Fill these in. Remove half the ambiguity from your AI project.

# Capability + boundary brief

User job:
Model-assisted task:
What the system can do reliably:
What varies by context:
What it cannot know:
What it must never imply:
Evidence required before showing a result:
Safe fallback when confidence is low:
Owner of the final decision:

# Model behaviour spec

Role in the product:
Primary objective:
Priority order when goals conflict:
Must do:
Must not do:
How to express uncertainty:
When to ask a question:
When to refuse:
When to escalate to a person:
Examples of good behaviour:
Counterexamples:
Version / owner:

# Grounding + source brief

Decision this evidence supports:
Allowed sources:
Source of truth:
Minimum freshness:
Required fields / claims:
How conflicts are handled:
What counts as missing evidence:
Citation shown to the user:
Fallback when a source is unavailable:
Data owner / review date:

# Prompt contract

Job to be done:
Inputs available:
Inputs that may be missing:
Required output:
Output schema / length:
Rules that cannot be relaxed:
Tools the model may use:
Actions it may never take:
Questions it should ask first:
Failure response:
Test examples:

# Evaluation rubric

Task / user segment:
Dataset or scenarios:
Criterion 1 + pass threshold:
Criterion 2 + pass threshold:
Criterion 3 + pass threshold:
Critical failure (automatic fail):
Reference answer or ideal behaviour:
Edge cases:
Human review sample size:
Release threshold:
Regression owner:

# Failure taxonomy

Failure name:
What the user sees:
Likely cause:
Frequency:
Severity / user harm:
Can the system detect it?
Recovery shown in the interface:
Data saved or rolled back:
When a person takes over:
How this becomes an eval case:
Owner / status:

# Human approval matrix

Action:
User impact: low / medium / high
Reversible: yes / partly / no
Money or external representation: yes / no
Evidence required:
Confidence threshold:
Auto-run allowed when:
Approval required when:
What the approval card must show:
Timeout / cancellation behaviour:
Audit record:

# Handoff + escalation rule

Trigger condition:
Why automation stops here:
Person or team receiving it:
Context already gathered:
Evidence attached:
Decision still needed:
Safe state while waiting:
What the user sees:
Response deadline:
Resume rule:
Audit owner:

# Feedback + control plan

Decision or output users can correct:
Control available before generation:
Control available after generation:
Undo / rollback:
What feedback is captured:
What feedback changes immediately:
What requires review before learning:
How the user sees the effect:
Privacy constraint:
Owner:

# Agent state map

State:
What the system is doing:
What evidence it has:
What the user sees:
Can the user interrupt?
Can the user edit the plan?
Tool or dependency in use:
Timeout behaviour:
Failure state:
Recovery action:
Next state:

# AI release + change record

Date / version:
What changed:
Model / prompt / data / tool / UI:
Reason for change:
Eval set run:
Before / after result:
New failure mode introduced:
Human review completed by:
Rollback trigger:
User-facing note:
Next review date:

# Reasoning disclosure

Status / current step:
Plan the user can inspect:
Evidence used:
Checks completed:
Assumptions:
Known limits:
What remains uncertain:
Next action:
User control: pause / edit / continue

# Tool call card

Tool / system:
Why it is being used:
Input summary:
Permission level: read / write
Status: queued / running / done / failed
Result summary:
Source or artifact link:
Side effect created:
Undo available:
Retry / fallback:

# Approval card

Proposed action:
Recipient / destination:
Exact content or change:
Evidence supporting it:
Cost / consequence:
Can it be undone?
What happens on approve:
What happens on reject:
Edit option:
Approval expires:

# Agent progress row

Step name:
State: queued / active / blocked / done
Started at:
Evidence or output:
Dependency:
What blocked means:
User action required:
Retry policy:
Completion condition:

# Source + evidence card

Claim supported:
Source title:
Source URL / record:
Publisher / owner:
Observed date:
Freshness requirement:
Exact evidence used:
Conflict with other sources:
Confidence:
Limit / caveat:

# Uncertainty + recovery state

What is uncertain:
Why it matters:
What is known:
What is missing or conflicting:
Risk of continuing:
Best next question:
Alternative safe action:
Can the user override?
How recovery is confirmed: