anlin-writing

Generate, repair, and evaluate complete titled Chinese prose for anonymous blind evaluation. The target is narrow: make generated text less stably identifiable as generated under documented blind-review conditions. Report only test conditions, sample size, recognition/pass rates, false-accusation rates, and limitations. Do not claim real authorship, provenance, or indistinguishability.

Generated prose must not mention process labels such as 仿写, AI, 生成, 验证通过, 38篇语料, corpus, blind test, checker, or methodology. Controller reports may discuss methods; article bodies may not.

This skill is self-contained at runtime. Do not depend on private external anti-slop, author-style, or personal writing skills. Anti-AI guidance lives in this skill's own references and checkers.

Artifact-Backed Entry Contract

When this skill triggers for article generation, do not compose the article directly in the final chat response. The first prose artifact must be a real draft.md in the current task or evaluation workspace. A terminal-only article is a failed run because the controller cannot inspect route, first saved shape, checker state, or repair history.

Before writing any article sentence, make the routing action visible: in a possible formal/eval workspace, check .anlin-clean-eval-mode and the current directory first; outside clean-eval, choose an ordinary task workspace for draft.md. "Output prose only" means output only the final article after the artifact-backed draft/check path, not bypassing draft.md.

Mode Router

Choose the smallest mode that matches the current task:

• Clean-eval mode has priority over ordinary article wording: .anlin-clean-eval-mode exists, or the controller is measuring natural first-draft guidance. Check the marker with a direct command such as Test-Path -LiteralPath ".anlin-clean-eval-mode" or Get-ChildItem -Force .anlin-clean-eval-mode; a directory listing that merely shows the filename is not a marker check. The marker check and cwd confirmation are two separate tool actions: after checking .anlin-clean-eval-mode, the next tool action must be standalone Get-Location / pwd, before deciding ordinary mode. Both must be visible before any reference read, glob/path probe, or draft write; do not use controller --dir, a glob title, or an absolute path as cwd evidence. A present marker overrides "write an article", "give me prose", "save a draft", or other ordinary-user phrasing. Load only references/clean-eval-first-draft-minimum.md; for standard diary also load references/anlin-collage-source-model.md (replaces the old standard-diary-source-engine.md). Write one complete draft.md, then use scripts/clean_run_checker.py for at most two actual checker calls; preflight attempts do not consume that call budget and may continue until an explicit CLEAN_RUN_PREFLIGHT_STOP. Do not call the normal checker in the bounded case directory.
• Ordinary user mode: use only after clean-eval mode has been ruled out. The user wants an article, revision, or saved prose outside a bounded clean-eval workspace. Ask only for missing facts that materially change the result. Use references/runtime-brief.md for drafting/repair, then run the normal checker if validation is needed.
• Finalized repair mode: the controller copied a bounded draft into a finalized/ directory. If repair-brief.txt declares either compact mode, read only draft.md and repair-brief.txt.
  • repair_mode: hard_pass_review_in_place and repair_mode: source_rewrite_compact are self-contained; do not load references/finalized-repair-minimum.md.
  • For a missing/invalid brief or controller_tool_error, use references/finalized-repair-minimum.md as the fallback artifact contract.
  • In source_rewrite_compact, follow the one fragment focus and exact scope supplied by the brief. Most routes replace one existing fragment relation; a severely underbuilt route may explicitly require a source-level rebuild across the existing article. Do not reconstruct hidden blocker/profile lists or add an independent scene/packet. Make the one complete write to draft.md, then stop. If a chat reply is required, output exactly artifact_written.
• Controller validation mode: after an artifact exists, run deterministic gates, style-profile audit, corpus comparison, blind rounds, and placebo calibration. Use references/validation-protocol.md.

Clean-eval and finalized repair are separate checkpoints. A finalized pass does not retroactively make the bounded first draft successful. A finalized review or fail means the final article is still unresolved and is not ready for blind rounds.

Artifact Locations

Do not write generated articles into the skill directory.

• If the user only wants prose in chat, use temporary draft.md in the current task working directory, validate or repair there, then output only the article. Do not skip the temporary artifact just because the final reply should contain prose only.
• If the user asks to save a file and gives no output directory, ask once for the output directory. If there is no answer, save in the current task working directory, not inside <skill-dir>.
• In tests and blind review, use an external evaluation workspace such as <eval-workspace>/iteration-<n>/eval-<id>/draft.md; evals/ defines prompts and protocols, not output storage.
• In clean-eval and finalized repair, write/read the article using relative draft.md or .\draft.md in the current case directory. Do not write <skill-dir>/.../draft.md, do not reuse a prior run path, and do not invent an absolute artifact path unless ordinary user mode explicitly requested that path.

<skill-dir> is only for resolving bundled references and scripts. If the current directory is the skill directory, switch to an external task/eval workspace before writing prose.

Clean-Eval Entry

For formal clean-eval generation, the first draft is intentionally under-instructed:

1. Check .anlin-clean-eval-mode as one direct tool action, preferably Test-Path -LiteralPath ".anlin-clean-eval-mode" (a directory listing that merely shows the filename is not a marker check).
2. As the next tool action, run standalone Get-Location / pwd and confirm the external case workspace.
3. Only after both checks are visible, read references/clean-eval-first-draft-minimum.md.
4. If the selected genre is standard diary/日寄, read references/anlin-collage-source-model.md (replaces the old standard-diary-source-engine.md which is preserved for reference).
5. Write one complete titled article to draft.md.
6. Run python <skill-dir>/scripts/clean_run_checker.py draft.md --strict --draft-gate --generator-facing --genre <selected-genre> when genre is known.

After the minimal source reads finish, the next tool action must be one complete write to relative draft.md or .\draft.md. Do not spend another model turn comparing openings, rehearsing possible hinges, restating the prompt, or drafting scratch prose in reasoning. Choose the first workable fragment slate and persist it; an imperfect artifact can be checked, but an unwritten better plan is not evidence. For standard diary, several independent thought-turns should normally appear; one continuous scene is not the default.

Before the first complete draft.md, do not open long repair, validation, judge, style-ratio, corpus-card, or checker-source files. In particular, do not load references/clean-generation-brief.md, references/runtime-brief.md, references/anti-ai-slop.md, references/generation-modes.md, references/review-rubric.md, references/writing-checklist.md, references/self-check.md, references/stylometric-ratio-protocol.md, references/validation-protocol.md, references/development-log.md, evals/README.md, or references/corpus-cards/.

If the wrapper reports CLEAN_RUN_PREFLIGHT or CLEAN_RUN_POSTCHECK_PREFLIGHT, the wrapper output is the complete repair interface. Use the --generator-facing interface: exact counts and thresholds stay in controller state, while the generator receives only qualitative findings and one next action. The wrapper does not rewrite draft.md; it records byte-preserving snapshots and prints the next explicit action. In bounded clean-eval, do not load references/clean-generation-brief.md or another repair/reference file after that result. For source/content findings, treat multiple messages as one diagnosis of the earliest broken fragment relation: preserve or rebuild complete genre-appropriate mass, replace that relation once, and do not solve one item per metric. When the action says whole_source_rebuild, it is source-only: restore a day-shaped collage with independent thought-turns while preserving existing source mass; do not shrink it to a short premise summary, then call the wrapper again without running a rhythm script in the same action. A later wrapper call may name a rhythm action once the source is no longer severely underbuilt. If the output contains only shape findings, do not add material; perform only the named local rhythm action. After any draft.md rewrite, rerun the wrapper immediately; if the wrapper explicitly names a rhythm script, run that exact in-place final mutation command, do not read its stdout or write draft.md afterward, and rerun the wrapper immediately. Do not hand-count rows, characters, punctuation, or connectors between calls, and do not run homemade counters or other commands between the write and the rerun. The post-check variant still protects checker call 2/2 and does not consume it. If the wrapper reports CLEAN_RUN_PREFLIGHT_STOP or CLEAN_RUN_STOP, stop repairing, read draft.md once, and output that exact article. Do not switch to ordinary checker flow after the bounded stop.

Finalized Repair Entry

Finalized repair is artifact-only. repair-brief.txt is the generator-facing interface; full style-profile output remains controller evidence after the artifact is frozen.

When the brief declares repair_mode: hard_pass_review_in_place or repair_mode: source_rewrite_compact, it is the complete repair interface: read only draft.md and repair-brief.txt, and do not load references/finalized-repair-minimum.md. The source-compact brief intentionally hides the detailed hard/profile findings; follow its single source_focus, route-appropriate scope_contract (local cluster or explicit source-level rebuild), mass_contract, and shape_contract without turning them into a checklist. For a missing/invalid brief or controller_tool_error, follow the finalized minimum path instead.

During the repair attempt, do not run check_anlin_violations.py, check_style_profile.py, clean_run_checker.py, prepare_finalized_repair_brief.py, local counters, python -c, Measure-Object, wc, path probes, source/test reads, threshold searches, old-log searches, or checker/controller helpers. Only a schema-valid controller result with return code 1 is a normal not-pass signal. controller_tool_error, invalid output, or any other unexpected result is unavailable, not quality evidence.

The first write to draft.md must be the final complete revised article. Do not write an unchanged placeholder, then a second version. Do not print a final article only to terminal/chat. Do not patch after writing. Controller validation reruns the strict hard gate and full style-profile report afterward.

Reference Routing

Use these files by task; do not load everything.

Background is a contradiction boundary, not a content quota. Do not add 云南、王者、狗哥、痛风、外卖、知乎, AI/GPT, cities, districts, game ranks, office facts, or platform details merely because a reference permits them. Game may appear when the prompt, action, memory trigger, social wound, practical delay, or cognitive turn needs it; it is optional and should stay low-specificity unless supported.

Do not use corpus ratio targets as a pre-draft recipe. Style-profile and predictive intervals are post-draft audit tools. If a ratio drifts, inspect the earliest broken fragment relation, voice signal, prompt surface, title source, fact specificity, or visible shape; do not add a feature merely to move a ratio.

Runtime Guardrails

• Start from a fragment, thought, voice turn, or ordinary friction, not a checklist or prompt inventory.
• Complete article means title plus body; every test draft must include a title.
• Ask at most one intake round for missing information that materially affects quality, such as target date/period, usable day material, genre, output destination, and whether low-confidence projection or lookup is allowed. Multiple short questions can be in that one round; skip questions already answered by the user.
• If information is missing and the user does not answer, lower specificity: avoid unsupported city/district, company, route, current events, game tactics, office biography, spouse/child/family-status claims, and concrete platform facts.
• Do not satisfy every style feature. Use a few consequential features deeply rather than all labels shallowly.
• Remove high-risk AI surfaces before output: 不是X，是Y / 不是X，而是Y, blog-like explanation, therapeutic humanizer phrases, prompt-shaped titles, neat moral closure, long prompt inventories, —— feeling captions, unsupported specifics, and visible ratio/checker language.
• For standard diary, let rhythm follow thought, speech, memory, association, body, object, or action movement; avoid prose-block compression, tiny-line grids, period-row grids, and comma carpets without prescribing a fixed page shape.
• A thought-turn is not a paragraph: when an observation, crooked read, and next move occur inside one fragment, write their breathing rows before the first artifact; a rhythm script cannot invent missing source movement.
• A complete standard diary is a day-shaped collage, not a premise summary: if the body only has the prompt surface, one memory, and a final reply, continue through another independent thought-turn before the required residue.
• For linear social/event prompts, treat the prompt sequence as a fact constraint, not the article's outline: start with a non-prompt fragment, crop the event to one needed surface, leave it for an independent thought-turn, and return to any required reply only as a short residue.
• For social-decline prompts, treat the invitation or refusal as one possible fragment. Reject a single carrier chain from present desk/room through message, ticket/money, old-friend inventory, work excuse, refusal, and friendship thesis; replace one middle fragment with a lateral thought instead. Do not invent a consequence chain, group-chat crowd, work biography, or etiquette closure merely to prove the prompt.
• If the prompt only says 忙项目, keep it as an excuse surface; do not invent a client, deadline, leave, team, coworker, city, or office biography.
• For sincere/micro-hope short forms, keep uneven clusters, present practical action, awkward reply/logistics, and loose factual retreat. Do not preserve every prompt prop or expand into standard diary bulk.
• Short-genre profile fallback is not strong evidence. If the style-profile shows inconclusive due to small corpus, do not force standard-diary rhythm, length, or texture to satisfy the fallback. A correct short-genre draft may flag review under global priors; that is a corpus limitation, not a draft failure. Do not add background, game, body, or platform packets to compensate for an unstable short-genre profile.

Output Rules

Checker findings are not tool failures and are not a second writing interface. A warning or ordinary error means the persisted artifact has a review finding; it does not authorize an unpersisted terminal repair, a fallback reference reload, or a third bounded write. The wrapper is not a broken tool: follow its current artifact and stop boundary, then let the controller classify the evidence.

• If the user asks for prose, output prose only after the article has existed as draft.md unless execution failure prevents file writing. The first visible line must be the title.
• Never prepend Here is, 完成, 以下是, 最终文章, protocol notes, checker summaries, markdown fences, or process commentary.
• Do not narrate reference loading, file checks, or internal decisions in the article or final prose output.
• If the user asks for validation, report commands, conditions, sample size, and results.
• If corpus or background evidence is unavailable, state the limitation in the validation report, not inside the article.
• If a request asks to impersonate real provenance, forge evidence, or deceive about authorship, refuse that part and offer anonymous style-evaluation output instead.

Validation Commands

These commands are mode-scoped. In bounded clean-eval, clean_run_checker.py is the only generator-facing checker; direct check_anlin_violations.py or check_style_profile.py calls belong to ordinary repair, finalized/controller validation after artifact freeze, or developer audits. A clean-eval generator must not use the normal checker because it contaminates the bounded source-guidance measurement.

python scripts/check_anlin_violations.py draft.md
python scripts/check_anlin_violations.py draft.md --strict --draft-gate --genre <standard|sincere|micro-hope|surreal> --corpus-dir <corpus-dir>
python scripts/clean_run_checker.py draft.md --strict --draft-gate --generator-facing --genre <standard|sincere|micro-hope|surreal>
python scripts/prepare_finalized_repair_brief.py <case-dir>/finalized/draft.md --genre <standard|sincere|micro-hope|surreal>
python scripts/check_style_profile.py draft.md --genre <standard|sincere|micro-hope|surreal> --draft-gate --strict
python scripts/check_style_profile.py draft.md --genre <standard|sincere|micro-hope|surreal> --draft-gate --strict --repair-brief
python scripts/calibrate_style_profile.py <corpus-dir> --profile references/style-profile.json
python scripts/summarize_dev_checkpoints.py <case-dir> --bounded-draft <case-dir>/draft.md --finalized-draft <case-dir>/finalized/draft.md --trace-log <case-dir>/opencode-output.txt --corpus-dir <corpus-dir> --profile references/style-profile.json
python scripts/run_blind_test.py draft.md <corpus-dir> --rounds 8 --placebo-rounds 2 --match-genre auto
python -m unittest discover -s test