Appendix G: Glossary
Status: drafted · Time: 10 min · Audience: everyone Outcome: Look up any acronym, tool, role, or concept used in the curriculum and find a short definition with chapter pointers.
Every acronym, tool name, role, and concept the curriculum uses, with a short definition and a pointer to the chapter that covers it in depth. The glossary is alphabetical. When a term has a richer treatment elsewhere, the entry says where.
The glossary is meant to be useful as a reference, not as a substitute for the chapters. A reader who needs the full context for a term should follow the cross-reference. A reader who just needs to remember what an acronym stands for should be able to find it here in under thirty seconds.
ADR (Architecture Decision Record). A small, immutable per-decision artefact that records the chosen path for an architectural decision. Lighter than an RFC. Originated by Michael Nygard in 2011. See C.3 — The RFC pipeline.
Agent. A program that uses a language model to take actions through tools, with some degree of autonomy across multiple turns. Distinct from a chatbot, which only generates text. See Y.1 — The Tool Atlas and B.4 — The Claude Agent SDK.
API (Application Programming Interface). The contract through which software components communicate. The shape of inputs, outputs, and errors a service exposes to its callers. See 0A.5 — What is an API?.
API Council. Razorpay’s existing body that reviews API designs against the API Design Guide. The AI-specific lens on API Council reviews is covered in B.15. Distinct from the Staff+ Council described in the Council section.
Auto-memory. A persistent file-based memory pattern Anthropic ships with Claude Code, where the agent maintains a small set of memory files that survive across sessions. See B.8 — Memory systems.
Badge. The artefact a builder claims when they complete a belt. Earned through quests and a boss fight, not self-reported. See Appendix L — Certification.
Belt. A level of competence in the playbook curriculum. The four earnable belts are White, Yellow, Green, and Black. The Council sits above the belts but is not itself a belt. See the Master Index.
Black Belt. The fourth belt. AI-native builder. Force multiplier. Authors skill packs and plugins; ships across layers; coaches. See the Black Belt overview.
Blade. Razorpay’s design system. Used in design and frontend chapters. See G.16 — Blade deep dive.
Boss Fight. The capstone artefact required to earn a belt. Larger than a quest. Each belt has one. See the relevant belt README.
Builder. Anyone working through the playbook curriculum. Used inclusively across roles (engineer, designer, PM, ops).
Builder Day. A program-wide event where teams work on AI-assisted projects together. The origin story in the Prologue covers what made the second Builder Day work. See Prologue §0.2.
Charter. The Council’s founding document, ratified by members and revised annually. See the Council charter template and C.2.
CLAUDE.md. A markdown file in a repository that gives Claude Code project-specific context: conventions, do-not-touch areas, common workflows, gotchas. See G.3 and the service-level template.
CLAUDE.local.md. A gitignored personal-overrides file for Claude Code. See G.5 and the template.
Claude Agent SDK. The SDK Anthropic ships for building custom agents. Covered in B.4.
Claude Code. Anthropic’s terminal-native coding agent. The program’s default coding tool. See Appendix A — Tool Atlas.
Claude Enterprise. Razorpay’s org-workspace path for approved hosted chat and knowledge work. It can coexist with separately provisioned Claude Team access and a LiteLLM code route. See Y.8.
Claude Team. A separately provisioned, seat-limited Anthropic subscription path used for selected Razorpay users. An Enterprise workspace or LiteLLM key does not imply Team access. See Y.8.
Claude.ai. Anthropic’s web chat interface. Useful for thinking out loud and working with text. See Appendix A — Tool Atlas.
CLI (Command-Line Interface). A text-based interface where commands are typed at a prompt. Terminal is the most common CLI surface. See W.2 — Terminal fluency.
Coaching. Situation-driven, shorter-term, more directive support for a specific situation. Distinct from mentoring. See C.4 — Mentoring and sponsorship.
Codex. OpenAI’s coding agent. At Razorpay it is an optional secondary surface, not the default quota fallback; use it only when current support guidance confirms access for a bounded task. See Appendix A — Tool Atlas.
Cohort. A group of builders progressing through a belt or program element together. The Ship-to-Learn cohort is the structured Green Belt path.
Cohort lead. The role responsible for running a cohort: scheduling, evidence review, escalation. Role-named, not person-named.
Compass plugin. The version-locked plugin bundle used internally to ship the program’s pinned skill and MCP set. See W.7.
Connector. An integration that lets an AI surface read from or write to an external service (Slack, Google Workspace, a tracker, etc.). MCP servers are the most common connector shape. See Appendix B — Environment Setup.
Context (the pillar). One of Simon Willison’s three pillars: the information the agent has access to. Includes the prompt, files in the working directory, retrieved documents, and conversation history. See G.1 — The Three Pillars.
Council. In this playbook, the Staff+ Council described in the Council section. The standing community of senior contributors. The API Council is a different body covered in B.15.
Cowork. Anthropic’s desktop mode for non-engineers. See Appendix A — Tool Atlas.
Cursor. An AI-native IDE. Still in use by some teams. See Appendix A — Tool Atlas.
Decision archive. The numbered, immutable, append-only archive of RFCs the Council maintains. Doubles as a teaching corpus. See C.3.
Devstack. Razorpay’s internal development infrastructure (CI, deploy, environment). Owned by the devstack team. The relevant chapter pointer is Appendix B.
Docs (vs the log). Documentation is the canonical answer for how something works. The office-hours decision log is a running ledger that points at the docs and surfaces gaps. Distinct artefacts. See B.12.
Embedded sprint. A time-boxed (one-week) pattern where a senior contributor embeds with a team outside their own and ships with the team rather than for it. See B.13.
Enablement Stack. The 9-layer org-level model the playbook is structured against. See Prologue §0.4.
Figma MCP. The MCP connector that lets an agent read Figma frames and produce code. Covered in Y.9 and G.15.
Fall-back. The named path an agent or skill takes when the primary path fails (rate limit, timeout, classifier flag). See B.11.
FAQ. Frequently asked questions. See Appendix D.
gstack. Garry Tan’s framework for AI-assisted development with specialist roles. Profiled in Appendix N.2.
Golden set. A small, curated set of inputs with known good outputs used to evaluate a prompt or skill. See B.9 — Prompt evals.
Grandmaster. Historical name (replaced in v0.16) for the senior-IC community above Black Belt. The current term is Staff+ Council. See the Staff+ Council overview.
Green Belt. The third belt. Team velocity. Engineers around AI rather than just using it. Authors skills, runs subagents, ships across surfaces. See the Green Belt overview.
GREEN / YELLOW / RED. The readiness language used across the playbook. White Belt applies it through a seven-check manual setup gate; later modules reuse it as a self-check pattern. See W.8.
Grandmaster. See Staff+ Council.
GSD (Get Shit Done, also TÂCHES). A meta-prompting and context-engineering framework. Profiled in Appendix N.3.
Harness (the pillar). One of Simon Willison’s three pillars: the runtime around the model — tools, sandboxes, agents, hooks. See G.1.
Hook. A pre- or post-action automation in Claude Code that runs on specific events (before tool use, after tool use). See G.10.
IDE (Integrated Development Environment). A graphical environment for writing code. VS Code, Cursor, JetBrains products are common examples.
IETF (Internet Engineering Task Force). The standards body that publishes Internet RFCs. The program’s RFC pipeline borrows from IETF practice. See C.3, specifically RFC 7282 on consensus.
IP (Intellectual Property). The legal frame around who owns code, writing, and inventions. See C.5 — External voice and GitHub’s Balanced Employee IP Agreement as the public reference.
KB (Knowledge Base). A structured, persistent body of context the program accumulates. The opposite of re-deriving context every session. See Prologue §0.7 and Appendix N.
KB-driven development. The operating philosophy that ties the playbook together: don’t re-derive, accumulate. See Appendix N.1.
Layer 0. The foundation layer: environment, access, setup. White Belt is the curriculum step that owns Layer 0 readiness. See Prologue §0.4.
Liaison (Council). The rotating Council member who interfaces with engineering leadership for a quarter. See C.2.
LiteLLM. The open-source LLM proxy the program uses for routing and policy. See G.23.
LLM (Large Language Model). The model class that powers Claude, GPT, and similar systems. See Prologue §0.3.
LLM Wiki pattern. Karpathy’s anti-RAG pattern: a persistent, compounding wiki the agent reads from. Profiled in Appendix N.4.
MCP (Model Context Protocol). Anthropic’s protocol for connecting models to external tools and data sources. See B.1 — Authoring an internal MCP server.
Mentor / Mentoring. Relationship-driven, longer-term support. Distinct from coaching (situation-driven, shorter) and sponsorship (advocacy, not advice). See C.4.
Minimum viable wiki. A starter knowledge-base structure: index, log, schema, kb folder. See Appendix N.7 and the seed template.
MUST / SHOULD / MAY. RFC 2119’s vocabulary for stating requirements unambiguously. Used in the program’s RFC pipeline. See C.3.
NDA (Non-Disclosure Agreement). The legal frame around what an employee can and cannot disclose externally. Relevant to C.5 — External voice.
OKR (Objectives and Key Results). A common goal-setting framework. The playbook’s success metric (distinct from program OKRs) is an open question.
Office hours. A published, recurring slot where senior contributors take in-flight blockers from the broader team in public. The cheapest propagation move available. See B.12.
OSS (Open-Source Software). Code released under a license that permits free use and modification. See C.5 for the senior-engineer contribution patterns.
PCI (Payment Card Industry). Regulatory frame governing how payment card data is handled. See G.24.
PII (Personally Identifiable Information). Data that can identify a person. Subject to regulatory protection. See G.24.
Plugin. A bundled set of skills, MCPs, and configuration that can be installed as a unit. The program ships a pinned plugin. See B.3 — Cowork plugin marketplace.
POD. Razorpay’s term for a small product team, typically owning a single product surface. Used as a synonym for “team” elsewhere. The playbook prefers “team” in conceptual passages.
PR (Pull Request). A proposed change to a code repository, opened for review before merging. See W.12 and Y.13.
Prompt (the pillar). One of Simon Willison’s three pillars: the actual text the model sees on a given turn. See G.1.
Prompt eval. A test that runs a prompt against a known set of inputs and checks outputs against expected behaviour. See B.9.
Prompt injection. An attack where an attacker’s content reaches the model and tries to override the system prompt. See G.25.
Progressive disclosure. A skill-authoring pattern that keeps the default body small and reveals depth on demand. See B.7.
Quest. A graded artefact required to earn a belt. Each belt has at least one. Distinct from a boss fight (larger). See the belt READMEs.
RBI (Reserve Bank of India). The Indian financial regulator. Relevant to fintech compliance throughout the curriculum. See G.24.
Razorpay API Design Guide. The internal canonical reference for API design at Razorpay. The razorpay-api-review skill automates a large part of the rubric. See B.15.
Razorpay Cowork. See Cowork.
Redline. A specific item that must never appear in a prompt or external surface. Defined in G.22.
RFC (Request For Comments). A written proposal for a change, reviewed and discussed before commitment. The program’s RFC pipeline is described in C.3. See also the RFC template.
RFD (Request For Discussion). Oxide Computer’s term for the same artefact. Cantrill’s “Requests for Discussion” is the foundational public reference. See C.3.
Sandbox. An isolated environment where code can run without affecting the wider system. Claude Code uses sandboxes for shell execution. See W.2.
SDK (Software Development Kit). A library or set of libraries for building on a platform. The Claude Agent SDK is covered in B.4.
Self-assessment. The Prologue chapter that helps a reader pick their starting belt. See Prologue §0.10.
Ship-to-Learn. A 7-week structured cohort track for Green Belt candidates. Treated in Appendix M (currently planned; see the Master Index for the section overview).
Skill (Anthropic). A reusable workflow defined in a SKILL.md file that an agent can invoke. See G.6, G.7, and the skill templates.
Skill pack. A bundled set of skills published as a unit. See B.2.
Slash. Razorpay’s internal AI copilot. See Appendix A.
Slash command. A keyword that invokes a specific skill in Claude Code. See G.10.
Sponsorship. Advocacy in rooms the sponsee is not in. Distinct from mentoring (advice). At the senior level, sponsorship is the more under-supplied role. See C.4.
Staff+. A collective term for engineering levels at Staff and above (Staff, Senior Staff, Principal, Distinguished, Fellow). The senior-IC track. Used as the framing term for the Council.
Staff+ Council. See Council.
Subagent. A worker agent invoked by a parent agent to handle a sub-task. See G.8.
Substrate. The supporting structure (manifest.yml, slugs.yml, INDEX.md, scripts) that lets the playbook content stay coherent across versions.
Three pillars. Simon Willison’s prompt-context-harness model. Used as a tagging system across the curriculum. See G.1.
Tool (in agent context). A function the agent can call to take an action (read a file, query a database, send a message). See B.6 — Tool design.
Track. A subdivision within a belt or section. Green Belt has Parts A, B, C; the appendices have multiple tracks (methodologies, templates, etc.).
Vertex AI. Google Cloud’s AI platform. Razorpay routed Claude through Vertex until the March-2026 migration to the LiteLLM gateway; it is no longer in the active model path. If you see Vertex env vars in a teammate’s setup, they are stale. See W.4 and W.6.
White Belt. The first belt. Setup, first commit, first PR. See the White Belt overview.
Working forum (Council). Layer 2 of the senior-IC community. Active participants who contribute to RFC review, mentoring, and cross-team alignment work. See C.2.
Worktree. A git feature that lets you check out multiple branches into separate working directories simultaneously. Used to run multiple Claude instances safely. See G.9.
Yellow Belt. The second belt. AI in the daily rhythm. Ships a merged PR. See the Yellow Belt overview.
Maintenance
Section titled “Maintenance”The glossary is revised as new terms enter the curriculum. Pull requests welcome: a new term should land here when it lands in any chapter, with the cross-reference set at the same time.
When a term has multiple meanings in the curriculum (e.g., Council), the entry disambiguates explicitly.
When a term is deprecated (e.g., Grandmaster, replaced by Staff+ Council in v0.16), the deprecated term’s entry stays as a pointer to the current term so historical references resolve.
See also: Appendix N — Methodologies & Frameworks for deep treatment of operating frameworks. Appendix J — Reading List for the public sources cited across the curriculum.