01 / Start with the job
What genuinely requires judgment?
Library / Working with AI
Reviewed Oct 2026Efficient AI Use for Product Work
Use the lightest capable tool and evaluate outcomes, not just token cost.
01 / flow
Pick the smallest sufficient approach
Define the jobSeparate deterministic rulesShrink contextChoose capable modelStructure outputVerify outcome
02 / cards
Ten practical habits
02 / Shrink context
Remove stale or irrelevant material first.
03 / Match tool to task
Fast model for routine work; reasoning for ambiguity.
04 / Reuse knowledge
Keep decisions, briefs, evals and workflows durable.
05 / Batch comparisons
Cluster similar tickets, cases or feedback.
06 / Structure outputs
Decision, evidence, unknowns, risks, next step.
07 / Bound agents
Goal, tools, stopping rule and escalation.
08 / Escalate gradually
Signal → source → adjacent evidence → broader search.
09 / Verify with software
AI interprets, code checks, people authorize.
10 / Measure full cost
Latency, retries, cleanup, review and bad decisions.
03 / note
Decision rule
Optimize for quality-adjusted outcome per dollar, not lowest tokens per dollar.
GOING DEEPER
Do less before scaling up
Practical escalation: Start with a known signal and the smallest sufficient context; expand to likely sources, adjacent evidence and only then broader investigation. Use a fast capable model for ordinary tasks and a reasoning model only when ambiguity merits it. Agents need explicit tool scopes, stop conditions and escalation.
Measure total outcome cost: Include tool calls, latency, retries, cleanup, human review and bad-decision risk. Cheap tokens do not imply a cheap successful outcome.
PRACTICAL FRAMEWORK / AI PRODUCT WORK
Cost per successful outcome and spending authority
Do not optimize tokens alone. Track model calls, paid APIs, tool and compute charges, retries, review time, cleanup, support and downstream correction costs.
| Measure or control | Decision rule |
|---|---|
| Cost per successful outcome | Total operating + human cost ÷ verified completed tasks; include failures in total cost. |
| Per-action and total budgets | Hard limits enforced by software, not by a reminder in a prompt. |
| Allowed services and escalation | Approved spend categories, approval above thresholds and transaction logs. |
| Stop condition | Terminate repeated calls, loops or spending when the outcome stops improving. |