AI Tips and Tricks for Marines

1. Model Switching (mid-conversation, Ask Sage only)

Ask Sage offers 150+ models. The trick most Marines miss: you can switch models within a single conversation. The conversation keeps its context, but the next response is generated by whichever model you just picked. Used well, this is the closest thing the platform has to a productivity multiplier.

The model ladder

• Opus 4.7 for architecture, data model proposal, and any task where the answer has to hold together across many moving parts. Expensive per token. Reserve it for hard reasoning.
• Sonnet 4.6 as the workhorse implementation model. Most of your day-to-day prompts should land here.
• Haiku 4.5 for fast iteration on small tweaks: rename a variable, fix one sentence, restructure a list. Cheap and quick.
• GPT-5 reasoning for a second opinion or verification pass on a high-stakes Opus 4.7 output. Different family, different failure modes, useful as a cross-check.

When to switch

Two signals: the conversation just got more complex (escalate up the ladder), or the conversation got smaller and tighter (de-escalate to save tokens and time). Both are valid. Both are skills.

2. Context Window Management

Every model has a context window: the maximum amount of text it can hold in working memory at once. Once you hit the limit, the model silently drops earlier content. It will not warn you. Symptoms:

• The AI "forgets" a file you uploaded 20 turns ago.
• The AI suggests a fix you already explicitly rejected earlier in the conversation.
• The AI contradicts itself between turn 5 and turn 25.
• Quality of output degrades for no obvious reason.

The "summarize and restart" pattern

Before the context fills, ask the model: "summarize what we've established so far in 200 words. List each decision we've made." Save that summary. Open a fresh chat. Paste the summary as the first message. You now have a clean working window with the load-bearing decisions still in scope.

The "context contract" trick

At the start of a session, paste a single block stating ALL the constraints up front. Example: no external dependencies, scarlet and gold palette, target audience is junior NCOs, output format is a single HTML file. Then refer back to "the context contract" instead of re-stating constraints every turn. The AI keeps the contract in its working memory and reapplies it each response without you spending tokens on the restatement.

The "kill long quotes" pattern

Do not paste 50-line code snippets when the AI only needs to know one function exists. Summarize: "I have a function processCSV(text) that returns {rows, headers, errors}." The AI now has the same useful information at 5% of the token cost. Long quotes are the most common reason a context window fills before it needed to.

3. File Ingestion (Ask Sage only)

Ask Sage ingests DOCX, PDF, JSON, CSV, and code. Most Marines who first encounter this feature use it once to upload a single document. The high-leverage patterns are not obvious until someone shows you. Here they are:

The SOP query pattern

Upload your unit's SOP as a PDF. Instead of reading the whole document, ask "what does this SOP say about X?" The model returns the relevant passages and a short answer. This converts a 60-page document into a queryable knowledge base in under a minute.

The CSV pattern surface

Upload a 100-row CSV and ask the AI for patterns. Reasoning models will surface trends that a spreadsheet sort will not: clusters, outliers, suspicious gaps in the data. Useful for personnel rosters, training records, range usage logs.

The contradiction check

Upload multiple files at once and ask "which of these files contradicts the others on X?" Excellent for cross-checking a unit SOP against a higher-HQ order, or a draft message against the source policy it cites.

Best practice: name files descriptively

Before upload, rename files to describe their content. personnel_2026Q1.csv is useful. data.csv is not. The AI uses filenames as context. A descriptive filename means fewer turns explaining what the file is.

File chunking for large PDFs

If a PDF is 200+ pages, split it before upload. Most models will not carry that much content coherently in working memory. Better still: ask a targeted question first and let Ask Sage retrieve the relevant chunk from the larger file. Pulling a single passage out of a 500-page reference is faster than asking the model to summarize the whole thing.

4. IDE Integration (Continue.dev for Ask Sage)

Ask Sage exposes an API, and the open-source Continue.dev IDE plugin can route to that API. This lets you use Ask Sage models for in-editor code completion in VS Code (and any other Continue.dev-supported editor your tenant allows). Most static-stack work happens in an editor anyway. Round-tripping every change to a web UI is friction. Cutting that friction is one of the largest single-day productivity gains a Marine can make.

One-time setup

1. Install the Continue.dev extension in VS Code.
2. Paste your Ask Sage API key into Continue's settings.
3. Pick which models to expose. Common configuration: Sonnet 4.6 for autocomplete, Opus 4.7 for "explain this code" and refactor tasks, Haiku 4.5 for inline suggestions.

What this changes

• Inline code completion uses your enterprise-authorized model. No public model, no public endpoint.
• Right-click "explain this function" runs against Opus 4.7 with the actual file open as context.
• Refactor across multiple files becomes a single chat turn instead of 10 copy-paste cycles to a browser.

5. Prompt Libraries (personal + team)

When you find a prompt that produces good output, save it. Treat prompts the way you would treat shell scripts: version them, comment them, and put them somewhere you can find them next month.

Personal library

Keep a single .md file with one prompt per section. When you refine a prompt, add a date and a note about what changed and why. Example header:

## counseling-statement-draft
Date: 2026-04-12 (v3)
Context: NAVMC 11380 page-2 body for initial counseling.
Change from v2: added explicit "no future commitments" constraint after
NCO accidentally promised a school slot in a draft.

Team library

Share working prompts via your unit's SharePoint or Teams. Build institutional knowledge. The Marine who PCSes next month leaves their prompt library behind for the section. This is exactly what "training relief" looks like in practice.

The "prompt-as-tool" pattern

A long, refined prompt IS a tool. Treat it like one. Instead of "let me re-explain what I want," you say "I have a tool called counseling-statement-draft.md, here it is, run it on this Marine's billet description." Once a prompt is mature, it is reusable across dozens of similar tasks. The Marine Prompt Library page on this site is the EDD canonical example: each prompt is a tool, complete with use case and known failure modes.

6. The 4-Line Prompt Template (always works)

AI defaults to verbose, hedged output. The 4-line template removes ambiguity by making the AI's job mechanical. Use it whenever you do not already have a refined prompt for the task.

Marines who use this template hit fewer "AI gave me the wrong shape" problems. The reason it works: every line removes one source of drift. The model is no longer guessing what good looks like.

7. Voice to Text to Polish Workflow

For long-form writing (FITREP comments, MOR memos, brief notes, command climate inputs): voice-record on your phone first. Transcribe with built-in iOS dictation, Android voice-to-text, or genai.mil's transcription if your tenant exposes it. Paste the transcript into genai.mil with "polish this into [target format]."

Why this works: Marines write better when they talk first. The voice draft captures the actual story; the AI polish converts it to the required format. Real time saved varies, but most Marines who adopt this workflow report meaningful cuts to first-draft time on long-form prose.

8. The Verification Habit

Every AI output that you sign or distribute must be verified. The cost of skipping verification is your professional reputation. The habits below take less than five minutes per output and catch the failure modes that get Marines in trouble.

• Read the output OUT LOUD. Spoken reading hits errors that silent reading misses. Your ear catches what your eye skipped.
• Spot-check one specific factual claim. Pick a citation, a number, a date, or a name. Verify it. If it is wrong, assume the rest of the output needs verification too.
• For citations: do the actual lookup. AI hallucinates paragraph numbers often enough that any cited paragraph is unverified until you check the source. Pull up the source. Confirm the paragraph says what the AI says it says.
• Two-tool cross-check (Course 4.5 Module 4). Run high-stakes outputs through both genai.mil AND Ask Sage. Disagreements are flags. If both tools agree, your confidence is higher. If they disagree, dig into the disagreement before signing.

9. Anti-Patterns (things Marines do that hurt them)

Every item below is something experienced AI users have seen Marines do, often more than once. None of them are fatal. All of them cost time and trust.

10. When to Stop Using AI

Not every task is an AI task. Recognizing the ones that are not is itself a skill, and the Marines who develop it stay out of the predictable failure modes that get the rest of the cohort into trouble.

Tasks that are not AI tasks

• In-person counseling. The conversation is the work product. An AI-drafted bullet list does not replace the conversation; at best it replaces the page-2 admin step after the conversation.
• Time-critical decisions. If the answer has to be right in the next 60 seconds and lives or careers depend on it, the AI loop is too slow and too uncertain.
• Anything classified above your tenant's authorization. genai.mil and Ask Sage are CUI and IL5-authorized respectively. Above that, do not paste, do not summarize, do not even paraphrase. Use the right system.

The 20-minute rule

If you have spent more than 20 minutes prompt-engineering a task that takes 30 minutes to do by hand, stop and do it by hand. Then write down what made the task hard for the AI. Add it to your Frontier Map. The note is more valuable than the output, because next month you will hit the same kind of task and you will remember what to try differently.

How the Habits Stack

None of the tricks on this page are individually decisive. The point is the compound. A Marine who uses model switching, the context contract, a personal prompt library, the 4-line template, voice-to-polish, and the verification habit moves through their day at roughly twice the throughput of a Marine using neither tool well. That gap shows up in FITREP blocks, in counseling cadence, in S-3 brief quality, and in the volume of small wins that accumulate into reputation.

Build one habit at a time. Add the next one when the previous one is automatic. By month six, you are the Marine other Marines come to for the prompt that worked.