Name the Outcome
Describe what a successful answer lets you do. “Compare these options so I can choose one” is stronger than “tell me about these options.”
Field Guide // skynet.tips
Useful AI Habits
Small changes that make model output clearer, safer, and easier to verify. The first section applies everywhere; the rest are grouped by the kind of system you are working with.
Section 01
Universal Habits
Supply Source Material
Give the model the text, data, or policy it should use. Ask it to separate sourced facts from assumptions.
Constrain the Shape
Specify audience, length, format, exclusions, and the decision criteria that matter.
Ask for Uncertainty
Request unknowns, weak assumptions, and what evidence would change the answer. Confidence is not proof.
Verify Outside the Model
Check consequential facts against primary sources. A polished citation can still be invented or misapplied.
Share Less Data
Remove secrets, personal information, private files, and identifiers before sending material to any AI service.
Break Work Into Checks
Ask for a plan, inspect it, then proceed in small steps with acceptance criteria. Do not delegate an opaque leap.
Keep the Human Veto
Require approval before messages, purchases, deployments, deletions, or decisions that materially affect people.
Section 02
Chat Assistants
Start a Fresh Thread When the Subject Changes
Long conversations carry old assumptions forward. Opening a new thread is often faster than correcting a polluted one.
Say Who the Answer Is For
Audience sets vocabulary, depth, and what can be assumed. A draft for a board of non-specialists is not the draft for your team.
Show One Example of the Result You Want
A single sample communicates format, tone, and level of detail faster than a paragraph describing them.
Ask It to Restate the Task First
A one-line restatement exposes a misread requirement before the model spends effort answering the wrong question.
Correct, Do Not Restart
Point at the specific sentence that is wrong and say why. Rewriting the whole prompt discards the parts that already worked.
Distrust Instant Agreement
Models often fold when challenged, including when they were right. Ask for the evidence rather than a revised opinion.
Section 03
Reasoning Models
Give the Goal, Not the Recipe
Systems that plan internally do better with a clear objective and hard constraints than with prescribed steps.
Spend Reasoning Models on Hard Problems
Extended reasoning costs time and money. Routine drafting and formatting rarely justify it.
Judge the Answer, Not the Narration
Displayed reasoning is often a summary, and even when shown in full it is not a faithful record of the computation. It can read as rigorous and still reach a wrong conclusion.
State What Must Hold
List the invariants: budget, deadline, format, forbidden options. Explicit constraints are honored more reliably than implied ones.
Ask for the Options It Rejected
What the model discarded, and why, often says more about the quality of the answer than the conclusion itself.
Section 04
Agents and Tool Use
Decide Permissions Before Capability
Settle what an agent may touch before deciding what it may do. Start read-only and widen deliberately.
Prefer Reversible Actions
Drafts instead of sends, staging instead of production, recoverable deletes instead of permanent ones. Make mistakes cheap.
Require a Plan You Can Read
Approve the intended sequence of tool calls before execution, not after the side effects have landed.
Put a Budget on the Loop
Cap steps, elapsed time, and spend. An agent without a stop condition will eventually find one you did not choose.
Log Every Tool Call
Record inputs, outputs, and timestamps. An agent you cannot audit is one you cannot debug, explain, or defend.
Treat Fetched Content as Untrusted Input
Web pages, documents, and tool results can carry instructions aimed at your agent. Retrieved data is evidence, never authority over the task.
Keep Production Credentials Out of Reach
Give agents scoped, short-lived, least-privilege access. Blast radius is a design decision you make in advance.
Section 05
Coding Assistants
Read It Like a Stranger Wrote It
Code that compiles can still be wrong, insecure, or inconsistent with the conventions of your codebase.
Ask for the Test With the Code
Tests state intent and surface the edge cases an implementation quietly skipped.
Paste the Real Error, Not a Paraphrase
The full message, the failing input, and the versions in use beat any description of the symptom.
Pin the Versions You Actually Run
Training data ages. Name the language, framework, and library versions or expect confident calls to deprecated APIs.
Check License and Provenance Before Shipping
Generated code can echo a licensed source. Verify before it becomes part of a product you distribute.
Keep Secrets Out of the Prompt
Keys, tokens, connection strings, and customer records do not belong in a chat window — or in the repository afterward.
Section 06
Document and Retrieval Assistants
Demand Quotes and Locations
Ask for the exact passage and where it came from, rather than a confident paraphrase you cannot trace.
Confirm Retrieval Actually Happened
When a search returns nothing useful, a model may answer from memory instead. Ask what it retrieved before trusting the answer.
Curate the Corpus
Superseded policies and abandoned drafts produce confidently outdated answers. Retrieval inherits your filing discipline.
Ask What It Could Not Find
Absence is information, especially when you expected a document to exist.
Remember That Chunking Decides Findability
A fact split across two fragments may never surface. When answers miss obvious content, inspect how the documents were divided.
Section 07
Image and Video Generation
Describe Subject, Composition, and Light
Concrete direction outperforms a stack of style adjectives.
Change One Variable at a Time
Iterating on a single element makes it clear what actually improved the result.
Do Not Imitate a Living Artist by Name
Describe the qualities you want instead. A signature style is someone’s livelihood.
Confirm Usage Rights Before Commercial Use
Ownership, indemnity, and training-data terms differ by provider and change often. Check the current terms for your plan.
Label Synthetic Media
Disclose generated imagery wherever a viewer could reasonably mistake it for a photograph.
Section 08
Voice and Audio
Get Consent Before Cloning a Voice
Recorded permission, for a stated purpose. A recorded voice can identify a person and, in many places, counts as a protected likeness.
Proofread Names, Numbers, and Jargon
Transcription errors cluster exactly where the stakes usually sit.
Expect Accuracy to Fall With Noise and Overlap
Crosstalk, unfamiliar accents, and poor microphones degrade transcripts. Sample the output before trusting it in bulk.
Do Not Accept a Voice as Proof of Identity
Convincing voice forgery is inexpensive. Confirm sensitive requests through a separate, known channel.
Section 09
Predictive and Scoring Systems
Ask What the Score Actually Predicts
Many systems predict a convenient proxy rather than the outcome you care about. The gap between the two is where harm hides.
Check the Base Rate First
A confident flag on a rare event can still be wrong most of the time. Frequency shapes what a probability means.
Watch for Drift
Conditions change after training. A model that was accurate last year can be quietly failing now.
Measure Error by Group, Not Just Overall
Aggregate accuracy hides harm concentrated in a subset of people.
Give People a Way to Appeal
Any automated decision about someone should carry a human route to review, explain, and correct it.
Section 10
Privacy and Accountability
Know Where Prompts Go and How Long They Stay
Retention and training-use differ by product, plan, and region, and they change. Verify before sending anything sensitive.
Keep Regulated Data in Approved Systems
Health, financial, student, and government records carry obligations that a convenient chat window does not satisfy.
Treat Output as Draft, Not Professional Advice
Generated text is not legal, medical, financial, or safety advice. Consult a qualified professional before acting on consequential matters.
Record Who Approved What
Name the person accountable for each consequential AI-assisted decision, and when they signed off.
Disclose AI Use Where It Affects Others
Candidates, patients, customers, and students deserve to know when a system shaped a decision about them.
Plan the Rollback Before the Rollout
Decide in advance how you would detect, correct, and communicate a wrong answer that already reached many people.