Vibe Coding & AI Agents
150 terms · English and Spanish
This is how people talk where code gets written with AI: about agents, about context, about what can go wrong.
Many of these are ordinary words that picked up a new meaning — agent, context, memory, checkpoint — and the sentence shows which meaning is in play. The Spanish pair sits next to the English one — sometimes it's the same word, sometimes Spanish found its own.
Vibe codingAgentic codingAI pair programmerCopilot (concept)Coding agentAutonomous agentAgenticAgent loopAgentic workflowReActTool callingFunction callingMulti-agent systemOrchestrationHuman-in-the-loopPrompt injectionJailbreakGuardrailsSandboxSemantic searchKnowledge baseContext engineeringMemory (agent)Prompt cachingLong contextMCP serverCursorGitHub CopilotClaude CodeAgent modeAutocompleteTab completionSlash commandReasoning modelExtended thinkingDiffCheckpointStreamingEvalToken costRate limitShip itAgentic engineeringSpec-driven developmentPrompt templateStructured outputCodebase indexingBackground agentTest generationCode review agentVibe coderYOLO modeAccept allOne-shotAI slopAutonomous codingHuman-on-the-loopOrchestratorSubagentSupervisor agentSwarmTask decompositionPlanning (agent)ReflexionSelf-healingTree of ThoughtsAgent handoffReflection loopHeadless modeMeta-promptingPrompt chainingRole promptingNegative promptJSON modeOutput parserGroundingPrompt libraryCustom instructionsContext compactionToken budgetCLAUDE.mdChunkingRetrievalRerankingHybrid searchAgentic RAGCitation (grounded)Top-kMCP clientMCP toolTool schemaRepo mapWindsurfGemini CLIAiderClineDevinReplit Agentv0Composer (Cursor)Ghost textFrontier modelTokens per secondTime to first tokenRollbackSkill (agent)HookPluginModel routingFallback modelWorktree (agent isolation)Test-time computePR agentBoilerplate generationPrompt-driven developmentPermission promptTool approvalSWE-benchLLM-as-judgeRed teamingPrompt kiddiePrompt-firstHallucinated APIAutomation biasFan-out/fan-inPlan-and-ExecuteScaffolding (agent)A2A protocolSelf-consistencyIndirect prompt injectionPrompt leakingMCP resourceElicitation (MCP)Sampling (MCP)stdio transportJSON-RPCStreamable HTTPContext rotLost in the middleKV cacheNeedle in a haystackMinimum viable contextEpisodic memoryScratchpadEffort settingContext injectionGraphRAGCorrective RAGBoltLovable
Tools
AI pair programmerProgramador en pareja con IA
An AI assistant that codes alongside you, suggesting lines and reacting to edits like a second developer at the keyboard.
My AI pair programmer caught the off-by-one error before I even ran the code.
Copilot (concept)Copiloto (concepto)
A generic name for an in-editor AI helper that offers suggestions and drafts while you stay in control of the work.
Every major IDE now ships some kind of copilot for autocompleting code.
Coding agentAgente de codificación
An AI system that autonomously reads a codebase, changes files, and runs tools to complete a programming task end to end.
We gave the coding agent a bug report and it opened a pull request an hour later.
Tool callingLlamada a herramientas
When a model asks the surrounding system to run a specific function, like a search or a file edit, and uses the returned result.
The assistant answered the weather question by tool calling a forecast API.
Function callingLlamada a funciones
A model feature that returns a structured request to invoke a named function with arguments, so code can execute it reliably.
We defined a create_ticket schema and let function calling fill in the fields.
SandboxEntorno aislado
An isolated environment where AI-generated code or agent actions can run without touching the real system.
The agent tested its script in a sandbox, so the risky delete never reached real files.
Prompt cachingCaché de prompts
Reusing the processed form of a repeated prompt prefix so later requests run faster and cost less.
Caching the long system message cut our latency in half.
MCP serverServidor MCP
A program that exposes tools, data, or prompts to AI apps over the Model Context Protocol so any client can use them.
We wrote an MCP server for our database and Claude could query it directly.
Cursor
An AI-first code editor built on VS Code, known for chat, inline edits, and an agent that works across a whole project.
She moved from plain VS Code to Cursor for its whole-repo edits.
GitHub Copilot
GitHub's AI coding assistant that suggests code in the editor and answers questions about a project.
GitHub Copilot autocompleted the whole test file from a one-line comment.
Claude Code
Anthropic's command-line coding agent that reads a repo, edits files, and runs tools from the terminal.
I asked Claude Code to add logging and it patched every service in one pass.
Agent modeModo agente
An editor or tool setting that lets the AI act on its own — editing files and running commands — instead of only suggesting.
Flip on agent mode and it will refactor the module without you approving each line.
AutocompleteAutocompletado
Real-time suggestions that finish the code you are typing based on the surrounding context.
Autocomplete guessed the entire loop body after I wrote the signature.
Tab completionCompletado con Tab
Accepting an inline AI suggestion by pressing Tab, letting you write code in quick confirmed bursts.
He wrote the parser mostly by tab completion, hitting Tab line after line.
Slash commandComando de barra
A shortcut typed with a leading slash that triggers a predefined action or prompt in an AI tool.
Typing /test told the agent to generate and run unit tests.
CheckpointPunto de control
A saved snapshot of project state you can return to if an AI's later changes go wrong.
Before the big refactor I set a checkpoint to roll back to.
Streaming
Delivering a model's output token by token as it is generated, so text appears gradually instead of all at once.
Streaming let us watch the explanation build instead of staring at a spinner.
Rate limitLímite de tasa
A cap on how many requests or tokens you may send in a period, enforced by an API provider.
The batch job hit the rate limit and had to back off and retry.
Structured outputSalida estructurada
Model responses forced into a fixed shape, like JSON matching a schema, so code can parse them reliably.
Structured output returned clean JSON we dropped straight into the database.
Codebase indexingIndexación del código
Building a searchable map of a project so an AI can quickly retrieve the files relevant to a task.
After codebase indexing, the assistant jumped to the right function instantly.
Background agentAgente en segundo plano
An agent that runs a task on its own away from the editor, reporting back when the work is finished.
I dispatched a background agent to update the docs while I kept coding.
YOLO modeModo YOLO
A setting that lets an agent make edits and run commands without asking you to approve each step.
He ran the agent in YOLO mode and it refactored, tested, and committed on its own.
Accept allAceptar todo
Applying every change an AI proposed at once, without reviewing them one diff at a time.
The suggestions looked fine, so she hit accept all and moved on.
Headless modeModo headless
Running an agent with no interactive UI, from a script or command line, for automation.
In headless mode the agent ran nightly in CI with no one watching.
JSON modeModo JSON
A setting that forces a model to reply with syntactically valid JSON.
JSON mode guaranteed the response parsed without a try/except around it.
Output parserParser de salida
A component that turns a model's free-form text into a structure a program can use.
The output parser pulled the fields out of the model's messy answer.
Custom instructionsInstrucciones personalizadas
Standing user directions applied to every request, like a project's house rules.
Custom instructions told the assistant to always answer in metric units.
CLAUDE.md
A project file of instructions an agent reads as standing context for the repository.
They put the build commands in CLAUDE.md so the agent stopped guessing them.
MCP clientCliente MCP
The side inside a host or agent that connects to an MCP server and calls its capabilities.
The MCP client discovered the server's tools and offered them to the model.
MCP toolHerramienta MCP
An executable capability an MCP server exposes for an agent to call.
We added a run_query MCP tool and the agent used it to read the database.
Tool schemaEsquema de herramienta
A formal description of a tool's inputs and outputs that a model uses to form a valid call.
A clear tool schema stopped the model from passing the date in the wrong format.
Repo mapMapa del repositorio
A compact map of a repository's structure fed to an agent to help it navigate the code.
The repo map let the agent find the right module without reading every file.
Windsurf
An agentic IDE whose Cascade agent plans and edits across a whole project.
She built the prototype in Windsurf, letting Cascade wire up the screens.
Gemini CLI
Google's open-source terminal agent that pairs a ReAct loop with tools and MCP.
He drove Gemini CLI from the shell to scaffold and test the service.
Aider
An open-source terminal agent that treats code as a Git repo, committing its own edits.
Aider made each change a clean commit, so rolling back was trivial.
Cline
An open-source agent inside VS Code that edits files and runs commands with your approval.
Cline proposed the shell command and waited for a click before running it.
Devin
An autonomous AI software engineer that takes a task end to end, from plan to pull request.
They handed Devin the backlog item and reviewed the pull request it opened.
Replit Agent
An agent that builds and deploys a working app from a plain-language description.
From one paragraph, Replit Agent stood up a live to-do app in minutes.
v0
A generator from Vercel that turns a description or screenshot into React UI components.
She dropped a mockup into v0 and got a styled component back to tweak.
Composer (Cursor)
Cursor's agent model, tuned for fast agentic coding across a whole project.
Composer edited a dozen files at once and kept the app compiling.
Ghost text
The grey inline suggestion an editor shows ahead of your cursor, accepted with a key.
The ghost text guessed the rest of the line and he pressed Tab to take it.
RollbackReversión
Undoing an agent's changes to return the project to its earlier state.
The refactor broke the build, so she triggered a rollback to last night's state.
Skill (agent)Habilidad (agente)
A reusable ability or extension that teaches an agent to handle a class of tasks.
They packaged the deploy steps as a skill the agent could invoke by name.
Hook
A handler that fires on an agent event, before or after an action, to automate a response.
A hook ran the formatter every time the agent saved a file.
Plugin
An installable extension that adds commands or integrations to an agent.
The Jira plugin let the agent close tickets as it finished them.
Fallback modelModelo de respaldo
A backup model switched to when the primary one fails or hits a limit.
When the main API timed out, the fallback model kept the app answering.
Worktree (agent isolation)Worktree (aislamiento de agente)
A separate Git working copy for a parallel agent so its edits don't clash with others.
Each agent got its own worktree, so three refactors ran without conflicts.
Permission promptSolicitud de permiso
An agent's request for your go-ahead before an action like writing a file or running a command.
A permission prompt appeared before it deleted the folder, and she declined.
Tool approvalAprobación de herramienta
A user's explicit go-ahead for an agent to run a particular tool call.
Tool approval was required for anything that touched production.
stdio transportTransporte stdio
A local MCP transport that carries messages over a process's standard input and output.
For a local tool, stdio transport was simplest — no ports, just pipes.
JSON-RPC
A format for remote calls over JSON that underpins the message exchange in MCP and A2A.
Every MCP request on the wire was a JSON-RPC message with a method and params.
Streamable HTTP
The current remote MCP transport, which replaced the earlier standalone SSE transport.
They moved the remote server to Streamable HTTP and dropped the old SSE setup.
Bolt
An AI app builder on WebContainers that assembles and hosts a project right in the browser.
In Bolt she described the dashboard and previewed it running seconds later.
Lovable
An AI app builder that turns a plain-language description into a full app with database and auth.
From a paragraph, Lovable produced a working app with login and a database.
Agents
Agentic codingCodificación agéntica
A style of development where an AI agent plans, edits files, and runs commands toward a goal with little step-by-step guidance.
She switched to agentic coding and let the tool fix the failing tests itself.
Autonomous agentAgente autónomo
A program driven by a model that pursues a goal on its own, choosing actions and tools without waiting for each instruction.
The autonomous agent booked the flight, filled the form, and emailed a summary unprompted.
AgenticAgéntico
Describing software that acts with initiative — planning, using tools, and adapting — rather than only answering a single prompt.
The new release feels agentic: it chases down dependencies instead of just suggesting them.
Agent loopBucle del agente
The repeating cycle in which an agent observes results, decides a next action, acts, and reads the outcome until the task is done.
Each turn of the agent loop ran a test, read the failure, and edited the file.
Agentic workflowFlujo de trabajo agéntico
A task broken into steps an AI agent carries out in sequence, calling tools and checking its own progress along the way.
Their release checklist became an agentic workflow that tags, builds, and drafts the notes.
ReAct
A pattern where a model alternates between reasoning in words and taking actions, using each observation to guide the next step.
With ReAct, the agent thought aloud, searched the docs, then answered.
Multi-agent systemSistema multiagente
A setup where several specialized agents collaborate or split a task, passing work and messages between one another.
In their setup a planner delegates coding to workers and a reviewer checks the output.
OrchestrationOrquestación
Coordinating multiple agents, tools, or steps so they run in the right order and hand results to each other.
The orchestration layer decides which agent handles a ticket and when to escalate.
Human-in-the-loopHumano en el bucle
A design where a person reviews or approves an AI's actions at key points before they take effect.
Deployments stay human-in-the-loop: the agent proposes, an engineer clicks approve.
GuardrailsSalvaguardas
Rules and filters placed around a model to keep its inputs and outputs within safe, allowed bounds.
Guardrails blocked the agent from running shell commands outside the project folder.
Memory (agent)Memoria (agente)
A store that lets an agent keep facts across turns or sessions, so it recalls earlier decisions and preferences.
Thanks to memory, the agent remembered our naming convention from last week.
Agentic engineeringIngeniería agéntica
A discipline of building software by directing AI agents, focusing human effort on specs, review, and coordination.
Their team reorganized around agentic engineering, with fewer hands writing code directly.
Code review agentAgente de revisión de código
An AI that reads a proposed change and flags bugs, style issues, or risks like a human reviewer would.
The code review agent caught a race condition two humans had missed.
Autonomous codingCodificación autónoma
Development where an agent carries a task from prompt to working code with minimal human input.
Overnight autonomous coding turned the ticket into a tested, merged fix.
Human-on-the-loopHumano sobre el bucle
A design where a person oversees an agent and steps in on problems, rather than approving each action.
The team stayed human-on-the-loop, watching the dashboard and intervening only when it drifted.
OrchestratorOrquestador
A component or agent that coordinates several agents or steps toward one goal.
The orchestrator split the job, sent parts to workers, and merged their answers.
SubagentSubagente
A helper agent an orchestrator delegates a sub-task to, working in its own separate context.
The main agent spun up a subagent just to search the docs.
Supervisor agentAgente supervisor
An agent that assigns work to others and checks their results before accepting them.
A supervisor agent rejected the draft and sent it back for another pass.
SwarmEnjambre
Many simple agents working together on a task without a strict central controller.
A swarm of scrapers divided the sites among themselves and reported back.
Task decompositionDescomposición de tareas
Breaking a large task into smaller sub-tasks an agent can tackle step by step.
Task decomposition turned 'build the API' into a dozen small, checkable steps.
Planning (agent)Planificación (agente)
The stage where an agent works out a sequence of actions before it starts carrying them out.
The agent spent its first turn planning, then executed without backtracking.
Reflexion
A technique where an agent writes out what went wrong and uses that note to do better on the next try.
After the failed run, Reflexion let the agent avoid the same dead end.
Self-healingAutorreparación
An agent's ability to detect and fix a failure, like a broken test, on its own.
The self-healing pipeline caught the crash and patched it before anyone woke up.
Agent handoffTraspaso entre agentes
Passing a task and its context from one agent to another partway through the work.
After research, a handoff moved the job to the agent that writes the code.
Reflection loopBucle de reflexión
A cycle where an agent judges its own output and redoes it until a criterion is met.
The reflection loop rewrote the summary twice before it read cleanly.
Agentic RAGRAG agéntico
RAG where an agent decides on its own when and what to search, refining retrieval as it goes.
Agentic RAG searched again after realizing the first results missed the point.
PR agentAgente de PR
An agent that packages changes as a pull request, with a description and edits from review.
The PR agent opened a tidy pull request and even answered the reviewer's comment.
Plan-and-Execute
An agent architecture that first builds a full plan, then carries out the steps in order.
Plan-and-Execute laid out all ten steps up front and ran them without re-planning.
Scaffolding (agent)Andamiaje (agente)
The software wrapper around a model — loop, tools, parsing — that turns it into an agent.
The model was fine; weak scaffolding was why the agent kept losing track.
A2A protocolA2A
Agent2Agent: an open protocol letting agents from different frameworks talk to each other.
Over A2A, the planning agent handed the task to a vendor's billing agent.
Elicitation (MCP)
An MCP feature where a server asks the user for missing input through the client.
Mid-task, elicitation popped a prompt asking which environment to deploy to.
Minimum viable contextContexto mínimo viable
The least context an agent needs to do a task well; extra can hurt more than help.
Trimming to the minimum viable context made the agent faster and more accurate.
Episodic memoryMemoria episódica
An agent's memory of specific past episodes or sessions that it can look back on.
Episodic memory let it recall how it solved the same bug two projects ago.
Corrective RAG
RAG that checks its retrieved material and searches again when the results look weak.
Corrective RAG noticed the passages were off-topic and re-ran the search.
How the work goes
Vibe coding
Building software by describing what you want in plain language and letting AI write the code, often accepting output without reading it closely.
He vibe-coded a landing page over coffee, just typing requests until it looked right.
Diff
A view of proposed changes showing added and removed lines, so you can review before applying them.
The agent showed a diff and I accepted only the parts I trusted.
Spec-driven developmentDesarrollo guiado por especificación
An approach where you first write a clear specification and let AI turn it into a plan, tasks, and code.
Spec-driven development meant arguing about the doc, not the pull request.
Prompt templatePlantilla de prompt
A reusable prompt with placeholders you fill in per request, keeping instructions consistent across calls.
We stored a prompt template so every summary followed the same format.
Test generationGeneración de pruebas
Having AI write test cases for existing code to check its behavior and guard against regressions.
Test generation gave us 40 cases for the parser we'd been too lazy to cover.
Prompt chainingEncadenado de prompts
Linking several prompts so the output of one becomes the input of the next.
Prompt chaining first summarized the doc, then turned the summary into slides.
Prompt libraryBiblioteca de prompts
A collection of reusable, tested prompts a team shares and reaches for.
New hires pulled from the prompt library instead of writing prompts from scratch.
Boilerplate generationGeneración de plantilla base
Generating the repetitive starter code — project scaffold, wiring — so you skip the busywork.
Boilerplate generation set up the routes and config in seconds.
Prompt-driven developmentDesarrollo guiado por el prompt
Development where the prompt is the main driving artifact, not hand-written code.
In prompt-driven development they versioned the prompts like source files.
Fan-out/fan-in
A pattern that sends sub-tasks to parallel agents (fan-out) and merges their results (fan-in).
Fan-out/fan-in reviewed twenty files at once and pooled the findings.
Context
Semantic searchBúsqueda semántica
Finding results by meaning rather than exact keywords, by comparing the vector representations of query and content.
Semantic search surfaced the auth code even though the query never said 'login'.
Knowledge baseBase de conocimiento
A curated store of documents or facts an AI can retrieve from to ground its answers in your material.
Support answers now pull from a knowledge base of past tickets and docs.
Context engineeringIngeniería de contexto
Deciding what information to place in a model's limited working window so it has just what it needs, and no clutter.
Good context engineering meant sending the diff and the failing test, not the whole repo.
Long contextContexto largo
The ability of a model to take in very large inputs at once, such as whole codebases or lengthy documents.
With long context we pasted the entire spec and asked for a review in one go.
Role promptingPrompting por rol
Assigning the model a role or persona to shape the tone and expertise of its answer.
Role prompting it as a strict reviewer made the feedback far sharper.
Negative promptPrompt negativo
Telling a model what it should not include in its output.
A negative prompt kept the marketing copy free of buzzwords.
Context compactionCompactación de contexto
Compressing accumulated context or history so it fits the window without losing the point.
When the chat grew long, context compaction summarized the early turns.
Token budgetPresupuesto de tokens
The cap on tokens allotted to a task or step, controlling cost and how much context fits.
A tight token budget forced the agent to retrieve only the top three files.
ChunkingFragmentación (chunking)
Splitting documents into smaller pieces so they can be indexed and retrieved in RAG.
Better chunking kept related paragraphs together and improved the answers.
RetrievalRecuperación
Fetching the pieces of a knowledge base most relevant to a query.
Retrieval pulled the three sections that actually mentioned the error code.
RerankingReordenamiento
Reordering retrieved candidates by relevance with a second, more precise ranker.
Reranking pushed the one useful result above ten near-misses.
Hybrid searchBúsqueda híbrida
Combining keyword search and vector search to get better results than either alone.
Hybrid search caught both the exact error string and its paraphrases.
MCP resourceRecurso MCP
Data or content an MCP server exposes to an agent as context, as opposed to a runnable tool.
The server offered the wiki as an MCP resource the agent could read but not change.
Context rot
The decline in a model's quality as the context grows longer, even on simple tasks.
Past a hundred pages, context rot set in and it started missing obvious facts.
Lost in the middle
The effect where a model uses information from the middle of a long context worse than the ends.
The key clause sat mid-document and, lost in the middle, the model skipped it.
KV cacheCaché KV
The cache of attention keys and values that speeds generation but grows with context length.
A huge KV cache ate the memory once the conversation stretched past its limit.
Needle in a haystack
A long-context test that hides a fact in a large input to see if the model can find it.
The needle-in-a-haystack test buried a code in 200 pages; the model retrieved it.
ScratchpadBloc (scratchpad)
A working area where a model jots intermediate reasoning before giving its answer.
On its scratchpad the model worked through the math, then wrote the clean answer.
Context injectionInyección de contexto
Inserting needed data into the context before a request; in attacks, a malicious insertion.
Context injection slipped the user's timezone in so the times came out right.
GraphRAG
RAG over a knowledge graph, retrieving by links between entities rather than vector nearness alone.
GraphRAG followed the ownership edges to answer who reported to whom.
Models
Reasoning modelModelo de razonamiento
A model tuned to work through problems in explicit steps before answering, trading speed for stronger logic.
For the tricky algorithm we switched to a reasoning model and it caught the edge case.
Extended thinkingPensamiento extendido
A mode where a model spends extra compute reasoning privately before it responds, improving hard answers.
With extended thinking on, it planned the migration before writing a single line.
Eval
A test that measures how well a model or agent performs on a task, used to compare versions and catch regressions.
Our eval scored each prompt on 200 cases before we shipped the change.
Token costCoste en tokens
The price of a request, driven by how many tokens of input and output it consumes.
Trimming the system prompt dropped our token cost per call noticeably.
AI slopSlop de IA
Low-quality, generic content or code churned out by AI without care or review.
The repo filled with AI slop: plausible functions that didn't quite work.
Grounding
Tying a model's answer to verifiable sources or data to cut down on made-up claims.
Grounding the bot in the manual stopped it from inventing menu options.
Top-k
A parameter capping how many candidates are considered, in retrieval or in sampling.
Lowering top-k to five kept the context focused and the cost down.
Frontier modelModelo de frontera
The most capable model available at a given moment, at the leading edge of the field.
For the hardest tasks they paid for the frontier model and cached everything else.
Tokens per secondTokens por segundo
A model's generation speed: how many tokens it produces each second.
The smaller model's higher tokens per second made the chat feel instant.
Time to first tokenTiempo hasta el primer token
The delay between sending a request and the first token of the reply appearing.
Cutting time to first token made the assistant feel responsive on long answers.
Model routingEnrutamiento de modelos
Choosing which model handles a request based on its difficulty, cost, or type.
Model routing sent quick questions to a cheap model and hard ones to the big one.
Test-time computeCómputo en la respuesta
Compute spent thinking at answer time rather than during training, trading effort for quality.
Giving it more test-time compute let the model catch its own arithmetic slip.
LLM-as-judge
Using a model to score or compare other answers against stated criteria.
LLM-as-judge ranked the two drafts so they didn't have to read both closely.
Hallucinated APIAPI alucinada
A call a model invents to a method or library that does not actually exist.
The code looked right until the hallucinated API threw 'module not found'.
Sampling (MCP)
An MCP feature where a server requests a model completion that the client runs.
Through sampling, the server asked the client's model to summarize the file.
Effort settingAjuste de esfuerzo
A control for how much thinking budget a model spends on reasoning.
He raised the effort setting for the proof and lowered it for small edits.
Prompts and attacks
Prompt injectionInyección de prompt
An attack that hides instructions in content a model reads, tricking it into ignoring its original task or leaking data.
A web page said 'ignore prior rules' — an injection the agent almost obeyed.
Jailbreak
A crafted input that gets a model to bypass its safety rules and produce content it would normally refuse.
The forum traded jailbreak prompts until the vendor patched the model.
One-shot
Getting a working result from a single request, with no follow-up rounds of correction.
He tried to one-shot the whole game and, to his surprise, it ran on the first try.
Tree of Thoughts
A reasoning method that explores a branching tree of intermediate ideas instead of a single line of thought.
With Tree of Thoughts, the model weighed three approaches before committing to one.
Meta-promptingMetaprompting
Using a model to write or improve the prompts you then feed to a model.
She used meta-prompting to have the AI rewrite her clumsy instructions.
Prompt-first
An approach where the prompt is the primary design artifact and code comes second.
Their prompt-first workflow argued over wording before anyone opened the editor.
Self-consistency
Generating several reasoning paths and picking the answer the majority agree on.
Self-consistency ran the problem five times and voted on the final number.
Indirect prompt injectionInyección de prompt indirecta
Injecting instructions through outside data — a web page or file — that an agent reads.
A comment on the page carried an indirect prompt injection the agent nearly ran.
Prompt leakingFuga de prompt
An attack that coaxes a model into revealing its hidden system prompt.
A clever question achieved prompt leaking, exposing the app's secret instructions.
Everything else
Ship itA producción
A rallying phrase for releasing quickly, often after AI did the heavy lifting and the author trusts it enough to deploy.
Tests pass, looks fine — ship it.
Vibe coder
Someone who practices vibe coding, steering development in plain language and leaning on AI, often without reading the code.
As a self-taught vibe coder, she shipped three apps without ever learning the syntax.
Citation (grounded)Cita (con fuente)
A source reference attached to a claim in an answer so a reader can verify it.
Each sentence carried a citation, so the reviewer could check every claim.
SWE-bench
A benchmark that grades coding agents on real GitHub issues using the projects' own unit tests.
Their agent's SWE-bench score jumped once it learned to run the tests first.
Red teaming
Deliberately attacking a system with injections and jailbreaks to find holes before release.
A week of red teaming surfaced three prompts that leaked the system message.
Prompt kiddie
A dismissive term for someone who builds only by prompting, without understanding the code.
Real bugs exposed him as a prompt kiddie the moment the AI's output broke.
Automation biasSesgo de automatización
The tendency to trust an automated or AI output without checking it.
Automation bias let a wrong number sail through because 'the AI said so'.
Conversations from this deck
Reviewing a pull request an agent opened
DiffPR agentTest generation
Deck: Vibe Coding & AI Agents
The standup about AI agents: ten words that come one after another
Agentic codingCoding agentAgent loop
Deck: Vibe Coding & AI Agents
The agent reached production: eleven words from one night
Headless modeYOLO modePermission prompt
Deck: Vibe Coding & AI Agents
In the app, these cards come as the day's set — already made.
