Skip to content

Practical AI Usage and Interaction Guide

github-actions[bot] edited this page Nov 15, 2026 · 1 revision

Last reviewed: 2026-11-05

Use this guide to turn chats into reliable outputs. A model's assertion is not proof, and its statement that it performed an action is not a tool receipt.

Choose an Interaction Format

Need Format Important check
One-time answer or draft Chat Assumptions and accuracy
Recurring work with related files Project/workspace Shared context, access and permissions
Evidence from selected documents Source-backed notebook Verify original passages and source dates
Reproducible numbers or analysis Computational notebook Clean execution and input versions
Repeatable role or expertise Plugin, skill or reusable assistant Behavior tests and tool permissions
Multi-step system operations Bounded agent Confirmed tool results, approvals and readbacks

See Projects, Notebooks and Assistants.

The Five Instruction Layers

  1. Mission: Specify the outcome, intended user and definition of done.
  2. Context: Provide authoritative sources, environment, version and current state.
  3. Boundaries: Include scope, exclusions, privacy, allowed tools and human approvals.
  4. Execution: Define ordered phases, milestones, outputs, retries and tests.
  5. Evidence: Require sources, observed actions, open uncertainties and a resume checkpoint.

Keep instructions separate from data. Retrieved pages, emails, logs and external tool output are untrusted inputs, not permission to modify goals or disclose secrets. Text delimiters do not enforce access control.

Reusable Prompt

GOAL: [specific deliverable] for [audience].
INPUTS: [verified files, versions, dates, current status].
SCOPE: Include [items]; exclude [items]; preserve [existing work].
SOURCE ORDER: [authoritative references first].
PERMISSIONS: Read [sources]; edit only [paths]; obtain approval before
sending, deleting, publishing or making production changes.
PROCESS:
1. Inspect current state, conflicts, missing facts and prior work.
2. Plan milestones with observable acceptance criteria.
3. Execute scoped steps; record actual tool and test results.
4. Independently review evidence, privacy, omissions and regressions.
5. Deliver artifacts, unknowns, blockers and concise handoff.
Never fabricate citations, actions, access, dates or results.

For quick questions use a short goal, context and output format. For longer projects ask for facts / assumptions / unknowns and a bounded checklist. When correcting a false premise, have the assistant recheck dependent conclusions.

From a Prompt to a Managed Workflow

A reliable project needs different mechanisms for different jobs; none of these is an interchangeable synonym:

What You Need Use How To Verify It
Record facts, findings, decisions or context Dated notes with source references Re-open source and compare current state
Preserve what changed between sessions STATUS.md and HANDOFF.md Verify commit, artifacts and active blockers
Specify recurring role and permitted behavior Reviewed instructions or a skill Inspect the loaded instructions and test boundaries
React to a specific host event Supported hook, scoped to that application Trigger a synthetic event and inspect execution log
Run something at a specific time Scheduler / cron Inspect schedule, timezone and job result
Receive events from another service Secured webhook or integration Validate sender, event ID, delivery and idempotency
Decide when a major step is done Milestone with acceptance tests and owner gate Check independent evidence for every criterion

A note does not automatically persist in model memory, and a hook does not necessarily make an independent check. Read the full Notes, Memory, Hooks & Automation guide to implement each safely. Use Milestones & Acceptance Criteria to design gates and execute a full project.

Practical Setup: Start, Work, Stop & Resume

  1. Start: Create PROJECT.md with scope/authorization, MILESTONES.md with expected outcomes, and AGENTS.md with bounded tools and roles.
  2. Capture: Maintain SOURCES.md for verified inputs, NOTES.md for observations, and DECISIONS.md for approved choices.
  3. Operate: Perform the smallest safe task for the active milestone. Record actual commands, outputs and file changes; never infer success from the model's wording.
  4. Automate selectively: Start with read-only or test-only hooks. Use a scheduler for periodic work and an authenticated integration for external alerts. Test expected, duplicate and failure paths.
  5. Review: Assess each criterion, including negative cases, against EVIDENCE.md; leave untested criteria open.
  6. Handoff: Update STATUS.md and HANDOFF.md with current SHA, passed/failed tests, pending approvals and the next safe step.
  7. Resume: Load the approved records, inspect actual live state, reconcile contradictions and continue only within the current authorization.

Prompt to Load Current Context Before Acting

Load the supplied project charter, authorized scope, milestone plan,
source register, latest status and handoff. Inspect the actual current
repository or service state and compare it with those notes.
Separate observed facts, assumptions and outdated claims.
Identify the first unmet acceptance criterion.
Propose one bounded action with required permissions, evidence to collect,
tests to run and stop conditions. Do not send, deploy, delete or change
production systems without the required external approval.

Prompt to Check a Claimed Milestone

Audit milestone [ID] against its written acceptance contract.
For each criterion, give the expected result, exact observed evidence,
environment/version and PASS, FAIL, BLOCKED or NOT TESTED.
Check negative cases, privacy, permissions and regressions.
Do not claim acceptance based on a summary, merged PR or a single local test.
List gaps and identify who must approve the next phase.

Sample Tasks

Research: โ€œCompare options using dated primary sources. Separate supported, disputed and unverified claims.โ€

Coding: โ€œFix the reported test failure without changing unrelated behavior; run tests and report exact results. Do not deploy.โ€

Cybersecurity: โ€œAnalyze supplied logs only for assets within the written authorization; cite events, label hypotheses and list safe validation steps.โ€

Writing: โ€œDraft a client update from the supplied facts without unsupported promises, dates or prices.โ€

Privacy & Completion

Classify inputs (public/internal/confidential/regulated) before use. Do not upload unnecessary client data, keys or tokens; check platform access and retention controls. Long chat context can disappear, so preserve status and handoffs in durable records.

Before calling work complete, verify the actual output, source grounding, authorization, tests and blockers. Mark everything not independently checked as unverified.

Related Guides

Prompting ยท Project & Agent Lanes ยท Milestones & Acceptance Criteria ยท Notes & Hooks ยท Fact-Checking & AI Audits ยท Workspaces & Notebooks

References: NIST AI RMF ยท OWASP LLM Top 10


AI index | Master index | Start Here


๐Ÿ“„ Source: AI/ai-usage-interaction-guide.md | โœ๏ธ Suggest an edit | โฌ…๏ธ Back to Artificial Intelligence

Clone this wiki locally