Optimizing Context for Quality and Cost

Your context files power every skill. Well-structured context = better output at lower cost.

Why Context Optimization Matters

Every word in your context files gets sent to Claude with every skill run. This affects:

• Quality: Too little context = generic output. Too much = noise drowns signal.
• Cost: More tokens = higher API costs
• Speed: Smaller context = faster processing

Goal: Maximum signal, minimum noise.

Context File Size Guidelines

Total recommended: 2,000-5,000 words across all files

What to Include (and Exclude)

company.md

✅ Include

• Mission (1-2 sentences)
• Business model (B2B/B2C, pricing tier)
• Team structure (who does what, key roles)
• Current priorities (this quarter's top 3)
• Stage (pre-seed, Series A, etc.)
• Key metrics (ARR, customers, growth rate)

Example:

## Company
**Luma** is AI-powered project management for creative agencies.
We're Series B ($25M raised), 85 employees, $4.2M ARR growing 85% YoY.

**2026 Priorities:**
1. Win the agency vertical (become default choice)
2. Expand AI capabilities (planning, scope detection)
3. Move upmarket (50-200 person agencies)

❌ Don't Include

• Full company history
• Detailed financial data beyond key metrics
• Every team member's bio
• Irrelevant past projects
• Office locations, perks, culture details

product.md

✅ Include

• What it does (elevator pitch)
• Target user (who and why)
• Core features (top 5-7 only)
• Key metrics (DAU/MAU, retention, NPS)
• Tech stack (high level only)
• Current roadmap (this quarter + next)

Example:

## Product
**Luma** helps agencies plan projects, track time, and manage resources.

**Core features:**
- Project templates for agency work
- Time tracking and budgets
- Client portals and approvals
- AI project planner (beta)
- Resource planning

**Key metrics (Q1 2026):**
- DAU/MAU: 45%
- NPS: 42
- Month 1 retention: 58%

❌ Don't Include

• Every feature detail
• Complete tech architecture
• Historical feature decisions
• Deprecated features
• Internal implementation notes

personas.md

✅ Include (per persona)

• Who: Role, company size, industry
• Jobs to be done: What they're trying to accomplish
• Pains: Frustrations and blockers
• Gains: What success looks like
• Tech comfort: Tool proficiency

Example:

## Persona 1: Agency Product Manager

**Who:** Project manager at 20-50 person creative agency
**Jobs:** Manage 5-10 client campaigns, report to clients, wrangle freelancers
**Pains:** Clients want real-time visibility, manual status reports take 4 hrs/week
**Gains:** Client trust through transparency, time back for strategy
**Tech comfort:** High (uses Notion, Asana, Slack daily)

❌ Don't Include

• Demographics (age, location, hobbies)
• Day-in-the-life narratives
• Full tool inventory
• Personal background stories
• Made-up quotes

competitors.md

✅ Include (per competitor)

• Name and category
• Key differentiator
• Pricing (high level)
• Target customer
• When they win vs us

Example:

## Competitor: Monday.com

**Differentiator:** All-in-one work OS, highly customizable
**Pricing:** $8-16/user/month
**Target:** General teams across industries
**They win when:** Customer wants all-in-one (CRM + PM + docs)
**We win when:** Agency needs vertical-specific features (client portals, retainer tracking)

❌ Don't Include

• Full feature lists
• Complete pricing tables
• Historical company info
• Lengthy analysis
• Subjective opinions

Token Usage Optimization

Tokens = cost. Every word in context gets sent to Claude with every skill run.

Example Cost Calculation

Context files: 3,000 words = ~4,000 tokens Skill prompt: ~1,000 tokens Your input: ~500 tokens Total input: ~5,500 tokens per skill run

If you run 50 skills per month:

• 50 × 5,500 = 275,000 tokens
• At Claude Sonnet rates ($3/million input tokens): **$0.83/month**
• At Claude Opus rates ($15/million input tokens): **$4.13/month**

With bloated context (10,000 words = 13,000 tokens):

• 50 × 13,500 = 675,000 tokens
• Sonnet: ~$2.03/month (2.4× more)
• Opus: ~$10.13/month (2.4× more)

Savings from lean context: $1-6/month, faster responses

Optimization Strategies

1. Remove Unnecessary Words

❌ Bloated:

Our company, which was founded back in 2022 by our CEO Jane Smith and CTO Bob Johnson, has grown significantly over the past few years and now employs approximately 85 people across various departments...

✅ Lean:

Founded 2022 by Jane Smith (CEO) and Bob Johnson (CTO). Now 85 employees, Series B stage.

Savings: 40 words → 15 words (62% reduction)

2. Use Bullet Points, Not Paragraphs

❌ Paragraph form:

Our target customers are primarily creative and marketing agencies that range in size from about 20 to 100 people. These agencies typically work with multiple clients simultaneously and struggle with project visibility, time tracking, and resource allocation. They often find that general-purpose tools like Asana are too generic, while enterprise tools like Workfront are too complex and expensive for their needs.

✅ Bullet form:

**Target customer:**
- 20-100 person creative/marketing agencies
- Pains: project visibility, time tracking, resource allocation
- Too generic: Asana. Too complex: Workfront.

Savings: 68 words → 26 words (62% reduction)

3. Remove Redundancy

❌ Redundant:

## Mission
Empower agencies to do their best creative work.

## Values
1. Agencies First — Every decision starts with agencies
2. Ship Weekly — Ship often
3. Earn Trust — Be transparent with agencies

✅ Concise:

## Mission
Empower agencies to do their best creative work.

## Values
1. Agencies First
2. Ship Weekly
3. Earn Trust

Savings: Removes redundant explanations

4. Link to Docs, Don't Copy Them

❌ Copying entire docs:

## Technical Architecture
[5 pages of system design]

✅ Summarize + link:

## Technical Architecture
- Next.js frontend, Node.js API, PostgreSQL
- Event-driven with queues for async work
- Details: [Link to architecture doc]

Savings: Keeps context focused, detailed docs accessible

Quality vs Cost Tradeoffs

Too Little Context = Generic Output

Thin context:

## Company
We make project management software.

Result: Generic output with lots of assumptions

• "Your product could help teams..."
• "Consider features like task lists..."
• No specific industry insights
• Misses your competitive positioning

Just Right Context = Specific Output

Rich context:

## Company
**Luma** is AI-powered project management for creative agencies.
We help 50-person agencies manage client projects without chaos.
Revenue: $4.2M ARR, Series B stage ($25M raised).
Current priority: Reduce churn (currently 8% monthly, target 5%).

Result: Specific, relevant output

• "For agency clients who need real-time project visibility..."
• "Unlike Monday.com's generic approach, Luma's agency-specific features..."
• Recommendations aligned to churn reduction goal
• Pricing/packaging fits Series B stage

Too Much Context = Noise

Bloated context:

## Company History
Founded in 2019 by Jane (CEO) and Bob (CTO).
Jane worked at Google for 10 years where she led...
[5 more paragraphs of history]

We tried a B2C approach in 2020 but pivoted...
[Full pivot story with timeline]

Our office is in San Francisco at 123 Main Street...
[Office details, perks, culture]

Result: Wastes tokens, doesn't improve output

• Historical context rarely affects current decisions
• Claude gets distracted by irrelevant details
• Slower processing, higher cost

Better:

## Company
Luma (founded 2019) makes project management for creative agencies.
Pivoted from B2C to B2B in 2021, now serving 50-person agencies.

Savings: 300 words → 25 words, same decision quality

Context Refresh Strategies

When to Update Context

✅ Update Immediately

• Pivot or major strategy change — Old context misleads skills
• New product launch — Core offering changed
• Updated OKRs/goals — Priorities shifted
• New competitive threat — Landscape changed

✅ Update Quarterly

• Company priorities — Align to new quarter goals
• Product roadmap — Reflect shipped features
• Team structure changes — New hires, reorgs

✅ Update As Needed

• New personas discovered — User research insights
• Competitor positioning shift — Market moves
• Backlog priorities change — Reprioritization

❌ Don't Update

• Minor feature tweaks (not core to product)
• Small team changes (IC hires)
• Temporary projects (one-off initiatives)

Examples: Good vs Bad Context

Example 1: personas.md

❌ Too Thin

## Persona 1
- Name: Sarah
- Role: Marketing manager

Problem: No actionable insights for skills

❌ Too Bloated

## Persona 1: Sarah the Marketing Manager

Sarah is 32 years old, lives in Austin, has 2 kids and a golden retriever named Max. She's married to Tom, who works in tech. They enjoy hiking on weekends and trying new restaurants.

She wakes up at 6am, checks email while making coffee, drops kids at school at 7:30am, arrives at office by 8:15am. She uses a MacBook Pro, iPhone 13, and AirPods. Her desk has a succulent plant and a framed photo of her family.

Her favorite tools are HubSpot (for email), Canva (for design), Slack (for team chat), Google Workspace (for docs), Asana (for tasks), and Zoom (for meetings). She also sometimes uses Trello and Monday.com...

[3 more paragraphs of life story]

Problem: 95% irrelevant details, wastes tokens

✅ Just Right

## Persona 1: Agency Marketing Lead

**Who:** Marketing lead at 20-50 person creative agency
**Jobs:** Manage 5-10 client campaigns, report to clients, wrangle freelancers
**Pains:**
- Clients want real-time visibility (currently manual updates)
- Team uses 6 different tools (context switching)
- Status reports take 4 hrs/week
**Gains:**
- Client trust through transparency
- Team alignment on priorities
- Time back for strategy work
**Tech comfort:** High (uses Notion, Asana, Slack daily)

Why this works:

• Focuses on jobs, pains, gains (actionable)
• Removes demographics (not decision-relevant)
• Quantifies pain (4 hrs/week)
• Shows tool proficiency (informs UX decisions)

Word count: 80 words (vs 250+ bloated version)

Example 2: company.md

❌ Too Thin

## Company
We're a SaaS company building project management software.
Series B, growing fast.

Problem: Too vague, no differentiation

❌ Too Bloated

## Company

Luma was founded in January 2019 by Jane Smith and Bob Johnson. Jane previously worked at Google for 10 years where she led the Google Workspace PM team. Bob was an early engineer at Dropbox and built their sync infrastructure. They met at a conference in 2018 and bonded over their shared frustration with project management tools.

The company originally started as a B2C productivity app called "TaskFlow" but pivoted to B2B in March 2021 after realizing agencies had a more acute pain point. The pivot was difficult and required letting go of 15% of the team, but ultimately led to product-market fit.

We raised a $1M seed round in February 2019 from Founder Angels. Our Series A was $6M led by First Round Capital in October 2021. Most recently, we raised our Series B of $18M led by Accel Partners in January 2025.

Our team has grown from 12 people to 85 people over the past 2 years. We're headquartered in San Francisco at 123 Main Street, but we're remote-first. About 60% of our team works from home across the US, 30% come to the SF office, and 10% are international.

Our office has an open floor plan with standing desks, a fully stocked kitchen, and a rooftop deck with views of the bay. We offer unlimited PTO, health insurance, 401k matching, and a $1000/year learning budget...

[5 more paragraphs]

Problem: Founder backstory, pivot details, office perks irrelevant to PM work

✅ Just Right

## Company

**Luma** is AI-powered project management for creative and marketing agencies.

**Stage:** Series B ($25M raised, January 2025), 85 employees
**Revenue:** $4.2M ARR, growing 85% YoY
**Customers:** 340 agencies (20-100 person size)

**Business model:**
- SaaS subscription, per-seat pricing
- Pro: $29/user/month, Business: $59/user/month
- Product-led with sales assist for Business tier

**2026 Priorities:**
1. Win the agency vertical (become default choice vs Monday.com)
2. Expand AI capabilities (planning, scope detection, resource optimization)
3. Move upmarket (target 50-200 person agencies, increase ACV to $30K)

**We win when:** Agency evaluates Monday.com but finds it too generic
**We lose when:** Agency wants all-in-one (CRM + PM + invoicing)

Why this works:

• Clear positioning (agency-specific PM)
• Stage and traction (informs strategy)
• Current priorities (guides roadmap skills)
• Win/loss criteria (competitive context)

Word count: 130 words (vs 500+ bloated version)

Optimization Checklist

Before adding text to context files, ask:

• Is this decision-relevant? (Does it affect current product management work?)
• Is it current? (Historical context rarely matters)
• Can I say it in fewer words? (Bullet points > paragraphs)
• Is it already implied? (Remove redundancy)
• Does it change often? (If yes, will I keep it updated?)

If answer is "no" to #1-2 or "yes" to #4-5, cut it.

Measuring Impact

Before Optimization

• company.md: 1,200 words
• product.md: 900 words
• personas.md: 600 words
• competitors.md: 500 words
• Total: 3,200 words ≈ 4,300 tokens

Cost per skill: ~5,500 tokens input 50 skills/month: 275,000 tokens = $0.83/month (Sonnet)

After Optimization

• company.md: 600 words (-50%)
• product.md: 400 words (-56%)
• personas.md: 300 words (-50%)
• competitors.md: 300 words (-40%)
• Total: 1,600 words ≈ 2,150 tokens (-50%)

Cost per skill: ~3,650 tokens input (-34%) 50 skills/month: 182,500 tokens = $0.55/month (Sonnet)

Savings: $0.28/month + faster responses + same quality

Advanced: Dynamic Context Loading

Future enhancement: Load only relevant context per skill.

Concept:

• PRD skills → load product.md, personas.md
• Competitive skills → load competitors.md
• Planning skills → load goals.md, backlog.md

Benefits:

• Further token reduction
• Faster processing
• Lower cost

Status: Not yet available in Second, coming in future release

Related Guides

• Using Agent Teams — Agent teams multiply token usage
• Enrich Your Context — Add more detail to your context files

Last updated: February 2026