Here is a test you can run right now.
Ask your AI tool: “Draft a follow-up email to a prospect who attended our recent webinar, expressing interest but not yet ready to commit.”
The output you receive; without any additional context; is what the AI produces for every company in every industry for every webinar follow-up. It is the mean. It is competent, professional, and completely generic.
Now imagine the same prompt with a context pack loaded.
The AI knows your company writes with a direct, peer-to-peer tone; that your prospects are COOs at manufacturing companies; that your webinar covered AI implementation for distribution companies; and that your standard follow-up tone is warm but never pushy.
The second output is not the mean. It is specific to your company, your prospect, and your moment.
A context pack is the document that converts the second scenario from a theoretical possibility into the default for every session your team runs.
It is the single most important document in your AI system; not because it is technically complex, but because everything else the system produces depends on it.
What a context pack is not: clearing the misconceptions
A context pack is not a system prompt
A system prompt is a technical instruction given to an AI model at the start of a session to configure its behaviour. System prompts are typically brief, technical, and focused on how the AI should respond.
A context pack is substantive. It contains the actual knowledge about the company, its clients, and its operations.
The practical distinction: “You are a professional business writing assistant” is a system prompt. The 1,500-word document that describes the company’s voice, its client types, its positioning, and its decision rules is a context pack.
A context pack is not a brand guide
A brand guide describes the company’s visual identity; logo usage, colour palette, typography, brand values at a high level. A context pack is operational.
The practical distinction: a brand guide might say “our tone is professional but approachable.” A context pack says:
“In client proposals, we lead with the specific operational problem the client described; not with our credentials. We use the client’s own language where possible. We make one clear recommendation rather than presenting options. Our closing always names the specific next step.”
A context pack is not a knowledge base
A knowledge base is the company’s accumulated documentation; SOPs, policies, product specs, client histories, FAQ entries. A knowledge base is designed for retrieval; answering specific questions when queried.
A context pack is designed for orientation; loading the AI’s foundational understanding of the company before any specific work begins.
The practical distinction: the context pack (1,500–3,000 words) is loaded at the start of every session. The knowledge base is queried when a specific question requires a specific answer. They work together; the context pack provides orientation, the knowledge base provides specific facts.
The six components of a context pack: what each one does
Component 1: Company identity
What it contains: a factual, specific description of the company; what it does, who it serves, what it has achieved, and what makes it different from alternatives.
What a basic version requires:
- One paragraph on what the company does and for whom (specific, not generic)
- One paragraph on the company’s results and proof (specific outcomes, named client types)
- Two to three sentences on what makes the company different (not what every company in the category would say)
What it enables: every AI output that references the company is grounded in accurate, specific information.
The AI does not describe the company generically as “a professional services firm.”
It describes Phos AI Labs as “an embedded AI strategy firm for $5M–$25M non-tech companies that builds AI foundations, trains teams, and installs AI-native operations through a four-phase engagement model.”
Length: 250–400 words.
Component 2: Client archetypes
What it contains: two to four detailed descriptions of the company’s best client types; their role, industry, specific concerns, how they communicate, what they are trying to achieve.
What each complete archetype contains:
- Role and industry (specific)
- Typical revenue range
- Current AI situation (where they are on the maturity curve; what they have tried that has not worked)
- The trigger that brought them to this conversation
- Their primary concern going into the first conversation
- How they communicate (direct vs. indirect, formal vs. casual, data-driven vs. intuitive)
- What a successful outcome looks like for them specifically
- One or two things they will say in a first conversation that signal good fit
What it enables: AI outputs calibrated to the specific person receiving them.
A proposal for a COO at a $22M engineering consultancy reads differently from one for a founder at an $18M distribution company; and the AI produces the difference when the archetypes are loaded.
Length: 100–150 words per archetype.
Component 3: Voice and tone standards
What it contains: specific, operational instructions for how the company writes across different output types; not abstract tone descriptors but concrete writing standards.
What a basic version requires:
- A three to five sentence description of the overall voice register
- A vocabulary guide: ten to fifteen words and phrases the company uses, ten to fifteen it avoids
- Format standards by output type (how a client email differs from a proposal from a LinkedIn post)
- Three to four “avoid these patterns” examples
What it enables: AI outputs that require minimal tone editing. The twenty minutes of tone adjustment that currently precede every AI draft is eliminated when the voice guide is loaded.
The output starts on-brand.
Component 4: Competitive positioning
What it contains: how the company honestly describes its position relative to the alternatives its clients are considering.
What a basic version requires:
- The one to two things the company does that alternatives consistently fail to provide
- The alternatives the company is most often compared to; and the honest distinctions
- What the company explicitly does not claim (more credible than listing what it does)
- The specific situation in which the company is the right choice
What it enables: AI outputs that make the right implicit comparisons without being defensive or aggressive.
Length: 200–350 words.
Component 5: Product and service descriptions
What it contains: accurate, current descriptions of what the company offers; in the specific language the company uses, with the specific outcomes each offering produces.
What a basic version requires:
- One to two paragraphs per offering: what it is, what it produces, who it is for, and the specific outcomes it delivers
- The language the company uses to describe each offering; and the language it does not
What it enables: AI outputs that describe the company’s offerings accurately and in approved language. Proposals that describe the engagement correctly. Sales emails that reference the right service for the prospect’s situation.
Component 6: Decision rules
What it contains: the documented logic for the most common judgment calls the AI will encounter; what the company does in specific scenarios that recur regularly enough to be worth documenting.
What a basic version requires:
- Fifteen to twenty common scenarios with the company’s standard position for each
- Written as explicit rules (“When a client requests a scope change mid-project, we: quote the change at standard rates, offer a discounted rate for existing retainer clients at the discretion of the account lead, or escalate to the founder for changes above $5,000”)
- The three to five scenarios that require escalation and should never be handled by the AI independently
What it enables: AI recommendations and draft communications consistent with how the company actually operates; not consistent with generic professional best practice.
Before and after: the same prompt with and without the context pack
The scenario: a $15M engineering consultancy is sending a follow-up email after a discovery call with a potential client; a VP of Operations at a $22M specialty manufacturer who described their main challenge as “AI has been on the agenda for two years but nothing has actually been implemented.”
Without the context pack
Subject: Following up on our conversation
“Thank you for taking the time to speak with us today. It was great learning more about [Company Name] and the challenges you are facing. At [Firm Name], we specialise in helping companies like yours implement AI solutions that drive efficiency and growth.
We would love to continue the conversation and explore how we might be able to support your AI journey. Please let me know if you would be open to a follow-up call next week.
Best regards, [Name]”
This output is grammatically correct, professionally formatted, and completely generic. It could have been sent by any firm to any prospect. The prospect has probably received fifteen emails that sound like this in the last month.
With the context pack loaded
Subject: What we heard on the call; and what we’d suggest
“Tom,
You’ve been at the AI question for two years without implementation. That’s one of the clearest signals that the problem isn’t awareness or intent; it’s the foundation work that hasn’t been done before the tools go in.
Most companies in your position have plenty of interest and zero AI Foundations. The result is every tool deployment that gets cancelled after 90 days when nobody’s using it.
What we’d suggest as a starting point isn’t a roadmap or a tool recommendation. It’s a four-week Foundations engagement that documents how your operations actually work and builds the context layer that makes AI useful for your team; not generic.
I’ll send a brief description of what that looks like. If it resonates, we can schedule a second call.
[Name]”
Same prompt. Same model. The difference is everything.
What produced the difference
- Company identity told the AI the firm is the embedded AI strategy firm for $5M–$25M non-tech companies
- Client archetype told the AI “Tom” is a VP of Operations at a manufacturer who has had AI on the agenda without implementation; a specific, documented archetype
- Voice guide told the AI the firm writes directly, leads with the client’s specific problem, makes one recommendation rather than options, and uses a peer-to-peer register
- Product description told the AI what the Foundations engagement is and when it is the right starting point
The AI assembled those four inputs into a specific, on-brand email in under 30 seconds.
Building your context pack: the process and the time investment
Who should write it
The founder or COO writes it. Not a copywriter, not a junior team member, not the AI itself.
The context pack reflects the company’s accumulated strategic judgment; who the clients are, how the company actually positions itself, what the company’s real decision logic is. Only the people who make those judgments can document them accurately.
The AI can assist with drafting once the founder has provided the raw material. But the raw content must come from the person who knows the company.
Build order and time estimates
| Component | Suggested order | Estimated time |
|---|---|---|
| Company identity | 1 | 45–60 minutes |
| Client archetypes | 2 | 60–90 minutes (two to three archetypes) |
| Voice and tone standards | 3 | 45–60 minutes + example selection |
| Competitive positioning | 4 | 30–45 minutes |
| Product and service descriptions | 5 | 30–60 minutes |
| Decision rules | 6 | 60–90 minutes (fifteen to twenty rules) |
| Total | 5–7 hours |
The test that tells you the context pack is working
Load the completed context pack into your shared AI workspace. Ask the AI to draft the follow-up email from the before/after example above (adjusted for your company and your prospect type).
If the output is specific, on-brand, and requires less than five minutes of editing: the context pack is working.
If it still reads like the generic version: specific components need more detail or more concrete example content.
The first update that tells you the context pack needs maintenance
When a team member makes a factual correction to an AI output; the AI referenced a service the company no longer offers, used pricing that has changed, described a process that was updated six months ago; that correction identifies a context pack entry that needs updating.
The update should happen the same week.
Common questions on AI context packs
”Can I use the context pack across multiple AI tools?”
Yes. The context pack is a plain-text document. It can be uploaded to Claude Projects, added to a ChatGPT custom GPT, loaded into any API-based workflow, or pasted into any AI session.
The context pack is a company asset; not a tool-specific asset. Migrating AI providers means re-uploading the document; not rebuilding the context.
”How long should a context pack be?”
A complete context pack with all six components: 1,500–3,000 words. This is short enough to load in every session without consuming the context window; long enough to cover the specificity that makes outputs company-specific.
The most common mistake is writing too short (800–1,000 words) and missing the decision rules and vocabulary guide entirely.
”What if my company is still defining its voice and positioning: can I still build a context pack?”
Yes; build it with what is known and mark the sections that are in progress. An 80%-complete context pack produces significantly better outputs than no context pack.
The risk of waiting: the voice and positioning evolve through use. The best way to clarify what the company sounds like is to produce outputs and react to them. The context pack is the tool for that iteration.
”Is a context pack the same as a Claude Project’s instructions?”
Claude Project instructions are one place where the context pack is loaded. They are not the same thing.
The context pack is the document that exists in the company’s own storage (Google Drive, Notion). The Claude Projects instructions field is one deployment point for that document.
If Claude is ever replaced by another tool, the context pack survives. If only the Claude Projects instructions were maintained, the context would not be easily transferable.
”How do I keep the context pack updated as the business changes?”
The AI system owner reviews the context pack as part of the weekly maintenance cadence. Any AI output that contains a factual error (wrong pricing, outdated service description, inaccurate policy reference) triggers an immediate update.
A formal review of the entire context pack happens quarterly; checking whether the company identity, client archetypes, and competitive positioning still reflect the current business.
”Should different team members have different context packs?”
The core context pack is shared; it describes the company, not the individual. Layer-three situational context (client-specific or project-specific information) varies by team member.
The layered framework: the shared context pack loads at every session (company identity, voice, decision rules). The team member adds situational context specific to their client or project. The combination produces outputs that are both on-brand and situation-specific.
Want your context pack built correctly: with every component at the depth that makes AI outputs specific from day one?
A context pack is the document that converts a general-purpose AI tool into a company-specific one.
It takes one focused day to build. The impact on output quality is immediate.
The before/after difference is not marginal.
It is the difference between a generic output that requires twenty minutes of editing; and a specific one that requires five.
Every AI session your team runs without a context pack loaded is a session producing outputs calibrated to the average company. Every session with one loaded is producing outputs calibrated to yours.
Path one: start with the company identity section today. Write 250–400 words answering: what does the company do, who does it serve, what makes it different, and what results has it produced? Load that one section into a Claude Project and run the before/after test. The difference tells you everything about why the context pack matters.
Path two: bring in a partner. If you want all six components built at the depth that makes AI outputs specific from day one; that is the Phos AI Labs Phase 1 Foundations engagement. Four weeks; context pack, voice guide, operating rules, and workflow documentation built, loaded, and tested against real work before handover. We have run 400+ AI engagements. Clients include Zapier, Coca-Cola, Medtronic, Dataiku, and American Express. Start here.