Learn

Productivity · 50 items · 3 min read

Easier AI Leverage

For one person working with AI models and agents as a small team. Format: easier > harder. AI multiplies whatever you give it. Clear intent gets multiplied, and so does confusion.

Sections
  1. Formulas
  2. What to delegate (1–12)
  3. Asking well (13–25)
  4. Reviewing & trusting (26–37)
  5. Building a system (38–50)

Formulas#

  • AI output quality ≈ context × clarity of the ask × quality of your review. The model is rarely the bottleneck. The brief usually is.
  • Leverage = tasks delegated × success rate − time spent fixing. A 90% agent on the wrong task still loses.
  • Delegate when describing the task + reviewing the result < doing it yourself. That gap gets bigger every model generation.
  • Your role moves from doer to editor to director. Taste and judgment become the job.

What to delegate (1–12)#

  1. Tasks you've done 3+ times and can describe > tasks you've never done yourself.
  2. Drafts, research and first passes > final decisions.
  3. Checkable output (code with tests, data with sums) > output you can't verify.
  4. Boring, repeated work (formatting, migrations, reports) > the creative core you enjoy.
  5. Parallel exploration (5 options at once) > one careful guess.
  6. Reading large piles (docs, reviews, logs) and summarizing them > reading everything yourself.
  7. Code in a stack you know > code you can't review.
  8. Low-risk tasks first, then higher-risk ones > handing over payments on day one.
  9. Customer research from real sources (reviews, forums) > the model's imagination.
  10. Internal tools for yourself > public products shipped without review.
  11. Translation and localization with a native check > no check at all.
  12. Keeping judgment, taste and final say > delegating those too.

Asking well (13–25)#

  1. Rich context (goal, audience, constraints, examples) > a one-line prompt.
  2. Showing an example of "good" > describing it in adjectives.
  3. Asking for a plan first, then execution > jumping straight to output.
  4. Clear definition of done > "make it better."
  5. Project files with standing instructions (CLAUDE.md, style guides) > repeating yourself every chat.
  6. Breaking big tasks into steps > one giant request.
  7. Asking it to question your assumptions > asking it to agree.
  8. Telling it what to avoid > hoping it guesses.
  9. Giving it tools and real data (files, APIs, search) > asking it to recall facts.
  10. Short feedback loops (run, check, adjust) > long unattended runs on unclear tasks.
  11. Asking "what would you need to know to do this well?" > guessing the context it needs.
  12. Saving prompts and workflows that work > rewriting them from scratch.
  13. Choosing the right model per task (fast for easy, strong for hard) > one model for everything.

Reviewing & trusting (26–37)#

  1. Reviewing the output like a senior editor > copy-pasting it.
  2. Tests, checklists and evals > eyeballing.
  3. Checking facts, numbers and citations > trusting confident prose.
  4. Keeping a human in the loop for anything public or irreversible > full autopilot.
  5. Version control and backups > letting agents edit without history.
  6. Limited permissions and API keys > giving agents access to everything.
  7. Reading the diff > reading only the summary.
  8. Small, frequent merges > one huge change you can't review.
  9. Learning where the model is weak > assuming it's good at everything.
  10. Noticing when fixing takes longer than doing > forcing delegation.
  11. Your own voice in public writing > generic AI prose. Readers can tell.
  12. Keeping sensitive data out of tools that shouldn't have it > pasting everything everywhere.

Building a system (38–50)#

  1. Documenting processes as instructions an agent can follow > knowledge living only in your head.
  2. Reusable skills, scripts and templates > one-off chats.
  3. Scheduled agents for routine work (reports, monitoring, digests) > remembering to do it.
  4. One source of truth (repo, docs, notes) > context scattered across chats.
  5. Memory and notes the agent can read > re-explaining your business each time.
  6. Measuring hours saved and quality > vibes.
  7. Upgrading workflows when new models ship > locking in last year's limits.
  8. Spending saved time on customers and thinking > filling it with more busywork.
  9. Learning the fundamentals yourself > being unable to judge the output.
  10. Several agents on independent tasks > one agent doing everything in sequence.
  11. Owning your data and workflows > being locked into one tool.
  12. Staying curious and experimenting weekly > fixed habits from 2024.
  13. Using AI to amplify what makes you you > replacing it. Your taste, relationships and judgment are the edge.

If you keep only 5: #1 (tasks you've done 3+ times), #3 (checkable output), #13 (rich context), #17 (standing instructions), #26 (review like a senior editor).