AI Guideby Bubblyfi

A practical guide to better conversations with AI

Turn a rough idea into a prompt that gets useful work done.

Learn the essentials, answer a few focused questions, and leave with a complete prompt you can paste into ChatGPT.

Your answers stay on this device. Nothing is uploaded.

The learning path

Eight steps from vague request to reliable result

023 min

Add useful context

Share facts, history, and constraints that could materially change the answer.

033 min

Supply evidence

Name the files, measurements, screenshots, dates, and sources ChatGPT should use.

042 min

Define the output

Specify the format, detail, tone, length, and decision the response should support.

052 min

Clarify before solving

For ambiguous work, let ChatGPT ask only the questions that would change its approach.

062 min

Review and refine

Keep the good parts and request precise changes instead of restarting the conversation.

073 min

Verify important work

Ask for source checks, calculation checks, assumptions, and unresolved uncertainty.

083 min

Continue long projects

Carry forward decisions, constraints, current status, files, and the next concrete step.

Built around official OpenAI prompting guidance. Independent guide; not an OpenAI product.

Context builder

Build the brief ChatGPT needs.

0 of 8 answered

Describe the finished result, not every step ChatGPT should take.

Include their role and how familiar they are with the topic.

Add the current situation, what you tried, and what happened.

Name attached files, screenshots, measurements, trusted sites, or applicable dates.

Keep this to the boundaries that prevent real problems.

Choose the format, detail, tone, length, and order.

Define what a complete and useful answer must contain.

Spec-driven development

Define the work before the agent writes the code.

Build a project constitution and a testable feature specification, then hand the Markdown to a coding agent.

Guide inspiration
  1. 01ConstitutionMission and rules
  2. 02SpecifyBehavior and scope
  3. 03PlanSmall implementation slices
  4. 04ImplementOne slice at a time
  5. 05ValidateProve acceptance criteria
  6. 06PackageReuse the workflow
01

Project constitution

Stable decisions the agent should preserve across every feature.

02

Feature specification

Describe observable behavior without prescribing code too early.

03

Plan and validation

Ask for small, reviewable slices and evidence that the result matches the spec.

Legacy codebase

Start with discovery, not a rewrite.

Have the agent map the current architecture, behaviors, tests, constraints, and risky areas. Turn confirmed reality into the first constitution before specifying changes.

Reusable workflow

Package repeated instructions as a skill.

When the same standards work across projects, save the steps, checks, and output format as a reusable skill. Keep project-specific facts in each project’s constitution.

Review discipline

The spec is a control surface—not magic.

Review the plan and evidence. Small slices, clear acceptance criteria, and explicit validation reduce drift; they do not remove the need for engineering judgment.

Prompt library

Start with a proven structure.

Choose a template, replace the brackets, and copy it into ChatGPT.

Tips and strategies

Give more signal, not simply more text.

The best context is the information that could change the result.

01

Separate facts from assumptions

Label what you observed, what you suspect, and what remains unknown. This reduces false certainty.

Try thisObserved: [facts]\nSuspected: [hypothesis]\nUnknown: [gaps]
02

Name the authority

Tell ChatGPT which file, person, standard, or current source takes priority if information conflicts.

Try thisTreat the approved specification as authoritative. Flag conflicts instead of choosing silently.
03

Use examples strategically

One good example can communicate tone and structure better than a long explanation. Say what to copy from it.

Try thisMatch the attached report's section order, but use the new project's facts.
04

Ask only useful questions

Clarification is valuable when an answer would change the work—not as a ritual before every task.

Try thisAsk up to three questions only if the answers materially affect your recommendation.
05

Steer instead of restarting

Point to the exact weakness, protect what worked, and request a targeted revision.

Try thisKeep the analysis and evidence. Make the opening more direct and move the recommendation first.
06

Add a final quality check

Before important work is finished, ask for a concrete check tied to how you will use it.

Try thisBefore finishing, verify every action has an owner and due date; flag anything unsupported.

Not sure what context matters?

Let ChatGPT interview you.

This prompt asks one focused question at a time, then produces a complete Markdown brief for your approval.

Vague

Stronger