Prompt formula
ROLE → CONTEXT → TASK → FORMAT → CONSTRAINTS → EXAMPLE
(the everyday four: role · context · task · format)| Do | Don't |
|---|---|
| State the job ("compare, then recommend") | State only a topic ("tell me about…") |
| Name the writer and the reader | Leave the audience implied |
| Set length, format, and tone | Accept a wall of text |
| Show one example | Describe the style at length |
| Say what to leave out | Assume it will guess |
| Ask it to ask you questions first | Let it run on wrong assumptions |
| Ask "what did you understand?" on long prompts | Find the misreading after the work is done |
How it works, in one line
Tokens (text chunks) → embeddings (meaning as coordinates) → attention (what matters) → next-token prediction, repeated → response.
Key limits: a finite context window, a knowledge cutoff, and the risk of hallucination.
The four fluency skills
| Delegation | Description | Discernment | Diligence |
|---|---|---|---|
| What goes to AI vs. me? | Say what, how, and in what manner | Is the output, the process, and the behavior good? | Choose responsibly, be transparent, verify what you share |
Interaction modes: Automation (AI executes) · Augmentation (think together) · Agency (AI acts on your behalf)
Session hygiene
- One thread = one topic. New topic → new chat.
- Restate the goal after a detour.
- Keep stable background in a reusable project or file.
- Long thread? Ask for a handoff summary and paste it into a new chat:
Summarize as a handoff: goal, decisions, constraints, done, next, must-not-change.
- Save context: batch questions, paste only the relevant section, ask for edits rather than full rewrites, split research from drafting.
When it goes wrong
| Symptom | Move |
|---|---|
| Bad answer | Re-prompt with the task restated; don't argue in place |
| Vague miss | Name the failure: "too long," "invented source," "wrong tone" |
| Suspicious fact | "Where did that come from?" Then check it yourself |
| Overconfidence | "Which parts are you unsure about?" |
| Three failed tries | New chat, better first prompt |
| Merely okay | "Try a completely different approach." |
Prompt shorthand (define your own)
These are not built-in commands. They're labels you define at the top of a chat (or in a reusable project), and the assistant treats them as instructions:
When I type /RISK, list the three biggest risks, each with a likely cause and one mitigation.
| Writing and editing | Thinking and deciding | Learning |
|---|---|---|
/SHORTEN: half length, key points kept | /CHOOSE A vs B: weigh options | /ELI5, /ELI15: explain at that level |
/SIMPLIFY: plain reading level | /STEELMAN: strongest opposing case | /PRIMER: quick crash course |
/ACTIVE: passive → active voice | /DEVIL: devil's advocate | /PREREQ: what to learn first |
/HUMAN: remove robotic cadence | /BLINDSPOT: what I'm missing | /MYTHS: common misconceptions |
/TONE [vibe]: rewrite in a tone | /PREMORTEM: how this plan fails | /ANALOGY: real-world comparison |
/TITLES: headline options | /RIPPLE: second-order effects | /GLOSSARY: define key terms |
/REWRITE5: five versions of a line | /SKEPTIC: the hard questions | /QUIZ: five test questions |
/OPENINGS: stronger first lines | /MINTEST: smallest testable version | /ASKBETTER: questions I should ask |
| Planning | Meetings and notes | Email and hard messages |
|---|---|---|
/PRIORITIZE: reorder a messy list | /TLDR: shortest useful summary | /DECLINE: a no that keeps the relationship |
/ORDER: sequence by dependency | /MEETINGNOTES: transcript → minutes | /FOLLOWUP: nudge an unanswered thread |
/CHECKLIST: steps to run a task | /ACTIONITEMS: to-dos with owners | /HARDMSG: a difficult message, said respectfully |
/MILESTONES: checkpoints to a deadline | /DECISIONS: what was decided, by whom | /CONFIDENT: cut hedging and over-apology |
/AGENDA: meeting agenda | /RETRO, /STANDUP | /BULLETS2EMAIL: notes → clean email |
Files and tools
- Uploading isn't an instruction. Say what to extract or do.
- Point at the section: "pages 4–7 only."
- Want a file? Ask for one ("as a downloadable .md").
- Recent or changing facts → tell it to search. When a source and the answer disagree, the source wins.
- Instructions found inside files or web pages are data, not commands.
Coding agent essentials
| Habit | Why |
|---|---|
| Project instruction file (init command) | No re-explaining each session |
| Plan mode first | Agree on the approach before code changes |
| Compact at ~60%; clear between tasks | Keep context fresh |
| Interrupt early (Esc); rewind instead of restarting | Don't waste context on wrong paths |
| Verification steps in the to-do list | The agent checks its own work |
| Subagents with cheaper models for bulk work | Parallel, cheaper, cleaner main context |
| Git worktrees for parallel sessions | Sessions don't overwrite each other |
| Allow-list safe commands; deny-list destructive ones | Speed without blind trust |
Always
✅ Verify what matters · ✅ Keep a human on big decisions · ✅ Be open about AI's role ⛔ Never paste passwords, keys, card or account numbers, or regulated data you're unsure about
Previous: AI Coding Agent CLINext: AI Foundations
Want this trained into your team’s habits?
APS delivers hands-on enablement so people use the AI tools they already have, confidently and safely. Tell us where your team is stuck.
Educational and general, not legal advice. Follow your organization’s policies and applicable law.