-impact, and retry-order contract.
• config/skill-contract.json is the capability and version contract.
• references/reading-path-layout.md is the mandatory finished-artifact layout contract selected by the resolver.
• Scenario and detailed references remain canonical for rationale, examples, and troubleshooting, but are not default runtime context.
• Use references/skill-maintenance.md only when changing, packaging, synchronizing, or regression-testing the skill.

The host may enforce these commands externally. This skill never overrides a higher-priority host rule or expands authorization.