AI generates plausible OPORD text. Without doctrine-aware constraints it fails structure, loses task coherence across paragraphs, and has no mechanism to catch failures before the order reaches a unit. Most AI-assisted OPORD efforts stop at the prompt. This one did not.
Core pipeline Ongoing
Orders Production
A three-stage AI pipeline that produces exercise-ready OPORDs for brigade and below — governed by doctrine, adversarially reviewed, and gated before unit release. The goal is not a perfect order. It is a 90% first draft that saves units significant prep time.
Where this fits
This is part of the supporting stack.
This project is part of the wider NextGen Wargame system. It strengthens the production chain around CADE by improving source-package completeness, controller usability, or multi-document workflow reliability.
Project Proof
The useful version first.
Every project is framed as a problem, design move, built system, supporting workflow, evidence set, and transferable skill.
Treat orders production as a governed workflow, not a conversation. Each stage — authoring, adversarial review, release gate — runs inside constraints the model cannot self-impose: an OPORD writing guide, quality standards, skeptic persona definitions, lint rules. AI drafts inside those constraints; human review is the gate.
Three sequenced skills (opord-writer, red-team, order-release) supported by Python scripts (lint, DOCX export, normalize), a shared doctrine reference library (FM 5-0, FM 6-0, NATO-COPD, AJP-5), and a Streamlit GUI with dedicated pages per skill.
This project is part of the wider NextGen Wargame system. It strengthens the production chain around CADE by improving source-package completeness, controller usability, or multi-document workflow reliability.
This shows AI-assisted staff work approached as a systems design problem — not just what to ask the model, but what constraints govern its output, what checks catch failures, and what pipeline moves draft material to execution-ready products. The techniques apply equally to real-world OPORDs.
90%
first-draft target
Exercise OPORDs Brigade and below FM 5-0 / FM 6-0 aligned Techniques apply to real-world orders
Pipeline Skills
Three stages. One governed workflow.
Each skill has a specific job in the sequence. None is a prompt wrapper — each enforces constraints AI generation cannot self-impose without them.
01 · Authoring
opord-writer
Drafts, refactors, expands, and red-teams OPORDs from BN to JTF/Corps. Five modes covering the full authoring cycle. Output: planner-source Markdown exported to DOCX via script.
DraftRefactorExpandAlign AnnexSkeptic Sweep
Guardrail: OPORD writing guide
A dedicated writing guide — not a system prompt — governs structure, task-to-purpose standards, coordination measure rules, and echelon-specific quality checks. The model writes inside this constraint. It cannot substitute generic military language for doctrine-specific structure.
02 · Adversarial Review
red-team
cross-project skillRuns structured adversarial assessment using domain-specific skeptic personas — CDR, S3, S2 for orders. Every finding is severity-tagged and paragraph-referenced before the order goes further.
CDR PerspectiveS3 PerspectiveS2 PerspectiveSeverity Tagging
Why it is a separate skill
The skeptic framework is not OPORD-specific. The same approach applies anywhere AI produces a product that needs adversarial pressure — plans, briefs, exercise materials, scenario documents. Separating it from opord-writer makes it reusable across projects.
03 · Release Gate
order-release
A sequential, largely automated pre-release workflow. Seven passes in fixed order: lint, planner skeptics, staff skeptics, specialized skeptics, skeptic synthesis, QC Gate against echelon quality standard, annex alignment. Produces a consolidated report and prioritized fix plan. Assessment only — no edits to the source order.
LintPlanner SkepticsStaff SkepticsQC GateFix Plan
Key design decision
The unit receives an order that has already cleared lint, adversarial review, and an MDMP-sufficiency check against the echelon quality standard. Their time goes into refining the 10%, not rebuilding the 90%.
Deeper Detail
Orders Production is a three-stage AI pipeline — opord-writer → red-team → order-release — that produces exercise-ready OPORDs for brigade and below. Each stage runs inside constraints AI cannot self-impose: an OPORD writing guide, adversarial skeptic personas, and a formal quality gate. The result is a first draft that has been validated before it reaches the unit.
Operating Constraints
- Orders must be FM 5-0 / FM 6-0 compliant at the declared echelon.
- AI generation is constrained by doctrine files — output that contradicts the writing guide is rejected and regenerated.
- order-release is assessment only — no automated edits to the source order.
- Current scope: exercise OPORDs at brigade and below. Techniques apply to real-world orders.
Production Approach
Each stage of the pipeline has a fixed job and runs in sequence. opord-writer drafts inside a doctrine-aware writing guide. red-team applies adversarial pressure from CDR, S3, and S2 perspectives, tagging every finding by severity and paragraph. order-release runs seven automated passes — lint, skeptic sweeps, QC gate, annex alignment — and produces a consolidated fix plan before the order goes to the unit.
Key Decisions
Build a pipeline, not a prompt
Reasoning:
Prompts produce plausible text. A governed pipeline — with writing guide guardrails, adversarial review, and a formal quality gate — produces a product the unit can actually use. The constraints are what make the output credible.
Alternatives considered:
- Prompt engineering with increasingly detailed system prompts
- Single-pass generation with manual human review
Separate red-team as a cross-project skill
Reasoning:
Skeptic-based adversarial review is not specific to orders. Keeping it separate makes it reusable across any AI-produced artifact — plans, briefs, exercise materials. The orders pipeline calls the red-team skill; the red-team skill does not depend on orders.
Alternatives considered:
- Embed skeptic sweep directly in opord-writer
- Run adversarial review only within order-release
Target 90% first-draft quality, not 100%
Reasoning:
A 100% ready order is not an achievable or honest target for AI generation at the current state of the art. A 90% first draft that has been validated and gated saves the unit the most time — they spend prep time on decisions, not structure.
Alternatives considered:
- Present the output as complete and require unit to identify gaps
- Require full human rewrite of AI output before delivery
Tech Stack
- Claude
- Python (lint, DOCX export, normalize scripts)
- Streamlit
- Markdown
- FM 5-0 / FM 6-0 / NATO-COPD / AJP-5
Result & Impact
Units receive a first-draft OPORD that has passed lint, adversarial review from multiple staff perspectives, and a formal quality gate — before it reaches them. Prep time shifts from building structure to refining decisions.
Learnings
- Doctrine files are constraints, not references. The writing guide works because the model is required to write inside it, not consult it.
- A separate quality gate before release is not redundant with in-process review — it catches different failure modes and provides a formal record of what was checked.
- Red-team personas must be domain-specific to be useful. Generic adversarial prompts produce generic findings. CDR, S3, and S2 each see the order through a different operational lens.