Self-Evolving Writing Workflow

BeeWeave's self-evolving writing workflow is coordinated by four groups of skills:

• writers: beeweave-article-writer and beeweave-social-writer handle writing, revision, current draft persistence, and trace recording.
• initializer: beeweave-writing-style-initializer initializes writing style asset templates.
• learner: beeweave-writing-style-learner extracts candidate rules from historical articles, traces, diffs, and feedback.
• evolver: beeweave-writing-skill-evolver reviews, validates, activates, rejects, rolls back, and compacts writing rules.

First Use

Writing style assets live in the current Workbench:

workbench/writing/
+-- style/
+-- traces/
+-- eval/

The style/ directory stores user style assets, including author profile, active rules, anti-patterns, examples, pending rules, rejected rules, and the evolution log. beeweave-writing-style-initializer is the explicit initialization entrypoint.

Templates live under beeweave-writing-style-initializer/references/. Article and social writers do not own these templates. If style assets are not initialized, they prompt the user to initialize them first. If the user still wants to continue writing, writers proceed with built-in defaults. Real user style assets are stored only under workbench/writing/.

Initialize Style Assets

On first use, run the initializer to create the fixed writing style asset files:

/beeweave-writing-style-initializer

After initialization, ask the learner to study historical articles:

/beeweave-writing-style-learner workbench/articles/published

The initializer only creates directories and the eight fixed template files. It does not learn from articles. The learner studies historical articles, traces, diffs, or feedback only after initialization is complete.

Daily Writing

When writing long-form articles or short social content, writers first read the applicable active style rules, anti-patterns, and examples, then create a draft. Drafts are still saved to:

workbench/articles/drafts/

By default, one piece of content maintains a single current working draft. When the user asks AI to revise, shorten, expand, change the opening, or revise based on feedback, the writer updates the current draft_path instead of creating a pile of -v2, -v3, or -final files by default.

Version history is written to:

workbench/writing/traces/YYYY-MM-DD-<type>-<slug>/
+-- trace.md
+-- trace.json

Traces record paths, summaries, version events, diff summaries, and learning state by default. They do not copy the full body text. Full body snapshots are stored under trace snapshots/ only when the user explicitly asks to save full process snapshots, preserve a specific version, or mark a version as a learning sample.

AI Revisions and Manual Edits

If the user asks AI to revise a draft, for example "the opening is too grand; make it start from a real experience", the writer records that revision as a revision_events entry containing the user instruction, diff summary, and a strong learning signal.

If the user manually edits a draft and says "this is my edited version; learn what I changed", the learner reads the original draft, edited draft, or trace and extracts stable editing patterns.

Signal strength is interpreted in this order:

• Manual user edits are strong signals.
• User-instructed AI revisions are strong signals.
• AI self-revisions without user confirmation are weak signals.
• Versions marked final, published, or explicitly accepted have stronger evidence.
• Rejected, abandoned, or unused versions can only become anti-pattern candidates.

Rule Learning and Activation

By default, the learner only writes candidate rules to:

workbench/writing/style/pending_rules.md

Each candidate includes scope, suggested layer, evidence, confidence, and validation suggestions. The learner does not modify active_style_rules.md or a writer SKILL.md by default.

When the user asks to activate, reject, or organize candidates, the evolver first decides which layer each candidate belongs to:

• routing layer: adjusts skill trigger boundaries.
• instruction layer: adjusts stable workflow and quality standards.
• resource layer: adjusts Workbench style assets, examples, narrow scenarios, or eval cases.

Changes that affect active rules, rejected rules, source SKILL.md files, or shared references require validation results or explicit user confirmation. Validation material lives under workbench/writing/eval/; it can compare outputs before and after a change using historical briefs, user final drafts, and rubrics.

The evolver only handles rule lifecycle and style asset maintenance. It does not publish articles, move drafts, trigger wiki ingestion, or manage trace publish metadata during normal publishing.

Publishing, Cleanup, and Compaction

When publishing, beeweave-article-publisher treats the current working draft as the candidate final draft, moves it to workbench/articles/published/, updates frontmatter, and triggers wiki ingestion.

If the draft has a corresponding trace, the publisher updates status, final_version, published_path, and the cleanup summary in the trace, then appends a published revision event. If the trace cannot be found or cannot be updated, publishing and wiki ingestion are not blocked, but the final response explains that the trace was not updated.

The publisher does not create a new base trace and does not manufacture a new writing version for publishing. final_version defaults to the trace's current version.

If a trace contains temporary snapshots/, the publisher does not delete any snapshot by default. It cleans temporary snapshots only when the user explicitly asks for post-publish cleanup, while preserving the final snapshot or snapshots explicitly marked as learning samples. No extra cleanup is needed when there are no snapshots.

As rules accumulate, ask beeweave-writing-skill-evolver to perform compaction. It proposes a plan to merge duplicate rules, move narrow rules down into assets, delete long-term ineffective rules, and preserve key rules. After confirmation, it updates style assets and writes to evolution_log.md.