y (%) | Review Consistency Score | Draft Over-Generation Rate | Session Quota Efficiency | Time-to-First-Draft (hrs) | |----------|---------------------------|--------------------------|----------------------------|--------------------------|---------------------------| | Manual/Git Baseline | 95% | N/A | 0% | N/A | 6.5 | | Monolithic CLAUDE.md + Sonnet 4.6 | 42% | 38% | 78% | 31% | 2.1 | | Dedicated Agents & Constrained Pipeline | 89% | 91% | 12% | 84% | 2.4 |

Key Findings:

• Isolating review logic from generation contexts eliminated contradictory feedback and stabilized critique quality.
• Constraining the CLAUDE.md to <200 lines and offloading phase-specific logic to dedicated agents reduced AI over-structuring by 66%.
• Structured bullet-point extraction during the ideas and knowledge phases preserved authorial control while maintaining AI-assisted research speed.

Core Solution

The stabilized workflow replaces monolithic prompting with a modular agent/skill architecture, enforcing strict phase boundaries and explicit style constraints.

1. Pipeline Architecture

The workflow enforces a linear progression with dedicated handlers: ideas -> knowledge -> output

2. Agent & Skill Decomposition

Instead of a single configuration file, responsibilities are distributed across specialized modules:

Review Skills (Phase-Specific Audits)

• review-voice.md: Audits output for banned patterns, journalistic tone compliance, and structural integrity.
• review-idea.md: Validates brainstorming entries for prose violations, pre-oriented framing, and missing conceptual fields.
• review-knowledge.md: Checks research notes for unsourced claims, factual drift, and unnecessary editorialization.

Core Agents

• redacteur.md: Handles drafting for ideas and knowledge phases. Outputs structured bullet points rather than full prose, proposes output skeletons with style guardrails, and actively compensates for model over-drafting tendencies.
• reviewer.md: Executes multi-angle reviews (journalist, cynical reader, bullshit detector). Runs in an isolated context to ensure fresh, consistent critique without generation bias.

3. Constrained CLAUDE.md

Reduced to <200 lines, containing only:

• Core tone directives
• Pipeline stage definitions
• High-level guardrails All phase-specific logic, antipatterns, and review criteria are externalized to agents/skills.

4. Explicit Antipattern Registry

Banned phrases and structural clichés are enforced programmatically via review skills. Detected patterns are logged and added to the registry:

- "The real..." / "The actual..." — Overused, signal too heavy
- "In other words:" — Condescending transition
- "This is exactly the trap/pitfall" — Formulaic AI intensifier
- "And this might be the most important point." — Theatrical padding
- "And that's normal." — Paternalistic framing
- "Here's why." (hook ending) — Clickbait structure
- "X doesn't disappear. It changes nature." — Over-smoothed antithesis
- "It's not just about X. It's about Y." — Cliché transition
- "What struck me..." / "What stayed with me..." — Naming emotion instead of evoking it
- "This is the signal." (conclusion) — Preempting reader interpretation

Pitfall Guide

1. Monolithic Context Bloat: Packing all instructions, tone rules, and pipeline logic into CLAUDE.md dilutes attention mechanisms. LLMs struggle to prioritize conflicting directives, leading to inconsistent outputs.
2. Unconstrained Generation in Early Phases: Allowing the model to freely draft ideas and knowledge sections causes over-structuring. The AI fills gaps with plausible but generic prose, eroding the author's unique analytical voice.
3. Context Contamination During Reviews: Running generation and critique in the same session causes "windshield wiper" feedback. The model's recent outputs bias its self-review, resulting in oscillating praise/criticism.
4. Ignoring Negative Prompting & Antipatterns: Failing to explicitly ban formulaic AI phrases guarantees a synthetic tone. LLMs default to high-probability transitional clichés unless constrained.
5. Over-Engineering Ideation: Applying strict output formatting or review criteria to brainstorming phases kills creative exploration. Early phases should prioritize raw signal extraction, not polished prose.
6. Neglecting Quota & Infrastructure Realities: Heavy, continuous AI sessions without pacing trigger rate limits and quota exhaustion. This breaks workflow continuity and forces manual recovery during critical drafting stages.
7. Skipping Cognitive Disconnection: Continuous AI dependency without breaks prevents perspective reset. Stepping away allows the author to recalibrate stylistic boundaries and detect AI drift that becomes invisible during prolonged sessions.

Deliverables

• 📦 Project Blueprint: Complete repository structure demonstrating the agent/skill decomposition, pipeline configuration, and review isolation patterns. Available at: github.com/agaches/starter-packs/tree/main/blog
• ✅ Implementation Checklist:
  • Extract phase-specific logic from CLAUDE.md into dedicated .md agent files
  • Implement isolated review contexts to prevent generation bias
  • Populate antipattern registry with detected AI clichés
  • Enforce bullet-point extraction for ideas/knowledge phases
  • Validate quota pacing and session boundaries before heavy drafting
• ⚙️ Configuration Templates: Ready-to-use redacteur.md, reviewer.md, and review skill skeletons with embedded guardrails and negative prompt hooks.