Write More, Spend Less: How to Use Tokens Efficiently
Before anything else: keep your SoloEnt client up to date
We keep optimizing context management and API caching in every release, so that as models change, you always get the most cost-effective setup.
Download the latest SoloEnt
The core logic
Token usage = input size × number of calls
With that in mind, we'll approach savings from two directions: controlling the size of each input, and cutting wasted calls.
High impact: save on every use
Limit the context window
Let the AI see only what it actually needs. When writing chapter 47, it doesn't need chapter 1. When polishing a line of dialogue, it doesn't need the whole chapter.
In practice:
- Activate only documents relevant to the current scene. To write a chapter, the AI only needs the directly relevant setting, chapter outline and limited context. Build a SOLOENT.md so the AI can understand the context quickly from that one file, instead of reading more documents every time.
- Use
@to reference precisely, or drag the relevant documents into the chat window (hold Shift), instead of opening or reading every document by default. - When revising, select only the target paragraph, not the whole chapter.
- After each scene, close document references you no longer need.
Estimated savings: 40–60%
Use short instructions instead of long descriptions
The AI doesn't need your backstory. It just needs to know "what to do" and "how." SoloEnt already provides a system prompt, so there's no need to restate the background in your conversation.
❌ Token-hungry:
You are a professional fiction writing assistant. Please rewrite this dialogue to make it more tense,
so readers feel the strained relationship between the two characters, while keeping their personalities consistent...
✅ Token-efficient:
Rewrite dialogue: raise conflict tension, keep each character's voice
Save your frequent instructions as a Skill and call them with one click, at zero description cost.
Estimated savings: 20–35%
Audit the Rules you're using
Rules are the most overlooked hidden token hog, because Rules are force-loaded into every request by default.
Trimming principles:
- Load drafting-related Rules only when drafting
- Delete "You are a..." role-play preambles (the AI already knows what it is)
- Use lists instead of paragraphs: same information, half the tokens
- Review your Rules once a quarter and remove anything that's been internalized and no longer needs stating
Estimated savings: 15–30%
Medium impact: build good daily habits
Use light models for light tasks
Not every task needs the strongest model.
| Task | Best model (when quality matters) | Light model (when you can lower the bar) |
|---|---|---|
| Brainstorming, outlining, consistency checks | Sonnet | Haiku, Gemini Flash |
| Drafting prose, polishing dialogue, expanding scenes | Gemini | Gemini Flash, Haiku |
| Complex plotting, deep style imitation, long-form logic | Opus | Sonnet |
| First drafts, rough outlines | Gemini Flash, Haiku | Open-source models |
Estimated savings: 50–70% (on light tasks)
Work in steps instead of asking for everything at once
Don't keep generating speculatively: asking the AI for a 2,000-word chapter and starting over whenever you're unhappy is the most wasteful pattern there is.
Recommended flow (chapter writing example):
- Step 1: Get a chapter outline first
- Step 2: Once you're happy with the outline, expand it into prose
- Step 3: Finally, fine-tune voice and style
Each step uses very few tokens, and you only move on once it's right, so the total ends up far below repeated regeneration.
Make good use of Plan mode: before executing, switch to Plan mode and use a few light rounds to confirm direction, structure and key details, then switch back to execution to generate content. Plan mode uses very few tokens, and aligning the direction once saves the huge cost of regenerating many times later.
[Plan mode] In this chapter I want A and B to reconcile, but plant a hint about C. What structures could work?
→ Align on direction, confirm the beats
[Execution mode] Write the chapter using the second structure
Estimated savings: 30–50% (iterative work)
Open new windows often. Don't keep going in an old conversation
Every chat window has a history: the more you go back and forth, the larger the input tokens for the next message, because the whole history gets sent along. A window that's dragged on for dozens of turns can burn a huge number of tokens on "historical baggage" alone.
Tips:
- After finishing one independent task, open a new window for the next
- Don't polish dialogue, discuss the outline and revise the setting all in the same window
- If a window is already long and you need to regenerate content, open a new window and bring in only the necessary context
- Reactivate the right context by calling SOLOENT.md or
@-referencing specific files
Good habit: one window, one task.
Estimated savings: 10–30% (over time)
Have the AI edit, not rewrite
Left unconstrained, the AI tends to output the whole passage again, so tell it exactly which part to change every time.
❌ Triggers a full rewrite:
Improve this passage
✅ Outputs only the change:
Only change the third paragraph. Slow down the sentence rhythm. Output just the revised paragraph, nothing else
Add "no explanation needed" and "no summary needed": the AI's preambles and sign-offs cost tokens too.
Estimated savings: 20–40% (polishing)
Advanced: deeper optimization
Lock in your frequent routines with a Workflow
If you do the same prep before every chapter (check the recap, confirm characters' emotional states, read the chapter outline), turn it into a Workflow. The only parameter is the chapter number; the system assembles the rest.
Each call's prompt tokens stay at a fixed minimum instead of varying with however you phrase it that day. It also keeps every run consistent.
Result: consistency and token savings, both at once
Use local models as a "draft layer"
Run an open-source model locally with LM Studio to generate first drafts (at zero marginal cost). Once the draft exists, use a cloud model only for the final polishing pass. That pass uses few tokens but makes a big difference.
Hardware reference:
- 16 GB RAM: runs 7B-parameter models (fine for drafting)
- 32 GB RAM: runs 13B-parameter models (more consistent quality)
For high-output writers, cloud usage can drop by more than 60%
The one-line summary
Control your context and state exactly what you need. Not prompt-tuning. That's the real key to saving tokens.
Keep Rules short, reference precisely, choose models by task. Do all three and your monthly token bill can drop by more than half, with no loss in writing quality.