An AI-ready project brief: the template that stops endless prompting
An AI project brief connects outcomes, scope, sources, constraints, and validation. Use this template before assigning work to an agent.
6 min read
Adopting AI in a business starts with a piece of work worth improving. Preparing a project update, turning requests into actionable tasks or finding the decisions behind a change are concrete candidates. Model selection comes after defining the outcome and the conditions for success.
A first AI project should matter to the team while remaining small enough to compare with the current method. It also needs an owner who can decide to expand, adjust or stop it.
Choose a task the team recognizes
Begin with an existing frustration. Which work repeats? Which sources get reread? Where do people lose time moving information between tools? Ask users to show a recent occurrence, including its inputs and result.
“Put AI everywhere” does not explain what to build or how to evaluate it. “Prepare a project report from cards and decisions for an owner to review” identifies an input, a reader and an observable outcome.
France Num's AI resources offer a starting point for small businesses. For the actual pilot, choose one work situation rather than a general catalog of possibilities.
Scroll to read all columns
| Candidate | Useful first version | What to inspect |
|---|---|---|
| Project report | Sourced draft without distribution | Accurate facts, visible risks and clear decisions |
| Customer request | Proposed cards without delivery commitments | Preserved scope, open questions and acceptance criteria |
| Documentation | Suggested update linked to a change | Correct information and current sources |
| Code | Bounded fix in an agreed environment | Working journey and relevant regression checks |
Write a short pilot brief
One page can identify the user, current task, desired outcome, sources, necessary access and acceptance criteria. Add situations where a person must take over.
Name the business owner, the user who will review results and the person responsible for tool access. A small business may assign several roles to the same person. Naming them still prevents a data question from waiting on someone who thought they were only proofreading text.
The AI-ready project brief provides a structure. Keep the pilot mission bounded. If the initial objective is a report, automatic customer distribution and schedule changes can remain separate decisions.
Each week, an owner gathers cards, reads comments and asks contributors for missing details. The AI pilot prepares a draft separating accepted deliveries, work awaiting review, blockers and decisions. A person checks and distributes it. The agent changes no deadlines and sends no messages.
The result is useful if preparation becomes easier without moving all the effort into corrections. Faster generation is not a gain when the report takes longer to verify.
Establish a baseline before adding AI
Observe several representative examples of the current method. Record preparation time, review time, errors discovered and necessary follow-ups. Keep the inputs so the same task can be attempted with AI.
Include an ordinary case and a difficult one: a missing document, ambiguous request, outdated source or conflicting instruction. A demonstration built entirely around perfect inputs says little about everyday use.
Measurements should inform a decision. If the problem is reporting accuracy, token counts are insufficient. If the problem is repetitive work, measure total human time, including data preparation and result correction.
Prepare the data and access
Provide useful sources with their dates and status. Distinguish an accepted decision from an idea under discussion. Make an old document identifiable before it becomes the basis of a result presented as current.
Check the access the mission needs. Reading cards to write a draft does not require deleting them or managing workspace members. Effective access should match the mission and the identity executing it.
Choosing local, cloud or hybrid AI depends on the processing and data involved. A desktop application may still call a cloud model. Local inference does not mean the entire collaborative project works offline. Describe the actual data path in the selected configuration.
Run the first cycle with an explicit review
Return the first result to someone who knows the task. Ask them to inspect criteria rather than give a general impression. For a report: are states accurate, links accessible, decisions real and unknowns visible?
Record discrepancies. An error can come from the model, an old source, missing access or a poorly defined mission. Rewriting the prompt after every failure can hide a data or process problem.
The agent work verification guide helps structure this review. The person accepting the result should be able to explain why the criteria passed without reconstructing the whole conversation.
Move from one assistant to an agent team
One agent is often enough for the first use case. Add a role when its contribution is clear: an independent reviewer can inspect the report, an explorer can gather sources and a corrector can address confirmed discrepancies.
Coordination becomes useful when missions depend on each other. Who can start, who needs an upstream result and who accepts delivery? Stellary's project piloting defines a mandate and an authorized team. The pilot delegates work and results go through review.
An agent supervisor workflow fits a repeatable circuit: prepare the report, check sources, correct discrepancies and request approval. Project coordination organizes the next work; a workflow reproduces a known method. They can be combined.
Decide using comparable outcomes
Compare the same inputs against the same criteria. Track time to an accepted result, remaining errors, observable costs and human interventions. Distinguish a partially accelerated task from one the system handles completely.
There are three reasonable decisions. Expand if outcomes are sufficiently consistent. Reduce or change scope if a specific problem blocks quality. Stop if review and maintenance cost more than the original task.
Do not invent a universal numeric target. The acceptable level depends on the task. An internal draft reviewed before use has different requirements from an action committing a customer, invoice or access right.
Help the team adopt the method
Explain what the agent does and what users need to inspect. Show a successful example and an example where it asks for information. Training should focus on the team's actual decisions, alongside the tool's controls.
Plan for AI being unavailable. Who resumes the task? Where are its inputs and last accepted result? A simple fallback prevents the pilot from creating an opaque dependency on a conversation or one person's computer.
Keep useful corrections in reviewed documentation. Project instruction files serve coding executors; business rules and decisions belong in shared documents. Repetition should make the system easier to use without accumulating contradictory instructions.
Start in Stellary
Create a project, add the brief and prepare a card with acceptance criteria. Configure an agent and its access, then run a bounded mission. Link the result to the card and keep the review decision with it.
The project status report template is a useful first use case when reporting is the problem. The coordination guide becomes relevant when several contributors need to work together. Evolve the system from accepted outcomes and the difficulties the team actually encounters.
Answer the business owner’s operational questions
Before expanding, make responsibilities understandable. Who can change instructions? Who reviews a result when the usual owner is absent? Where should an error be reported? Keep these answers accessible without relying on the one person who knows the entire setup.
Also inspect costs absent from model billing: preparation, review, training and maintenance. A pilot can be valuable because it improves continuity or prevents information loss. Do not force those benefits into an unsupported productivity percentage.
Separate personal convenience from a shared process
An employee may already use AI to draft notes or inspect code. Turning that habit into a team process requires shared inputs, agreed access and a place to review the result. A successful personal experiment does not automatically tell colleagues how to handle failures or customer information.
Invite the people doing the work to revise the method after the first trials. Their corrections often reveal missing definitions: what counts as a delivery, when a decision is authoritative or which exception requires a person. Add those definitions to shared documents instead of leaving them in a private conversation.
When a task changes, repeat a small set of reference examples. This can expose an obsolete assumption before the workflow spreads to additional projects. Expansion should preserve the result users trusted in the initial pilot, with an explicit reason for any new access or broader authority.
