| 1 |
Describe the task exactly as you would to a competent temp |
Clear, literal instructions outperform clever role-play or riddles. |
Write the first draft the way you'd brief a new hire, then paste it verbatim. |
|
| 2 |
Iterate ruthlessly & read every model output |
The "engineering" part is the trial-and-error loop. Outputs teach you what the model misunderstood. |
After each run, ask "what was misinterpreted?"—tweak, rerun, repeat. |
|
| 3 |
Probe edge cases up front |
Prompts that only work on happy paths break in production. |
Feed the model empty input, malformed data, or "no matches" scenarios and specify how it should respond (e.g., unsure tag, fallback text, etc.). |
|
| 4 |
Respect the model's intelligence—don't baby it |
Modern models can digest papers and long context. Hiding complexity often makes them worse. |
Supply full docs, schemas or even research papers, then ask for the transformation you need. |
|
| 5 |
Let the model critique your prompt |
It can spot ambiguities you missed. |
Add a pass: "Ignore the task—only list unclear/ambiguous parts of my instructions." Fix those. |
|
| 6 |
Use example-based prompting strategically |
Too many similar examples overfit; illustrative, diverse ones teach the concept without constraining creativity. |
1-3 varied examples > dozens of near-identical ones. |
|
| 7 |
Give the model an "out" |
Prevents forced, low-quality guesses on inputs outside spec. |
Add: "If criteria aren't met, reply ONLY with 'UNSURE'". |
|
| 8 |
Test on tasks you think are barely possible |
Pushing boundaries reveals the techniques that matter and the limits worth coding around. |
Pick a hard case first, once it works, the easy ones will follow. |
|
| 9 |
Use models to write (and refine) prompts |
High-bandwidth, meta-prompting speeds you up. |
Ask Claude to interview you about the task, then turn that Q&A into the prompt skeleton. |
|
| 10 |
Keep prompts tidy but don't obsess over typos |
Grammar helps humans read; models cope fine. Spend effort on conceptual clarity. |
Final pass: spell-check once, but prioritize meaning over polish. |
|
| 11 |
Tell it what not to do, be specific about what to preserve during edits |
Negative constraints prevent unwanted changes and preserve important elements during modifications. |
Add explicit "Don't change X" or "Preserve Y formatting" instructions when editing existing content. |
|