→ un peldaño no reemplaza al anterior. cada uno se apoya en el que sigue.
● rungs 01 + 02 · the loop you can build tonight
A MODEL THAT CRITIQUES ITSELF ALREADY WINS.
SELF-REFINE
~20% average gain over single-shot output across 7 tasks. Three prompts: generate, critique, revise. No training, no infra. Madaan et al. 2023.
APE
24 of 24 tasks matched or beat human-written instructions. Machine IQM 0.810 vs human 0.749. Zhou et al. 2022. El modelo ya escribe instrucciones a nivel humano.
● rung 03 · OPRO · Google DeepMind
"TAKE A DEEP BREATH" scored 80.2% on GSM8K math. the human classic scored 71.8%.
→ the model found that prompt by reading its own score history. nobody wrote it. · Yang et al. 2023 · PaLM 2-L · no-instruction baseline: 34.0%
● rung 04 · evolution eats the evolver
PROMPTBREEDER MUTATES THE MUTATOR.
A genetic algorithm over a population of prompts. The twist: the mutation operators are themselves prompts, and they also get mutated and selected. Self-referential all the way down. La evolución también evoluciona.
58.4%
hand-written Chain of ThoughtGSM8K zero-shot · PaLM 2-L
BACKPROP WORKED FOR WEIGHTS. NOW IT WORKS FOR WORDS.
DSPY
33% to 82% on a GPT-3.5 math pipeline, with zero hand-written prompts. You declare input and output signatures; a compiler bootstraps the demos and optimizes the pipeline. Khattab et al. 2023.
TEXTGRAD
GPQA 51% to 55% on GPT-4o, plus a 20% relative gain on LeetCode Hard. Every component is a graph node; natural-language "gradients" flow backward telling each node how to change. Yuksekgonul et al. 2024. Los gradientes son oraciones.
● rung 07 · 2025 · the frontier
GEPA: REFLECTION BEATS REINFORCEMENT LEARNING.
+20%
over GRPO reinforcement learning, at best+6% on average · Agrawal et al. 2025 · ICLR 2026 oral
35x
fewer rollouts than the RL it beatlanguage feedback is denser than a scalar reward
+10%
over MIPROv2, the prior best prompt optimizerincluding +12% on AIME 2025 math
→ reflexiona en lenguaje natural sobre sus propios trazos de ejecución, muta el prompt, y guarda una frontera de Pareto de ganadores diversos. No necesitas RL. Necesitas un buen bucle.
● rung 07, en detalle · words vs. weights
WORDS BEAT WEIGHTS. SAME RESULT, FEWER TRIES.
RL pipeline · GRPO24,000 rollouts
Thousands of gradient rollouts to fit one task. Miles de intentos, una señal numérica cada vez.
GEPA · reflective evolutionup to 35× fewer
Reflect on the trace, mutate the prompt, keep the Pareto winners. Reflexionar, mutar, seleccionar. Eso es todo.
→ GEPA beats GRPO by 6% on average, up to 20%, using up to 35x fewer rollouts. Agrawal et al. 2025, ICLR 2026 oral. El lenguaje es un medio de aprendizaje más rico que un premio escalar.
● rung 07, in plain english · how it learns
IT READS ITS OWN FAILURES. THEN REWRITES ITS OWN INSTRUCTIONS.
TRYrun the task, and keep the whole story of what happened — not just the score
READan LLM reads that story and says, in plain words, why it failed
REWRITEit edits its own prompt to fix that exact mistake, and tests the new version
KEEPevery version that wins somewhere survives — and the best ideas get combined
→ your error messages are training data. Tus mensajes de error son datos de entrenamiento. Un número te dice que fallaste; el lenguaje te dice por qué.
EL BUCLE SE CIERRA.
The machine now searches prompt space better than you do. The question is no longer how to write a prompt. It is what loop you hand your criterion to.
● ACT III se lo hicimos a nuestro propio repo
THE PLAYBOOK, IN PRODUCTION.
CRITERIONa judgment you keep making by hand becomes a versioned prompt with a cache
TEMPLATEa request you keep typing becomes a named .md file with slots to fill
ROUTERa 3-tier detector maps a new terse ask to the right template. alias, keyword, then LLM
IMPROVERanything that fits no template gets rewritten against the same 10-directive rubric
COMPOSERone goal fans out into named sub-prompts for teammate agents
ONE UPagents read telemetry and edit the instructions that produce future agents
● paso uno · mine your own history
YOUR HISTORY IS TRAINING DATA.
6,366
total Claude Code sessions on disk1.18 GB of JSONL transcripts
5,963
machine sessions, filtered outautomation calling claude -p under the hood
403
prompts a human actually typedthe only ones that teach you how you really ask
18
named, parameterized pattern templatesplus a 10-entry directive vocabulary, each with an evidence trail
→ tu historial ya es un dataset. Los prompts que repites son plantillas esperando nombre.
● paso dos · route, honestly measured
CHEAPEST TIER FIRST. MEASURED, NOT VIBES.
Alias hit, then keyword signature, then an LLM for the long tail. The signatures live once, in the pattern file. Tres niveles: alias, keyword, LLM.
56%
of 397 real prompts claimed by keyword signaturesthe other 43% is a genuine long tail. it goes to the LLM tier
36/36
P02 improve-after-build routing precisiona strong keyword signal earns its fast path
0/8
P06 events-workflow keyword precisiontoo broad. it should always fall through to the LLM. knowing this is the point
● demo en vivo · recorded output as the safety net
ONE TERSE ASK BECOMES THE FULL PROMPT.
terminal · recorded 22 minutes before this talkprompts.route
The conductor decomposes the task, writes each expert's instruction, and calls a clean instance for every one. No expert sees the task, the transcript, or the other experts — only its own brief. Un director, expertos frescos: cada uno ve solo su instrucción.
→ experts chosen per task, briefs written by the conductor · synthesize + flag contradictions · runnable tonight: /examples/04_conductor.py
This deckbuilt by a metaprompt · today · in one session
The prompt that built this deck ships inside the deck. Fetch it yourself, right now.
Moodboard: a Pinterest board, reduced to a palette by a script
Research: two agents read 9 papers and mined 1 repo
Copy + build: one conductor session, CC-PRO method
El prompt que construyó este deck viaja dentro del deck
terminal · liveawaiting
> curl /.well-known/ai-agent.json
● one rung up · and one honest confession
AGENTS THAT EDIT THE AGENTS' INSTRUCTIONS.
CALIBRATION
Reads session telemetry, finds what an agent keeps getting wrong, edits the agent's instruction file. Evidence-gated. Never touches working code. Un agente que corrige el manual de otro agente.
OUROBOROS
Plan, critique through adversarial lenses, revise, K layers, stop on convergence. Its own docs name the lineage: Self-Refine, Reflexion, Debate, CRITIC. Honestidad: ingeniería bien hecha, no un mecanismo nuevo.
THE FRAME
Most "new" agent frameworks are one of these loops wearing a trench coat. Knowing the ladder means you can name what you're actually running. Y saber qué estás corriendo es la mitad del criterio.
● CALIBRATION el manual, no el código
THE AGENT THAT REWRITES THE MANUAL.
Same mistake, three sessions running — then the instruction file changes. Nunca el código de trabajo.
→ evidence-gated · never touches working code · un agente que corrige el manual de otro
● el plan que se corrige a sí mismo · four lenses, three layers
OUROBOROS. PLAN, CRITIQUE, REPEAT.
→ Self-Refine, Reflexion, Debate, CRITIC corren este mismo bucle · ingeniería bien hecha, no un mecanismo nuevo
● ACT IV abre el saco · name what you're running
OPEN THE COAT. NAME THE LOOP.
the hype stackthe ladder · la escalera
AgentForge Proself-critique loop
Self-Refineself-critique loop
AutoPilot Xscore-history optimizer
OPROscore-history optimizer
SwarmMindconductor + experts
Meta-Promptingconductor + experts
PromptOS 2.0evolutionary search
Promptbreederevolutionary search
→ Most "new" agent frameworks are one of the ladder's loops, renamed. Y saber qué estás corriendo es la mitad del criterio.
● la frontera, este mes · six papers, ten months
THE FRONTIER, THIS MONTH.
01
The Optimizer That Remembers Its Own Training
SePO trains one agent to evolve both your prompt and its own optimizing skill together, keeping past attempts as stepping stones instead of throwing them away.
SCOPE evolves several candidate prompts at once from an agent's own execution traces, routing lessons to quick fixes or long-term memory. Nearly 3x'd success on a hard benchmark.
SPEAR lets the optimizer run actual code to cluster its own errors before rewriting a prompt, with automatic rollback the moment a change makes things worse.
p1 shows most training examples just measure the model's own noise, not whether a better prompt helps — keep only the few examples that truly discriminate, and beat the field with two.
A LangChain build of Promptbreeder — the mutator that mutates itself.
todos verificados hoy en GitHub · star counts as of tonight · graduate to these when the six starter scripts stop being enough
● ACT V empieza esta semana
THREE MOVES. ZERO INFRA.
01
> Self-Refine tonight: generate, critique, revise. Three prompts, any model.esta noche · ~20% average gain in the paper · no infrastructure, just the loop
02
> Version your criterion: put your judgment in a prompt file with a version number and a cache.tu criterio, versionado · bump the version, the cache invalidates itself
03
> Mine your own history: the prompts you keep retyping are templates waiting for a name.mina tu historial · nuestra versión: 6,366 sesiones a 18 plantillas con slots
all six examples, runnable tonight, zero API keys: vcn-latam-metaprompting.vercel.app/examples/ · graduate to DSPy or GEPA when you have a real metric. no antes.
● cierre
El acceso al frontier no debería depender de tu código postal.