### [](#le-d%C3%A9clencheur)Le déclencheur

By Codcompass Team·2026-05-10·4 min read

AI-Assisted Technical Writing: Constraint-Driven Agent Architecture & Pipeline Optimization

Current Situation Analysis

The transition from manual IDE/Git-based writing to AI-assisted drafting initially improved velocity but introduced critical failure modes. A monolithic CLAUDE.md configuration attempted to handle ideation, drafting, and critique simultaneously, causing context dilution and style drift. Without phase-specific constraints, the LLM optimized for fluency over fidelity, resulting in:

Voice Contamination: Drafts increasingly exhibited formulaic AI phrasing (theatrical hooks, condescending transitions, unearned intensifiers), making the output sound generic rather than author-specific.
Review Inconsistency: Running critiques within the same session as drafting caused "windshield wiper" feedback (contradictory suggestions) and review fatigue.
Context Bloat & Quota Exhaustion: Unstructured linear pipelines (ideas -> ébauches -> posts) led to unbounded token consumption, frequent API quota exhaustion, and infrastructure throttling during peak sessions.
Why Traditional Methods Fail: Standard prompt files lack architectural separation of concerns. LLMs inherently conflate creation and evaluation when placed in a single context window, leading to over-writing, loss of authorial control, and degraded output quality.

WOW Moment: Key Findings

Restructuring the workflow into dedicated agents with explicit anti-pattern constraints and phase-gated reviews yielded measurable improvements in output fidelity, review stability, and resource efficiency.

Approach	AI Voice Detection Rate	Review Consistency	Draft Iterations
Monolithic CLAUDE.md + Linear Pipeline	65-70%	Low (contradictory feedback)	4-6
Dedicated Agents + Constraint-Driven Pipeline	15-20%	High (homogeneous, actionable)	2-3

Key Findings:

Separating drafting (redacteur.md) from reviewing (reviewer.md) eliminates context contamination and stabilizes feedback quality.
Explicit anti-pattern enforcement drastically reduces AI-isms, preserving authorial voice.
Structured bullet-point knowledge extraction prevents over-writing and accelerates final composition.
The sweet spot lies in phase-gated constraints: strict guardrails during ideas/knowledge generation, followed by multi-angle critique in isolated sessions.

Core Solution

The architecture shifts from a monolithic prompt file to a modular agent/skill ecosystem with a strict `ideas -> knowled

ge -> output` pipeline.

1. Pipeline Restructuring

ideas: Capture raw concepts, enforce bullet-point structuring, prohibit full prose generation.
knowledge: Research, source validation, and structured extraction. Guardrails prevent narrative drift.
output: Final composition using extracted knowledge, applying tone and style constraints.

2. Agent & Skill Architecture

redacteur.md: Handles ideation and knowledge structuring. Outputs a skeleton with style guardrails. Compensates for AI over-writing by enforcing concise, scannable data points.
reviewer.md: Multi-angle critique (journalist, skeptical reader, bullshit detector). Runs in isolated sessions to ensure fresh, homogeneous feedback.
review-voice.md: Audits final output for banned patterns, journalistic standards, and strong passages.
review-idea.md: Validates ideas for prose violations, pre-oriented blocks, and missing conceptual fields.
review-knowledge.md: Checks knowledge bases for unsourced claims and narrative drift.

3. Context Optimization

CLAUDE.md reduced to <200 lines, containing only essential tone definitions, pipeline explanations, and core constraints.
Anti-pattern list maintained as a living constraint file, updated post-publication or during review cycles.

4. Anti-Pattern Enforcement (Living Constraint List)

"Le vrai..." / "La vraie..." — overused, signal too heavy
"Dit autrement :" — condescending
"C'est exactement ce qui/ça/le piège" — formulaic AI intensifier
"Et c'est peut-être le point le plus important." — theatrical
"Et c'est normal." — paternalistic
"Voici pourquoi." (hook ending) — clickbait
"X ne disparaît pas. Elle change de nature." — over-smoothed antithesis
"Ce n'est pas qu'une question de X. C'est une question de Y." — cliché transition
"Ce qui m'a frappé" / "Ce qui m'a retenu" — naming emotion instead of evoking it
"C'est le signal." (conclusion) — concluding for the reader

Pitfall Guide

Monolithic Context Bloat: Overloading CLAUDE.md with drafting, reviewing, and style rules dilutes context and causes inconsistent model behavior. Keep core configuration under 200 lines and delegate responsibilities to dedicated skills.
Conflating Creation & Critique: Running reviews in the same session as drafting contaminates context and triggers contradictory feedback. Always isolate reviewer agents to preserve analytical freshness.
Unfiltered AI Phrasing Patterns: LLMs default to formulaic intensifiers, theatrical conclusions, and condescension. Maintain a living anti-pattern list updated from published drafts and enforce it via dedicated review skills.
Over-Reliance on Linear Ideation: Allowing AI to freely expand ideas/knowledge results in over-written, AI-sounding content. Enforce bullet-point structuring and strict guardrails during the knowledge phase.
Quota & Infrastructure Blind Spots: Long, unstructured sessions exhaust API quotas and trigger rate limits. Implement phase-gated workflows to control token consumption and prevent mid-session throttling.
Neglecting Reviewer Freshness: Context carryover from drafting sessions biases critique. Use dedicated reviewer agents with clean context windows to ensure homogeneous, unbiased feedback.

Deliverables