Content
# agente-daktus-content
> **Stage 1 of the Daktus CDSS Pipeline.** Skill-based context harness that transforms a clinical briefing into a validated, EHR-compatible JSON protocol. Progressive disclosure over markdown skills, deterministic validation around a probabilistic core, model-agnostic.
**Status:** in production · Cardiology and Rheumatology delivered · Gynecology and Psychiatry in audit
---
## The Pipeline
This repository is **Stage 1** of a three-stage Clinical Decision Support pipeline:
```
[agente-daktus-content] → [agente-daktus-qa] → [daktus-conduct-engine]
Stage 1: Production Stage 2: Validation Stage 3: Decision Engine
briefing → JSON protocol validates, corrects, consumes encounter bundle
versions, learns → structured anamnesis
+ management + evidence
```
- Stage 2 → [agente-daktus-qa](https://github.com/daanmt/agente-daktus-qa)
- Stage 3 → [daktus-conduct-engine](https://github.com/daanmt/daktus-conduct-engine)
---
## The Problem
Clinical content production at scale has a structural tension: LLMs are powerful for evidence synthesis and protocol drafting, but clinical safety is non-negotiable. Pure prompt-engineering produces good prose; it cannot reliably produce structured clinical decision trees with traceable evidence and EHR-compatible output.
The pipeline encodes the answer: a deterministic harness around a probabilistic core, where each phase has explicit contracts and the LLM's job is constrained, auditable, and verifiable.
---
## The Architecture
**Context engineering via skill-based progressive disclosure.** No framework — the pattern emerged from the problem. Each phase of the pipeline loads only its active skill (a `.md` file with structured instructions), preserving context-window efficiency and operational focus across long production runs.
```
SKILL.md (orchestrator)
│
├── briefing-arquitetura → maps clinical briefing into node architecture
├── ingestao-evidencias → ingests OpenEvidence reports into bibliographic bank
├── auditoria-banco → classifies references into TIER 1/2/3
├── redacao-playbook → writes clinical clusters with evidence traceability
├── auditoria-playbook → citation scan + semantic scan + coverage scan
├── codificacao-json → paper design → TUSS → Daktus JSON → validation
└── qa-entrega → 28-point pre-delivery checklist
```
Each skill is specialty-agnostic. The same pipeline runs for any clinical area.
---
## Key Design Decisions
**Markdown skills over Python orchestration framework.** State and instructions live in human-readable `.md` files that any LLM can read. The pipeline is portable across Claude, GPT, Gemini — model-agnostic by construction.
**Progressive disclosure over full-context loading.** Loading the entire spec into context every phase wastes tokens and degrades focus. Loading only the active skill keeps the model on-task and the audit trail clean.
**Deterministic validation layer over probabilistic core.** Structured JSON output is validated against schemas, citations against bibliographic banks, logic against scripted audits — before output reaches Stage 2. Probabilistic where it adds value (synthesis, reasoning); deterministic where it is non-negotiable (output contract, evidence linkage).
**State externalized in canonical files.** `ESTADO.md` is the single source of truth, read at session start by every agent. Session history in `history/session_NNN.md` ensures continuity across LLM instances and human sessions.
---
## Multi-Agent Operation
| Agent | Interface | Responsibility |
|---|---|---|
| **Antigravity** | Claude chat | Clinical reasoning, content drafting, evidence audit |
| **Claude Code** | CLI / terminal | Git, scripts, JSON validation, QA automation |
Synchronization via `ESTADO.md` and `history/session_NNN.md`, read by all agents on session start.
---
## Repository Structure
```
/
├── ESTADO.md # Source of truth — read on every session start
├── CLAUDE.md # Session boot protocol
├── SKILL.md # Orchestrator — full pipeline
│
├── tools/
│ ├── skills/ # 7 sub-skills — loaded per phase
│ └── *.md # Technical instructions, guardrails, patterns
│
├── especialidades/ # Active specialty work in progress
├── referencia/ # Delivered specialties (read-only)
├── history/ # Session logs — full traceability
└── scripts/ # validate_json, audit_logic, audit_references
```
---
## Deliverables
| Specialty | Status |
|---|---|
| Cardiology | Delivered |
| Rheumatology | Delivered |
| Gynecology | JSON audit — vdraft2, corrections C1–C4 |
| Psychiatry | Phase 3 — awaiting clinical review |
---
## Tech Stack
Python · Claude API (Anthropic) · Claude Code CLI · Structured outputs (JSON) · TUSS coding · HL7/FHIR-compatible schemas · Markdown-based agent contracts
---
## About
Built and maintained by [Daniel Martins](https://www.linkedin.com/in/danielmartinsf) at Daktus Health Tech.
Background: engineering (EFOMM — systems modeling, automation) + medicine (UFJF, final-year). I build AI systems for high-stakes domains where the cost of error is real. Healthcare was the proving ground; the architecture travels.
