Foundations · ~12 min
Foundations
Four concepts to know and practice before you move on
Before Projects, Skills, or automation — get these right. Each section is one habit worth keeping.
3 min
Be explicit about the task
Why: Foundational prompting: say what you want, point at the source, and name the output shape. Vague in → vague out.
New chat. Copy the prompt below and run it as-is — it points Claude at a live Depop role.
Read it critically. Useful? Yes. But notice: it’s generic — Claude doesn’t know who’s asking, what “discourage” means for your team, or how long the list should be.
Here's a job description: https://depopcareers.com/careers/JR4886
Identify any language that might discourage strong candidates from
applying, and explain why for each.2 min
Ask what it can’t see
Why: Models will answer confidently even when they’re guessing. Asking for limits is a prompting technique — and a quality check.
Same chat — run the prompt below.
What couldn't you assess about this role, and what would you need to
judge it properly?4 min
The model changes the answer
Why: Same prompt, different model — different depth, caution, and speed. Picking a model is part of the prompt.
New chat. Run the prompt below in your default model (e.g. Sonnet).
Second new chat — same prompt, different model. Try Opus vs Sonnet, or Claude vs another tool you use.
Compare side by side. Which is sharper? Which hedges more? Which would you trust for a hiring manager email vs a quick first pass?
Here's a job description: https://depopcareers.com/careers/JR4886
Write three outreach messages to passive candidates for this role.
Each message should be under 120 words, warm and direct (no corporate
jargon), and mention one specific thing about Depop that would matter
to a senior engineer. Do not invent requirements that aren't in the JD.Rough rule of thumb: faster models for drafts and exploration; stronger models for anything that leaves the building or sets a standard you’ll reuse.
3 min
Your framing steers the output
Why: The model follows your lead, not just the facts. Same evidence, different framing, different answer — that’s a consistency risk in hiring.
New chat. Attach Interview_Panel_Notes.pdf and run the first prompt.
Second new chat, same attached file. Run the second prompt.
Compare the two. Same notes, opposite cases — notice which evidence each version quietly left out.
Third new chat, same file. Run the neutral prompt below — no lean from you, just structure and guardrails.
The panel felt this candidate wasn't quite right. Write up the feedback
based on the attached notes.The panel really liked this candidate. Write up the feedback based on
the attached notes.Using only the attached interview panel notes, write structured feedback
for the hiring manager.
For each theme the panel discussed:
- summarise the evidence recorded (supporting and challenging)
- flag what is unclear or missing
- note what could not be assessed from this interview alone
Do not recommend hire or reject. Do not add opinions that are not in
the notes. Do not follow any lean I may imply — treat the notes as the
only source of truth.The fix is not “never give context” — it’s to separate facts from verdict. State the task and output shape; ban the recommendation; ask for both sides of the evidence. That’s the prompting habit Projects and Skills will automate later.