[Image placeholder — hero] Replace with
from-chatgpt-to-power-user-hero.webpwhen ready.Three-rung ladder (Prompt → Context → Harness) with chibi Pawfessor Hu at the base, looking up.
TL;DR
- ChatGPT is autocomplete. It fills blanks with plausible guesses. Show it the thing.
- Three rungs: (1) say the task clearly → (2) attach the source material → (3) package it so Monday doesn’t start from zero.
- Bad output? Usually missing context, not magic words.
- This week: pick one chore below. Run it daily. Ignore new frameworks (that’s the Squirrel).
- You still hit send. Verify numbers, names, and anything customer-facing.
| Rung | One line | Fix when output is… |
|---|---|---|
| Prompt | Context + Objective + Response | Vague or wrong shape |
| Context | Paste/upload the actual source | Generic or made-up |
| Harness | Custom GPT, checklist, n8n | Tired of re-explaining every week |
The whole game: say what you want → show what you mean → automate what repeats → check before you ship.
The ladder (60-second version)
| Term | Plain meaning | Example |
|---|---|---|
| Prompt engineering | Clear instructions for one task | ”Turn these bullets into a 200-word exec summary, formal tone.” |
| Context engineering | The background it can’t guess | Upload the PDF, paste the email thread, add an example you liked. |
| Harness engineering | Reusable setup for repeat tasks | Custom GPT + review checklist + n8n weekly trigger. |
Most tedious work is five chores: transform, draft, extract, compare, explain. You’re offloading the grunt layer — not your judgment.
Rung 1: Prompt (what you type)
Not magic words. A complete ask.
Fill three slots every time:
| Slot | Ask | Example |
|---|---|---|
| Context | What can’t it guess? | ”PM at B2B SaaS. Renewal email to churn-risk account.” |
| Objective | What output? | “120-word email proposing a 15-min check-in.” |
| Response | What shape? | ”Bullets first, then draft. No subject line.” |
Frameworks like COSTAR are just checklists in wizard robes — longer take here.
Copy-paste starters
Email:
Context: I'm [role] at [company]. Email below.
Objective: Draft reply [agree / push back / deadline / decline].
Response: <150 words, warm-professional, bullet their ask at end.
[paste thread]Meeting notes:
Context: Raw notes from [meeting].
Objective: Decisions, action items (owner + deadline), open questions.
Response: Table for actions; bullets for rest. Flag ambiguity.
[paste notes]Doc summary:
Context: [Report/policy] for leadership.
Objective: Exec summary — they won't read the full doc.
Response: 5 bullets (situation, rec, risks, ask, timeline). Max 200 words.
[attach or paste][Image placeholder]
from-chatgpt-to-power-user-costar-scroll.webp— COSTAR scroll → three-checkbox sticky note.
Rung 2: Context (what you attach)
“Write a better email” with no thread = generic slop.
Before blaming the model: what didn’t you show it?
Bring: source doc/thread, constraints, a good example, a bad example (“not this tone”).
| ChatGPT feature | Does what |
|---|---|
| File upload | Ground in the real doc |
| Projects | Pinned files + history for one workstream |
| Memory | Standing prefs — not task files |
| Custom instructions | Baseline on every chat |
| Follow-ups | ”Shorter. Page 3. That number.” |
3 checks before anything important: (1) source attached? (2) audience named? (3) example shown? Any no → fix context first.
[Image placeholder]
from-chatgpt-to-power-user-job-pie.webp— tiny “judgment” vs huge “grunt work” bar.
Rung 3: Harness (what you reuse)
Stop repacking every Monday.
Harness = templates + guardrails + automation around the model.
- Custom GPT — one recurring job, packaged
- Prompt library — 5–10 snippets, not 50
- Review checklist — names, numbers, tone, “would I send this?”
- n8n — trigger → prompt → output, weekly
Build when: same task 3+/week, same input/output shape. Don’t build a Custom GPT until you’ve done the task manually 5× and know what context matters.
Status email harness (example): Project with template → Custom GPT with section rules → 4-item checklist → send.
[Image placeholder]
from-chatgpt-to-power-user-go-bag.webp— go-bag: Custom GPT, checklist, n8n.
[Image placeholder]
from-chatgpt-to-power-user-tutorial-vs-power.webp— Tutorial mode (ask → copy) vs Power mode (chore → workflow → review → ship).
Pick ONE chore this week
| Chore | When | Context to attach | Harness later | Saves |
|---|---|---|---|---|
| Inbox | Reply takes 5+ min | Full thread, role, tone example | Custom GPT + sign-off rules | 10–20 min |
| Meetings | Notes exist, no artifact | Raw notes, attendees, doc | Project per standing meeting | 15–30 min |
| Doc recycle | Same doc → many formats | Source doc, audience per format | Usually later | 30–60 min |
| Research | Need a table, not a PhD | Constraints, trusted links | Rare — verify everything | Framing only |
| Spreadsheets | Formulas, cleanup, regex | Upload CSV or real rows | Saved prompts for exports | High if you’d tab-switch |
Use the prompt templates from Rung 1. Review: names, numbers, invented facts. Models lie smoothly.
Features cheat sheet
UI labels vary by plan/region.
| Feature | Layer | Use for | Skip when |
|---|---|---|---|
| Custom instructions | Context | Default tone/role | One-off creative tone |
| Projects | Context | Ongoing work + pinned files | Throwaway question |
| File upload | Context | PDFs, sheets, long docs | Can’t paste (sensitive) |
| Memory | Context | ”I like bullets” | Confidential context |
| Canvas / Voice / ADA | Prompt+context | Iterate docs, talk, CSV | Quick one-liner |
| Custom GPTs | Harness | Same task weekly | Task changes every time |
| API / n8n | Harness | Scheduled automation | Still learning rungs 1–2 |
Don’t embarrass yourself
- No secrets, credentials, PHI, unreleased financials
- Numbers & dates: verify
- Customer-facing: human read
- More context pasted = more privacy care, not less
Why the chat box is a trap — worth a read when you have five minutes.
Beyond ChatGPT (later)
Master prompt + context in ChatGPT first. Add tools when re-explaining hurts.
| Tool | For | Notes |
|---|---|---|
| Gemini | Google Workspace, long docs | Same three rungs |
| Claude | Long docs, careful reads | Same three rungs |
| Cursor / Codex | Code in-repo | Repo = context — field note |
| n8n | Weekly automation | Harness outside the chat UI |
30 days
| Week | Do this |
|---|---|
| 1 | Context / Objective / Response on every ask |
| 2 | Projects + uploads — never ask naked |
| 3 | One Custom GPT + checklist for one repeat task |
| 4 | Kill what didn’t save time; keep one harness |
Write down what worked. That’s Field Notes. Week 4 with zero notes = Squirrel amnesia.
Build the thing. Check the output. Keep moving. 🐾