A useful startup context file is a compact record of your current product, customer hypothesis, constraints, evidence, and next decision. It helps an AI assistant work from the same business facts across tasks without treating every previous conversation as current truth.

The minimum useful context

You do not need a finished idea to write the context. Start with your experience, reachable people and a question. Label unknown customers or problems as open rather than filling them with guesses. Begin with enough information to change the recommendation. More text is not automatically more useful. A long business plan can hide the one constraint that matters: you can only spend ten hours this week.

  • Product: a working direction, possible problem and intended user, even if these are still unknown.
  • Founder resources: available hours, skills, budget, and reachable audience.
  • Evidence: observations with their source and date.
  • Hypotheses: beliefs you still need to test.
  • Decisions: accepted choices and why they were made.
  • Current focus: one question or experiment for the next task.

Omit credentials, unrelated personal information, and raw customer transcripts. Summarize only what is necessary for the decision and what you have permission to use.

An annotated example for a solo founder

The following is a hypothetical AgentCRM context file. Its limits are intentional: there is no invented market size, customer quote, or claim that demand is validated.

# AgentCRM — current context

## Product hypothesis
Help solo real-estate agents track inquiries and the next follow-up.
The problem and willingness to pay are not validated.

## Founder facts
Team: one engineer; limited sales experience.
Capacity: 20 hours/week. Budget: $500/month.
Audience: no owned audience; can reach local agents individually.

## Evidence
No customer interviews completed yet.

## Decisions
Start with one segment: independent residential agents.
Do not build a full CRM before testing the follow-up problem.

## Unknowns
How often do agents miss a follow-up?
What workaround do they use today?
Will they try a manual pilot?

## Current focus
Arrange five problem interviews this week.

## Update rule
Review after interviews. Keep assumptions labeled.

The numbers are planning inputs for this example. They are not a claim about typical founders. “No interviews yet” is more useful than a confident demand statement because it tells the next workflow where evidence is missing.

Load the context with a purpose

For a manual workflow, provide the file in your chosen project or attach it to the task. Ask the assistant to confirm the current facts before giving advice. ChatGPT Projects can provide shared files and instructions to related chats; consult the official Projects guide for the current interface.

Use this context to plan the next validation experiment.
First summarize the product, constraints, and evidence you used.
Keep facts, hypotheses, and unknowns separate.
If documents conflict, show the conflict before recommending a plan.
Give one next experiment that fits this week's time and budget.

With 1PC, the bundled workflows read your founder profile and relevant product files. Idea Intake uses the overview and idea brief; Validation also looks for market, competitor, ICP, and validation files. Missing evidence should remain a gap, not become an invented fact.

Keep context current without building a documentation job

Update after a material decision, rather than after every message. Write down what changed, the evidence behind it, and what the next task should do differently.

If interviews change the intended buyer from an individual agent to an agency owner, update the customer hypothesis and affected scope. Preserve any still-valid constraints. Do not create several files called final-v2 and let the next assistant guess which is current.

One compact context file works at the beginning. Split it into product overview, customer hypothesis, and validation findings when each has a clear purpose. Read the existing content before replacing it. 1PC stores current content and retains its previous revisions.

Run a context check before the next task

If your next task happens in another AI tool, use the guide to sharing product context across AI tools. It explains the portable file method, saved-state boundaries and MCP connection checks.

  1. Can you identify the current customer and decision in one minute?
  2. Does each important claim have a source or a hypothesis label?
  3. Are the resources still realistic for this week?
  4. Are rejected directions clearly separated from accepted ones?
  5. Would an assistant know what to ask, rather than fill a gap with a guess?

If those answers are clear, use the founder workflow hub to choose the next task. If demand is still uncertain, start with an evidence-led validation plan.

Write your context once. Keep it useful.

Use the free template to capture your strengths, constraints, a working direction and the question you want to investigate.

Get the founder context template ↗