AI Coding Guide

DataHack Summit · 21 Aug 2025 · Bangalore
Anand S · LLM Psychologist · Straive
Slides · Transcript

Jagged edge. Don’t promise. Parallel-prototype, then share outcomes

Why: AI capability is uneven and unpredictable across tasks (“the jagged frontier”).

Don't commit to your manager/client. Build it. Then share it.
Fixing is 10X effort.

Or, commit to broad improvement, not specific apps.

Do next: run k variants, then talk.

Specs/tests are the product. Code is a build artifact

Why: AI makes typing cheap; the bottleneck becomes unambiguous intent. Structured instructions, diffs, and schemas consistently reduce error and review time.

Do next (CLI-first): keep PROMPTS.md/SPEC.md; use schema-constrained outputs (JSON Schema/Typed objects), and PR diffs only.

Optimize the validation loop, not the coding loop

Why: Copilot-style tools boost raw speed (55% faster on a benchmark task), shifting value to verification. Evals/property tests and "fail-first" tests pay compounding returns.

Do next: add property tests (Hypothesis/fast-check), hidden tests, metamorphic checks; gate merges on evals.

Make repos "LLM-friendly"

Why: Models perform better with consistent structure, fast tests, and machine-readable context (e.g., repo "manuals" for LLMs).

Do next: add llms.txt (or Context7), keep small files, consistent names, fast lint/test targets; ship sample I/O fixtures.

Prefer typed languages & DSLs

Why: Structured outputs (schemas/grammars) and typed APIs minimize ambiguity and regression surface; typed stacks (TS/Rust/Go or Python+typing/Pydantic) tend to vibe-code more reliably.

Do next: define JSON Schemas/OpenAPI for tasks, emit DSLs (SQL, GraphQL, Vega-Lite) instead of free-form code where possible.

Re-design process around short feedback loops

Why: With AI in the loop, the classic "tighten the loop" wisdom matters more-fast, frequent checks beat elaborate prompts.

Do next: pre-commit hooks to run lint && test && typecheck; CI must be <2 minutes; default to many small PRs.

Always build evals-in-the-loop

Why: Evolutionary/iterative systems that mutate code and auto-evaluate consistently outperform single-shot prompting.

Do next: spin N variants per change; auto-benchmark; keep a league table; merge winners only.

Expect uneven gains: seniors benefit differently than novices

Why: Experienced devs get sizable productivity boosts, but novices can develop an illusion of competence without robust verification. Plan training, not just tooling.

Do next: teach code review and failure-mode literacy; require short "how I verified" notes in PRs.

AI Coding Guide

DataHack Summit · 21 Aug 2025 · Bangalore
Anand S · LLM Psychologist · Straive
Slides · Transcript