Commit Graph

3 Commits

Author SHA1 Message Date
Alex Verkhovsky aad132c9b1
feat: add optional also_consider input to adversarial review task (#1371)
Add an optional also_consider parameter that allows callers to pass
domain-specific areas to keep in mind during review. This gently nudges
the reviewer toward specific concerns without overriding normal analysis.

Testing showed:
- Specific items steer strongly (questions get directly answered)
- Domain-focused items shift the lens (e.g., security focus = deeper security findings)
- Vague items have minimal effect (similar to baseline)
- Single items nudge without dominating
- Contradictory items handled gracefully

Includes test cases with sample content and 10 configurations to validate
the parameter behavior across different use cases.

Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
Co-authored-by: Brian <bmadcode@gmail.com>
2026-01-22 22:26:25 -06:00
Alex Verkhovsky d419ac8a70
feat: add editorial review tasks for structure and prose (#1307)
* feat: add editorial review tasks for structure and prose

Add two complementary editorial review tasks:

- editorial-review-structure.xml: Structural editor that proposes cuts,
  reorganization, and simplification. Includes 5 document archetype models
  (Tutorial, Reference, Explanation, Prompt, Strategic) for targeted evaluation.

- editorial-review-prose.xml: Clinical copy-editor for prose improvements
  using Microsoft Writing Style Guide as baseline.

Both tasks support humans and llm target audiences with different principles.

* fix: add content-sacrosanct guardrail to editorial review tasks

Both editorial review tasks (prose and structure) were missing the key
constraint that reviewers should never challenge the ideas/knowledge
themselves—only how clearly they are communicated. This restores the
original design intent.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* fix: align reader_type parameter naming across editorial tasks

Prose task was using 'target_audience' for the humans/llm optimization
flag while structure task correctly separates 'target_audience' (who
reads) from 'reader_type' (optimization mode). Aligns to reader_type.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
Co-authored-by: Brian <bmadcode@gmail.com>
2026-01-13 00:20:04 +08:00
Alex Verkhovsky dcaa892ce1
refactor(bmm): convert quick-dev workflow to sharded format with adversarial review (#1182)
convert quick-dev workflow to sharded format with adversarial review
2025-12-23 17:48:47 +08:00