Self-paced learning
Token efficiency
Measure token use and design lean, reliable AI workflows.
Your lessons
- →
Measure a baseline before you optimise
Record input, output, and cached tokens plus tool calls and retries for real work, and keep context-window occupancy distinct from billed usage.
- →
Write a compact brief and bound the output
Shorten the task brief and cap output without dropping acceptance criteria, safety rules, or the checks that prove the work.
- →
Send relevant context, not the whole repo
Choose targeted files, search hits, and short handoffs instead of resending the repository or full chat history by habit, and know when a summary is too lossy.
- →
Reuse stable prefixes; verify cache hits
Use prompt caching as a provider-specific billing and latency tool, keep prefixes stable, and confirm cache reads in usage — without assuming a universal discount or a smaller request.
- →
Bound loops, tools, models, and batch jobs
Cap retries and expensive tool calls, pick a model that matches the task, and use asynchronous batch APIs only for independent, latency-tolerant work — without removing checks.
- →
Evaluate before/after; keep the cheaper workflow only if quality holds
Run the same tasks on the baseline and optimized workflows, report every measured category plus quality, and accept the cheaper path only when the pass criteria still pass.
Your project
Measure token use and design lean, reliable AI workflows.
Optional saved submission: share an HTTPS URL to a sanitised token-usage report and explain your comparison. Include task criteria, observed or explicitly simulated categories, tool calls, and quality results. No API keys, private code, or raw personal logs. Learning and practice are free without a purchase; signing in is only needed to save. Saved evidence is self-reported, not verified or certified.
Check your work
- Measured baseline
At least three tasks have baseline rows marked observed or supplied simulated data, correct accounting, and unknown fields retained. Simulation is not presented as measured savings.
- Criteria and checks retained
Compact briefs and thin context still include acceptance criteria, safety rules, and the named quality check. Tests were not deleted to save tokens.
- Token reduction vs discount vs latency vs quality
The report does not treat cache hits or batch pricing as fewer request tokens or as a guaranteed percentage. Occupancy and billed categories are separate columns.
- Bounded loops and justified model/batch
A step or attempt cap and stop rules exist. Batch is used only for independent, latency-tolerant jobs or is explicitly skipped. Model choice is tied to task difficulty, not a frozen price list.
- Cheaper only if quality passes
Each optimized task is kept only when its pass criteria still pass. Failures keep the baseline and say so.
- Self-reported public proof
Public https URL, secrets redacted, and the learner treats the save as pending review — not certified, not instructor-approved, and not a claim of credits, accreditation, jobs, or income.