Writers Loop

Writer's Loop

Core Rule

Extract preferences only from explicit user decisions:

• Approved or rejected plans
• Accepted, rejected, or undone edits
• Manual rewrites that change style, voice, or structure (not fact corrections)
• Explicit standing preferences declared by the user

Never extract preferences from unreviewed drafts, one-off comments, fact corrections, or current-task constraints. Learn from decisions, not raw drafts.

When to Use

Use for writing where quality depends on structure, audience fit, correctness, style, or iteration — typically 300+ words or any writing where feedback and refinement are expected:

• Coding plans, implementation plans, and technical proposals
• Business reports, research summaries, executive updates, and memos
• Product specs, design docs, documentation, and tutorials
• Fiction chapters, essays, scripts, speeches, and narrative outlines
• Poetry (limited support — use custom rubric; see references/artifact-types.md)
• Academic papers and whitepapers (limited support — adapt to venue requirements)
• Brand-governed or legally constrained writing (legal, clinical, regulatory, corporate)
• Style learning or style distillation from the user's own writing, permitted reference samples, files, pasted text, chapters, reports, docs, or codebase prose
• Translation where meaning, source-language writing style, voice, rhythm, formatting, terminology, or cultural effect matters

Do not use for tiny one-off text edits unless the user asks for a reusable process or preference learning.

Core Loop

Follow the loop in order. Do not emit a later-stage section before its gate has passed:

1. Frame: Identify artifact type, audience, purpose, constraints, success criteria, and desired tone.
2. Question Gate: Ask only blocking questions. If context is enough, state assumptions and continue.
3. Plan: Create the outline, argument, task sequence, scene plan, or section structure before drafting.
4. Plan Checkpoint: Stop and wait for the user to approve or request changes before drafting. ⚠ Do not draft before approval. Exception: fast draft path (see Entry Modes).
5. Replan: If the plan is rejected, revise only the criticized parts, reissue the full plan, and ask again.
6. Draft: Write from the approved plan. Preserve declared constraints.
7. Critique: Evaluate against the artifact-specific rubric. Every critique must point to a specific location and name a concrete problem — not a generic observation.
8. Propose: Suggest targeted revisions with reason and expected improvement. Stop and wait for the user's decision on each. ⚠ Do not apply changes before the user responds.
9. Decide: Record the user's apply, reject, or adjust response. Do not skip this step — preference learning depends on explicit decisions.
10. Revise: Apply accepted changes. Keep rejected changes out.
11. Evaluate: Check the revised artifact against the original success criteria.
12. Learn: Extract preference signals from decisions collected above. Record the signal type, applies-to scope, and evidence.
13. Distill: Promote repeated, high-confidence patterns into reusable rules.
14. Reuse: Apply stable rules in later tasks, scoped to artifact type and stage.

Entry Modes

• New artifact: start at Frame, then Question Gate, then Plan.
• Fast draft (questions waived): Frame → state Assumptions → compact inline Plan → Draft → Learning Status. Use only when the user clearly waives questions and asks for an immediate draft. Do not use this path when the user asks to bypass, ignore, or override a known plan checkpoint; in that case, present the plan and stop at PLAN CHECKPOINT. Skipped questions are weak signals and will not become preferences.
• Multi-agent: default to single-agent. Use multi-agent only for high-stakes, long, ambiguous, or multi-audience artifacts where independent planning or critique would materially improve quality. Read references/multi-agent.md first.
• Existing draft: start at Frame, then Critique; plan only if structure is unclear. If source text is missing or placeholder-only, ask for it and use this wording: "preserve user intent, voice, plot facts, and continuity."
• Targeted revision: start at Frame, then Propose; keep changes at sentence or paragraph level unless the user explicitly asked for section-scale changes.
• Style learning: start at Frame, then read references/style-distillation.md. Output Frame, Style Versus Content, Style Pack, and Storage Decision. If the user wants to clone another person's style, confirm permission or keep the pack session-only.
• Using a learned style: before asking artifact questions or drafting, load the style pack from the conversation or from .writers-loop/styles/ only when the user opted into local storage. State which pack is loaded. Draft with it, then critique content quality and style match separately. Do not copy source passages or facts from the style pack evidence. If the style pack is missing, unavailable, or the user says not to load it first, use the missing-style-pack template in references/style-distillation.md exactly, including Style Pack Status, Content Plan, and Style Application Plan. Do not draft and do not output QUESTION GATE or general artifact questions yet.
• Translation: start at Frame, then read references/translation.md. Output Frame, Translation, Review, and Learning Status; preserve source formatting inside the Translation section, not by omitting the loop metadata.
• Preference update: start at Learn. For storage-mode requests, confirm whether .writers-loop/, journal.jsonl, prefs.md, or style packs will be created before asking artifact questions. Treat tone, length, detail level, or storage mode set for the current task as a constraint, not a learned preference, unless the user says it applies in future work.

Planning Gates

Ask up to five questions only when the answer would materially change the plan. If the user skips questions, state working assumptions and continue. Skipped questions are weak or neutral signals, not strong preferences.

Fast draft exception: if the user clearly waived questions while asking for a draft, show a compact Plan inline and draft immediately. Do not pause for plan approval. End with Learning Status.

Checkpoint pressure override: if the user asks to bypass, ignore, or override a known plan checkpoint, the checkpoint still applies. Output a plan, explain that fast draft only waives blocking questions, then stop at PLAN CHECKPOINT.

After presenting a plan in the standard path, output the checkpoint format (see references/checkpoints.md) and wait. Do not draft until the user replies. If the user's response is ambiguous, ask: "Does that mean approve, or would you like changes?"