A 90-day development plan
Four phases, real deliverables and a criterion for advancing. Track your own progress.
14.1Your roadmap
Four phases on a real project
An adaptable proposal, not a promise of proficiency in three months. Use a real project with authorized access and controlled impact. Phases can overlap — the criterion for advancing is the evidence obtained. Set the day you are on and tick deliverables as you go; everything is saved only in this browser.
Days 1–15
Establish a complete flow and a baseline
What to do
- 1Choose one main tool; learn to operate context, execution, resumption and limits.
- 2Run a few real tasks and record where you intervened, the time consumed and which checks supported the delivery.
- 3Talk with a user or product owner to choose an improvement with recognizable value.
- 4Reproduce a failure and complete a small delivery with AI, including testing and observing the result.
Deliverables · 0/4 done
Criterion for advancing
You can explain the goal, demonstrate the result and distinguish the agent’s contribution from your interventions.
Just starting out?
Reduce the scope of the system and increase the room for explaining and experimenting.
Already senior?
Choose a bottleneck that affects other people.
High-consequence area?
Concentrate the work in isolated environments until the relevant controls are demonstrated.
Key takeaways
- 1
Start with one tool, real tasks and a baseline of where you intervene.
- 2
Make work delegable: repository map, task contracts, shared procedures.
- 3
Increase autonomy only with evidence: a small eval suite and an autonomy policy.
- 4
Finish by improving the team’s system — the benefit must survive your absence.
Check your understanding
Question 1 of 2
What is the criterion for advancing after days 16–30?
Read the full section of the original reportThe complete, unabridged text — with every source and nuance.
This roadmap is an adaptable proposal, not a promise of proficiency in three months. Use a real project, with authorized access and controlled impact. The phases can overlap; the criterion for advancing is the evidence obtained.
Days 1–15: establish a complete flow and a baseline
Choose one main tool and learn to operate context, execution, resumption, and limits. Select a few real tasks and record where you intervened, how much time was consumed, and which checks supported the delivery.
Talk with a user or product owner to choose an improvement with recognizable value. Reproduce a failure or a limitation and complete a small delivery with AI, including testing and observing the result.
Phase deliverables: task log; first completed change; list of the main limitations of the flow; one personal technical gap identified.
Criterion for advancing: you can explain the goal, demonstrate the result, and distinguish the agent's contribution from your interventions.
Days 16–30: make the work more delegable
Improve the repository map and the setup and validation commands. Use task contracts with invariants and completion criteria. Turn a repeated instruction into a shared resource, removing outdated information that gets in the way.
Hold a deliberate session on the chosen gap. Complete a variant that requires applying the understanding, and record what changed in your judgment.
Phase deliverables: short project orientation; tested task template; reusable procedure; comprehension exercise.
Criterion for advancing: another person, or a new agent session, can repeat the process without depending on verbal guidance from the author.
Days 31–60: increase autonomy with evidence
Build a small collection of tasks to evaluate the flow. Include previous errors, edge cases, and tasks with different kinds of difficulty. Compare two configurations when that answers a real question, keeping criteria fixed.
If there is a limitation that parallelism could solve, try it on a decomposable task, defining interfaces, isolation, and integration responsibility. Check permissions, accessible data, and recovery procedures. Automate one class of repetitive checks.
Phase deliverables: initial evaluation suite; automated repetitive checks; simple autonomy policy by class of change. Add a comparative log and an integrated parallel task if those experiments were useful for the chosen problem.
Criterion for advancing: there is a demonstrable reduction in repetitive intervention, accompanied by evidence of quality and control of consequences.
Days 61–90: improve the team's system
Choose an observed bottleneck: review queue, environment startup, CI, fragmented context, or manual validation. Implement an improvement and track its effect. If it depends on another team, write a proposal with the problem, impact, evidence, and required investment.
Share the procedure and train at least one other user of the flow. Reassess the value delivered from a product perspective. Remove automations and instructions that have not proven useful.
Phase deliverables: adopted flow improvement; contextualized before/after comparison; usage documentation; next development priority.
Criterion for advancing: the benefit survives your absence and shows up in other people's work or in a better result for the user.
How to adapt it without turning the roadmap into bureaucracy
If you are just starting out, reduce the scope of the system and increase the room for explaining and experimenting. If you are already senior, choose a bottleneck that affects other people. In higher-consequence areas, concentrate the work in isolated environments until the relevant controls have been demonstrated.
The plan can be repeated in cycles. In each cycle, eliminate some repetitive execution and expand one capacity for judgment. That combination keeps the transition connected to the work.