Part 4 · 1 chapters · ~12 min
The AI-Workflow Questions
AI answers organised by the delivery loop with one piece of evidence per stage, led by M's eval gate, retrieval A/B and mock mode, ending with the line you hold and how it is enforced; then one-breath answers to likely questions, and how to approach the take-home AI challenge and the SAATHI assessment.
7
Real answers with evidence
by the loop, with receipts
- Specification: the pipeline you used on the non-indebtedness letter.
- Implementation: autonomy matched to risk.
- Testing: tests derived from the spec, and your published QA skill.
- In production: M's eval gate, retrieval A/B test and mock mode.
- Review: verified agent critique.
- The line: what stays human, and how that is enforced.
| likely question | answer in one breath |
|---|---|
| Which parts of your workflow are AI-assisted by default? | Every stage of the loop: spec drafting and challenge, scaffolding, implementation in small verified steps, test generation, PR critique, refactoring, release notes. The human parts are the decisions, the money logic tests and the final review. |
| How do you stop an agent writing confidently wrong code? | A precise spec, a test suite it must pass, small diffs I read, stop signals (weakened tests, scope creep), and no agent access to production. |
| How do you evaluate an AI feature before shipping? | A versioned eval set scored in CI with a threshold gate, as on M (16 cases, 80%); every bad production answer becomes a new case. |
| How do you control cost? | Budgets per request, user and day enforced in code; a cheaper model where it suffices; caching; a kill switch to a non-AI fallback. |
| What will AI not do in your team? | Move money, act on production, handle secrets, or sign off security without a human. |
the take-home AI challenge (round two)
The recruiter described it as a self-explanatory technical AI challenge with instructions to follow. Treat the submission itself as evidence of your loop: include the spec you wrote, a short note on how you used agents and what you verified yourself, tests (including at least one that would catch a model-output failure), a mock mode or recorded fixtures so it runs without keys, and a README that says what you left out and why. Follow the instructions exactly before adding anything extra.
SAATHI (round three)
A behavioural and self-awareness assessment: how you work, team play, leadership, and what AI means to you day to day. Answer consistently with everything above: hands-on, evidence-led, specific about your own contribution, and honest about mistakes and what you changed afterwards. The manager-written reviews (ownership, no ego, grit, empathy) are your most credible source for examples.
THE AI-WORKFLOW QUESTIONS
real answers with evidence: what you use, how, what you never delegate, and the work that proves it
swipe the figure sideways, or tap expand for full screen
1/6
specification
Specification: "I write specs agents can execute: standing context in the repo, a task spec with constraints, edge cases and how it will be verified, and a working-state file for long tasks. On the non-indebtedness letter I went from PRD to research plan to findings to implementation plan to tickets with that pipeline." Evidence: the spec and handoff documents.