![](/blog-assets/ai-collaboration-playbook.webp) Have you ever been in this situation? - You ask AI to fix a bug, and it “helpfully” changes unrelated code—then you end up reverting by hand - You write a mountain of prompts, and AI still can’t find the right files; it just “guesses” - Halfway through a long chat, AI forgets the constraints and quality drops off a cliff If you treat AI as “a faster search box,” these issues stay hidden. But the moment you treat AI as “a collaborator,” they hit delivery quality head-on. This post walks through an **actionable engineering workflow** that turns AI collaboration from “prompt witchcraft” into a **repeatable process**. After adopting it in PrivyDrop, shipping features and fixing bugs got noticeably faster and steadier—not by taking bigger risks, but by doing less rework. By the end, you’ll have a minimal structure you can copy into any repo: - `AGENTS.md`: repo-level hard constraints (red lines, defaults, definition of done) - `docs/ai-playbook/index.md`: a one-page, high-signal entry point - `docs/ai-playbook/code-map.md`: a code map (where to change) - `docs/ai-playbook/flows.md`: key flows (how it runs) - `docs/ai-playbook/collab-rules.md`: collaboration rules + change-plan template (how we work) All examples come from the open-source project PrivyDrop—feel free to follow along: [**https://github.com/david-bai00/PrivyDrop**](https://github.com/david-bai00/PrivyDrop) Also, OpenAI recently published a hands-on write-up with a very similar mindset: [**https://openai.com/index/shipping-sora-for-android-with-codex/**](https://openai.com/index/shipping-sora-for-android-with-codex/) --- ## Step 0: Define boundaries and “done” (don’t start with prompts) This step does only one thing: make “what done looks like” explicit. Otherwise AI will do everything it can to make the code “run,” not to make it “run in the maintainable way your team expects.” Start with three minimal constraints: 1. **Boundary**: what must never happen (privacy/architecture red lines, protocol compatibility, guardrails on key parameters) 2. **Scope**: one change, one goal—no “while I’m here” 3. **Done**: build/tests/manual regression checklist must be written down Compress them into a single sentence and put it at the top of every request: ```text Single goal, plan first; never cross privacy/architecture red lines; done means it builds and includes a regression checklist. ``` --- ## Common anti-patterns (avoid these traps) Before we begin, here are the three most common wrong approaches: 1. **Tell AI to “change code” without a plan** - Result: 10 files changed, then you realize the direction is wrong—rollback becomes painful - Better: require a change plan first; implement only after approval 2. **Dump every document into the prompt** - Result: context overload; AI can’t see the signal (it can’t even locate entry points) - Better: provide a high-signal index + code map 3. **Let AI “optimize things on the side”** - Result: one PR mixes multiple goals; review cost doubles; bugs get harder to roll back - Better: single-scope changes that are easy to revert --- ## Step 1: Write `AGENTS.md` (repo-level hard constraints, consistently reused) Think of `AGENTS.md` as a machine-readable version of your team’s “defaults” and “red lines.” Its job isn’t to explain theory—it’s to be **reapplied in every session**. In PrivyDrop, five lines are enough to cover the core of engineering collaboration constraints: - Plan first: `AGENTS.en.md:7` - One change, one purpose: `AGENTS.en.md:8` - Privacy & architecture red line: `AGENTS.en.md:9` - Docs must stay in sync: `AGENTS.en.md:12` - Verification required: `AGENTS.en.md:13` File: [**AGENTS.en.md**](https://github.com/david-bai00/PrivyDrop/blob/main/AGENTS.en.md) If you want a minimal starter, here’s a good shape (keep it short, strict, and executable): ```md # AGENTS — Repo Rules First Principles - Plan-first: Propose a change plan and get approval before writing code - Single-scope: One PR solves one goal; avoid “while I’m here” fixes - Redlines: Never cross privacy/architecture/protocol/key-parameter guardrails - Docs-sync: If flows/entry points/interfaces change, update the playbook docs - Validation: Must include build/tests and key manual regression checklist ``` ### Multi-language support If your repo needs multilingual collaboration, a pragmatic pattern is: - Keep `AGENTS.en.md` as the canonical version - Add localized variants as needed (e.g. `AGENTS..md`) - After cloning, each contributor can create a language-specific symlink locally: ```bash # English users ln -s AGENTS.en.md AGENTS.md ``` - Add `AGENTS.md` to `.gitignore` to avoid symlink conflicts > **Key insight** > The secret of reliable AI collaboration isn’t “better prompts.” It’s “making constraints part of the repo.” `AGENTS.md` makes rules reusable, and the AI Playbook makes context durable. --- ## Step 2: Write `docs/ai-playbook/index.md` (a high-signal entry point) One of the most common reasons AI goes off track: **it can’t find your real entry points**. Then it “guesses” where to change things. Your index page should do two things: - Readable in 30 seconds: only “project snapshot + link index” - One-click navigation: send readers/AI to code-map / flows / collab-rules Reference implementation: [**docs/ai-playbook/index.md**](https://github.com/david-bai00/PrivyDrop/blob/main/docs/ai-playbook/index.md) Minimal template (copy-pasteable): ```md # AI Playbook — Index ## Project Snapshot - Stack: Next.js / Node / ... - Red lines: ... ## Document Index - Code map: docs/ai-playbook/code-map.md - Key flows: docs/ai-playbook/flows.md - Collaboration rules: docs/ai-playbook/collab-rules.md ``` --- ## Step 3: Write `code-map.md` (where to change: entry points + one-line responsibilities) The code map is for “fast navigation,” not “teaching you how to implement.” The writing rules are simple: - List only key directories and key entry files - One sentence per entry file: what it’s responsible for - When a new request arrives: hit 3–8 candidate files in the code map before deep reading Reference implementation: [**docs/ai-playbook/code-map.md**](https://github.com/david-bai00/PrivyDrop/blob/main/docs/ai-playbook/code-map.md) Optional: add a “common requests → entry points” cheat sheet to reduce search cost further: ```md Common request routing - New page / SEO: frontend/app/\*\*/page.tsx + metadata.ts - i18n copy: frontend/constants/messages/\* - Blog: frontend/content/blog/\* + frontend/lib/blog.ts ``` How to generate (and iterate) a code map: - First version: ask AI to summarize “directories + key entry points” for navigation—not completeness. - Iteration: treat it as a living doc. Update it incrementally per PR/commit (file list), instead of rewriting it from scratch. --- ## Step 4: Write `flows.md` (how it runs: key sequences + debug points + micro-plan template) If code-map answers “where to change,” flows answers “how it runs.” This is hugely valuable for AI: - If you write down the sequence and invariants, AI stops “patching by vibes” - You can compress past pitfalls into a reusable debug checklist Reference implementation: [**docs/ai-playbook/flows.md**](https://github.com/david-bai00/PrivyDrop/blob/main/docs/ai-playbook/flows.md) (with split deep dives) At minimum, include: 1. **Key flow / sequence** (Mermaid if useful) 2. **Debug checklist** (the most important logs/states) 3. **Micro-plan template** (force plan-first before coding) How to generate (and iterate) flows: - First version: ask AI to restate the end-to-end flow + key sequence + invariants. You review/correct (especially red lines and invariants), then commit it to docs. - Iteration: keep it up to date incrementally when interfaces/sequences change. --- ## Step 5: Make “plan first” enforceable (a plan = a mini design doc) This is where the speed-up really comes from: move the review earlier—from “reading diff” to “reading the plan.” Put the plan template in `collab-rules.md` and treat it as a hard constraint. PrivyDrop template: [**docs/ai-playbook/collab-rules.md**](https://github.com/david-bai00/PrivyDrop/blob/main/docs/ai-playbook/collab-rules.md) You can reuse this structure directly (goals/scope/approach/risks/acceptance/rollback/validation): ```text Title: Goals - Scope / Files - Approach - Risks & Mitigations - → Acceptance Criteria - Rollback - Docs to Update - docs/ai-playbook/index.md / code-map.md / flows.md / collab-rules.md / others? Validation - Build: next build - Manual: ``` In practice, a steadier workflow is: 1. Read playbook index + code-map + flows (read-only, no code changes) 2. Restate the current state and constraints in your own words (you correct once) 3. Produce a change plan (implementation only after approval) > **Pitfall avoidance** > Correct direction during the plan stage; avoid building and then tearing down. Single-scope changes keep rollback cheap and merging easier. --- ## Step 5.1: Context endurance (checkpoint → new chat) In long tasks, output quality dropping is almost inevitable. Make “endurance” a standard move: when you see guessing, forgotten constraints, or drift—write the state into a file, then continue in a new chat; or run a “compress/summarize” step to shrink context before continuing. Minimal handoff template (put it in `docs/ai-playbook/handoff.md` or a temporary file): ```md # Handoff ## Problem statement (3–5 sentences) ## Confirmed plan (bullets) ## Done / Not done ## Key files and entry points ## Red lines and invariants ## Acceptance & regression checklist ## Next-step checklist ``` The goal isn’t “beautiful docs.” It’s moving context from the chat window into a file that the next session can read reliably. --- ## Step 6: Close the collaboration loop (treat the agent like a new teammate) Once you have the pieces above, collaboration becomes a stable pipeline: 1. Request → constraints (cite `AGENTS.md`) 2. Navigation → entry points (cite `index + code-map`) 3. Alignment → sequences (cite `flows`) 4. Planning → mini design doc (cite `collab-rules` template) 5. Implementation → small, single-scope change (easy rollback) 6. Verification → `next build` + key manual regression points 7. Sync → keep playbook docs up to date The most visible benefit for me: shipping features and fixing bugs is faster and steadier. More importantly, “fast” comes from **less rework**, not more risk: - Correct direction at plan time; avoid throwing work away - Single-scope changes keep rollback cheap and merges easy - Flows turn past pitfalls into a checklist you can reuse If you want a gate, put these two questions into your PR template: - Does this PR include a change plan link/summary? - Did you update `docs/ai-playbook/*` (if entry points/flows/interfaces changed)? ## Prompt examples (ready to copy) Once your AI Playbook exists, you can start a collaboration session with something like this: --- **Role** You are a senior Next.js full-stack engineer with strong product instincts. Your collaboration quality determines whether this repo can iterate sustainably—be thorough and professional. **Task kickoff** Please read `docs/ai-playbook/index.md` to understand the project context, code map, and collaboration rules. The current request is: "xxx". **Working style** Please deeply read relevant docs/code. Think systematically, ask clarifying questions, then propose analysis + a change plan for review. Implement only after approval. --- **Why this works** - **Role**: sets quality bar and responsibility boundaries - **Kickoff**: forces reading the playbook instead of guessing - **Working style**: makes “plan first” part of the workflow --- ## Industry reference: How OpenAI uses Codex to run a sprint In “How we used Codex to build Sora for Android in 28 days,” OpenAI describes a workflow that matches this playbook closely: [**https://openai.com/index/shipping-sora-for-android-with-codex/**](https://openai.com/index/shipping-sora-for-android-with-codex/) Key points to align with: - Treat the agent like a new senior engineer: capable, but needs clear architecture/constraints - Externalize rules: maintaining a strong `AGENTS.md` pays off - Plan before real changes: plans are mini design docs—debug the plan before debugging the code - Context endurance: when you hit context limits, write plans to files for the next session - Multi-session parallelism: it’s closer to “managing a team” than using a single tool Android and web are different, but the method transfers: **improve inputs, and outputs stabilize**. --- ## A minimal directory structure you can copy ```text AGENTS.en.md # Canonical rules (English) AGENTS..md # Optional localized rules AGENTS.md # Symlink (created locally after git clone) docs/ ai-playbook/ index.md # High-signal entry point code-map.md flows.md collab-rules.md ``` If you already have docs scattered everywhere: start with `index.md` to unify entry points; then fill in code-map/flows/templates. --- ## Next steps 1. **Start now**: copy the minimal structure into your repo and begin with `AGENTS.md` 2. **Reference implementation**: visit [**PrivyDrop GitHub**](https://github.com/david-bai00/PrivyDrop) and browse the full AI Playbook 3. **Feedback**: if this helps (or you hit pitfalls), open an issue on GitHub or leave a comment 4. **Star**: if it’s valuable, consider starring PrivyDrop 🌟 --- ## Closing AI-assisted development doesn’t reduce the need for rigor—it increases it. The sustainable speed-up doesn’t come from longer prompts, but from stronger engineering constraints: externalize rules, plan first, codify flows, and make context durable. If you want to go further, this can evolve into a “copyable repo scaffold”: PR templates, issue templates, and a ready-to-use `AGENTS.md` + playbook starter kit.

Tag: Codex

Keep AI on Track: A Practical, Repeatable Collaboration Playbook (with Templates)