Edison SinaniAI implementation architectBook a call

AI Team Playbook / Rule 12 of 12

Protect the people who are still learning

Developers who are new to the stack or to the craft do some work by hand first, and must be able to explain every line of a change without the agent before it merges.

September 16, 2026 · About 2 minutes

Contents

The playbook

  1. Overview

Before code

  1. 1Write the constitution first
  2. 2One instruction file, under 200 lines
  3. 3Prose is advice, hooks are enforcement
  4. 4Three gates before any code
  5. 5A walkable demo before any screen

Building

  1. 6One task, one worktree, one pull request
  2. 7Every contract has a file and two tests

Reviewing

  1. 8AI review first, and filtered
  2. 9Human review by risk tier
  3. 10Done means the evidence is in the pull request

Over time

  1. 11Measure five numbers from day one
  2. 12Protect the people who are still learning

Appendix

  1. AInstall checklist
  2. BOther tools, limits and sources
  3. CTightening the process
  4. DRun the agent with no entitlements

Get the kit

Who owns it
The lead decides who the rules apply to and for how long, and writes it down so it is policy and not a personal comment.
In the kit
docs/onboarding-and-learning.mddocs/roles-and-rituals.md
Enforced by
Explain-before-merge is checked by the reviewer. The rest is policy, and the milestone retrospective asks whether it held.

Why

Agents remove the struggle that learning comes from. In a randomized trial, 52 mostly junior engineers learned an unfamiliar library. The group using AI scored 50 percent on comprehension. The group without it scored 67 percent. The largest gap was in debugging [7]. The long term version of this now has a name, comprehension debt: the gap between how much code a system holds and how much of it any person understands [16].

Comprehension with and without AIQuiz score after learning an unfamiliar library. 52 mostly junior engineers, randomized. Anthropic, 2026. 7
  1. 67%Without AIComprehension score for the group that learned the library by hand.
  2. 50%With AI17 points lower. The largest gap was in debugging questions.

How

Five rules for juniors and for anyone new to the stack.

  • Manual first. For the first two weeks in a new area, implement one task there without an agent, then review the agent’s version of the next task against your own. Comparison is the fastest teacher.
  • Explain before merge. On Standard and High tasks, the author must be able to explain every line to the reviewer without the agent. If they cannot, the pull request waits. This is the one rule with the biggest effect on understanding, and it costs nothing when the author already understands the change.
  • Debug by hand first. Form a guess and check it. Fifteen minutes, then the agent.
  • Read the agent’s plan, not just its diff. Disagreeing with the plan is where judgment forms.
  • One review a week without the first pass. Review one pull request before reading the automated findings, then compare.

Onboarding a developer, day one, before any agent session. Read the constitution, the system design and one implemented spec. Run the system from the README, and if a step fails, fixing the README is the first pull request. Read the instruction file, because the developer should know at least what the agent knows. Walk one merged pull request with the person who reviewed it. Day two is a Low task with an agent, the lead reviewing. The first week ends with one spec written and approved.

Bringing a whole team on, in four weeks. Week 1, the lead installs the kit and runs one feature alone, fixing the templates. Week 2, one developer joins, on a Low task and then a Standard one. Weeks 3 and 4, the rest of the team, one feature each. The weekly checkpoint starts, and by the end of week 4 the measurement file has three weeks of baseline. At the second milestone, review the numbers, decide on automatic approval for Low, and retire any rule nobody needed.

Adoption spreads by watching, not by mandate. A study of Microsoft’s rollout across tens of thousands of engineers found that the strongest predictor of trying an agent was whether the people around you already had, not a mandate or a training course [17]. Pair on the first task. Show the worktree and the hook firing. That does more than a document.

There is no “AI champion” role, on purpose. No company with a published, measured setup has one. On a small team that job is the lead’s ownership of the configuration.

References

  1. 7Anthropic, "How AI assistance impacts the formation of coding skills", February 2026. Randomized controlled trial, 52 engineers. www.anthropic.com/research/AI-assistance-coding-skills
  2. 16Farhan Thawar, VP of Engineering at Shopify, on "comprehension debt", as reported in Addy Osmani, "Comprehension Debt: The Hidden Cost of AI-Generated Code", O'Reilly Radar, 2026. www.oreilly.com/radar/comprehension-debt-the-hidden-cost-of-ai-generated-code/
  3. 17"Adoption and Impact of Command-Line AI Coding Agents: A Study of Microsoft's Early 2026 Rollout of Claude Code and GitHub Copilot CLI", arXiv 2607.01418, July 2026. arxiv.org/abs/2607.01418