Aprende

aprende — Learn from this conversation / Aprende de esta conversación

EN — A skill that turns finished conversations into durable, structured learnings: memories, anti-patterns (Reflexion-style), skill stubs, and project-doc updates. Confirmation-first. Never auto-writes.

ES — Un skill que convierte conversaciones terminadas en aprendizajes durables y estructurados: memorias, anti-patrones (estilo Reflexion), stubs de skills, y actualizaciones a project-docs. Confirmación primero. Nunca escribe automáticamente.

1. Purpose / Propósito

EN. Coding agents repeat mistakes across sessions because the corrections a user makes in one conversation evaporate when the session ends. aprende fixes that. When invoked, it reviews the current conversation, identifies what is worth preserving across four well-defined categories, and writes those learnings to the right files — in the format the user's existing memory system already reads — only after the user picks the items to keep.

Guiding principle: a false positive locked into memory is worse than repeating a correction three times. Be liberal at surfacing candidates, strict at confirming them, and conservative at writing them. Prefer false negatives.

ES. Los agentes de código repiten errores entre sesiones porque las correcciones que el usuario hace en una conversación se evaporan cuando esa sesión termina. aprende arregla eso. Al activarse, revisa la conversación actual, identifica qué vale la pena preservar a través de cuatro categorías bien definidas, y escribe esos aprendizajes en los archivos correctos — en el formato que ya usa el sistema de memoria del usuario — solo después de que el usuario elija qué guardar.

Principio rector: un falso positivo cristalizado en la memoria es peor que repetir una corrección tres veces. Sé liberal al proponer candidatos, estricto al confirmarlos, y conservador al escribirlos. Prefiere los falsos negativos.

2. When to invoke / Cuándo activarse

EN — Trigger on any of:

• The user types /aprende or /learn (the English alias).
• The user types phrases like: "reflect on this", "save what we learned", "remember this for next time", "let's capture that", "make a note of this", "don't forget X next session".
• The user has corrected the agent on the same thing twice in a session — proactively suggest running /aprende (one line, in the assistant turn); do not run it without permission.
• The session is wrapping up (the user says "we're done", "good, ship it", "commit and push") and signals have been captured by the optional PostToolUse hook (.aprende-signals.md exists with content).

ES — Activa con cualquiera de estos:

• El usuario escribe /aprende o /learn (alias en inglés).
• El usuario escribe frases como: "reflexiona sobre esto", "guarda lo que aprendimos", "recuérdalo para la próxima", "captura esto", "que no se nos olvide X".
• El usuario corrigió al agente sobre lo mismo dos veces en la sesión — proactivamente sugiere correr /aprende (una línea, en el turno del asistente); no lo ejecutes sin permiso.
• La sesión está cerrando ("ya terminamos", "listo, súbelo", "commit y push") y el hook PostToolUse capturó señales (existe .aprende-signals.md con contenido).

3. The four categories / Las cuatro categorías

One example each / Un ejemplo de cada:

• memory — "User prefers pnpm over npm in all JS projects." / "El usuario prefiere pnpm sobre npm en todos los proyectos JS."
• lesson — "Assumed localStorage works inside a WebView purchase flow; it does not — caused silent purchase failures. Use IndexedDB instead." / "Asumí que localStorage funciona dentro de un flujo de compra en WebView; no funciona — causó fallos de compra silenciosos. Usa IndexedDB."
• skill — /verify-rls — a 6-step Supabase RLS verification workflow that came up three times today. / un workflow de 6 pasos para verificar RLS en Supabase que apareció tres veces hoy.
• project-doc — "Run tests with pnpm run test:unit, not npm test. The latter triggers the e2e suite which needs a running DB." / "Corre tests con pnpm run test:unit, no npm test. El último dispara la suite e2e que necesita una DB activa."

4. Workflow (5 passes) / Workflow (5 pases)

Run these in order. Do not skip. Each pass has a hard output shape.

Ejecuta en orden. No saltes pasos. Cada pase tiene una forma de output estricta.

Pass A — Scan / Escaneo

EN. Read the conversation transcript silently. Look for the patterns in references/signal-patterns.md (multi-language). The high-signal classes:

1. Explicit user corrections (high confidence) — "no, do X instead", "actually that's wrong", "always do Y", "never do Z", "stop doing W", "the convention here is", and their Spanish counterparts in references/signal-patterns.md.
2. Error → fix sequences (high) — a tool call failed, a different call succeeded shortly after. The delta between them is the lesson.
3. Repeated attempts (medium) — three or more Edits to the same file in a row, three or more Bash retries of the same command with variations.
4. User explanations of why (high — Reflexion gold) — "the reason is", "because", "the gotcha is", "porque", "el detalle es", "la trampa es".
5. Workflows the user described step-by-step (medium, candidate for a skill stub) — four or more sequential tool calls that the user explicitly walked through.

If ~/.claude/projects/<current-project-slug>/.aprende-signals.md exists and is non-empty, prepend its contents to the scratch list — those are signals captured by the PostToolUse hook earlier in this session.

ES. Lee la transcripción de la conversación en silencio. Busca los patrones de references/signal-patterns.md (multi-idioma). Las clases de alta señal:

1. Correcciones explícitas del usuario (high) — "no, mejor X", "eso está mal", "siempre Y", "nunca Z", "deja de hacer W", "la convención aquí es".
2. Secuencias error → arreglo (high) — una llamada falló, otra distinta funcionó después. El delta es la lección.
3. Intentos repetidos (medium) — tres o más Edits al mismo archivo en fila, tres o más reintentos de Bash con variaciones.
4. Explicaciones de por qué del usuario (high — oro de Reflexion) — "porque", "el detalle es", "la trampa es", "the reason is".
5. Workflows que el usuario explicó paso a paso (medium, candidato a skill stub) — cuatro o más llamadas secuenciales que el usuario detalló.

Si ~/.claude/projects/<slug-del-proyecto>/.aprende-signals.md existe y no está vacío, anteponé su contenido a la lista — son señales capturadas por el hook PostToolUse antes en esta sesión.

Pass B — Generate / Generar

For each signal, draft a candidate item with this internal shape:

Para cada señal, draftea un candidato con esta forma interna:

{
  "category": "memory" | "lesson" | "skill" | "project-doc",
  "title": "<short imperative title, ≤ 60 chars>",
  "rationale": "<one sentence; what this captures and why>",
  "confidence": "high" | "medium" | "low",
  "source_excerpt": "<short verbatim quote or tool-result line, optional>"
}

Confidence derivation (mandatory):

• high — explicit user feedback in the conversation, OR error→fix with a user explanation, OR the same correction repeated by the user.
• medium — inferred pattern, single user comment without elaboration, error→fix without explanation.
• low — guess, single observation the model thinks might matter.

Be liberal at this pass. It's cheaper to filter at confirmation than to miss a real learning. Cap the candidate list at 15 items per run; if you find more, keep the 15 highest-confidence and tell the user to re-run.

Sé liberal en este pase. Es más barato filtrar en la confirmación que perder un aprendizaje real. Limita la lista a 15 items por corrida; si encuentras más, conserva los 15 de mayor confidence y dile al usuario que vuelva a correr.

Pass C — Dedup / Deduplicar

Before showing the list, check overlaps without dropping silently:

Antes de mostrar la lista, chequea overlaps sin dropear en silencio:

1. Read ~/.claude/projects/<slug>/memory/MEMORY.md if it exists. For each candidate, do a title + topic similarity check against existing bullets.
2. For skill candidates, ls ~/.claude/skills/ and ls ./.claude/skills/ (if the project-local folder exists).
3. For project-doc candidates, grep -i <keyword> against ./CLAUDE.md and ./AGENTS.md (if either exists).
4. If a candidate looks like a near-duplicate of an existing entry, do not drop it. Annotate it [overlaps with: <existing-name-or-path>] in the numbered list so the user can decide whether to skip, edit, or replace.

Pass D — Confirm / Confirmar

Output exactly this template (substitute values, keep the structure):

Imprime exactamente esta plantilla (sustituye valores, mantén la estructura):

Found N candidate learnings from this conversation.
Encontré N aprendizajes candidatos en esta conversación.

Reply with: numbers (e.g. 1,3,5), ranges (1-4), "all", "none", or
"edit N: <new wording>" / "drop low" / "skip N".

Responde con: números (ej. 1,3,5), rangos (1-4), "all", "none", o
"edit N: <nueva descripción>" / "drop low" / "skip N".

 1. [memory]      <title> — <rationale>  (<confidence>)
 2. [lesson]      <title> — <rationale>  (<confidence>)  [overlaps with: <name>]
 3. [skill]       <title> — <rationale>  (<confidence>)
 4. [project-doc] <title> — <rationale>  (<confidence>)
 ...

>

Then wait. Do not call Write, Edit, or any file-mutating tool until the user replies. This is the most important rule in this skill.

Después espera. No llames Write, Edit, ni ninguna herramienta que modifique archivos hasta que el usuario responda. Esta es la regla más importante del skill.

Accepted inputs / Inputs aceptados:

• Comma-separated numbers: 1,3,5
• Ranges: 1-4 or 1-3,7,9-11
• all — accept everything as shown
• none — accept nothing, exit cleanly
• edit 5: <new wording> — replace the candidate's rationale/title before saving
• skip 3 — accept all except the listed numbers
• drop low — filter out everything below the named confidence band (drop low, drop medium)

If the input is ambiguous, ask once. Do not guess.

Si el input es ambiguo, pregunta una vez. No adivines.

Pass E — Execute / Ejecutar

For each confirmed candidate, run the category-specific write procedure:

Para cada candidato confirmado, ejecuta el procedimiento de escritura por categoría:

memory

1. Determine <slug> for the current project: read pwd, then convert /Users/soyenrique/Desktop/aprende-skill to -Users-soyenrique-Desktop-aprende-skill (replace / with -, drop leading dash, keep dashes). The folder is ~/.claude/projects/<slug>/memory/. Create it if missing (mkdir -p).
2. Pick a filename: kebab-case slug from the title, prefixed by metadata.type (e.g. feedback_pnpm-over-npm.md, project_supabase-rls-policy.md). If a file with that name exists, append -2, -3, ... Never overwrite.
3. Write the file with the frontmatter and body spec from references/memory-format.md.
4. Append a one-line entry to ~/.claude/projects/<slug>/memory/MEMORY.md: - [<Title>](<filename>.md) — <one-sentence hook>.

lesson

1. Same folder as memory. Filename starts with lesson_ (e.g. lesson_webview-assumed-localstorage.md).
2. Use the strict frontmatter and four mandatory subheads from references/lesson-format.md:
   • **What happened / Qué pasó:**
   • **Why it happened / Por qué pasó:**
   • **How to avoid / Cómo evitar:**
   • **Detection signal / Señal de detección:**
3. Append to MEMORY.md under a ## Lessons / Lecciones subsection (create the subsection if missing).

skill (stub)

1. Decide scope: user-global (~/.claude/skills/<slug>/) by default; project-local (./.claude/skills/<slug>/) only if the workflow references repo-specific paths or tools.
2. Create the folder. Write SKILL.md as a stub, not a full skill — use references/skill-stub-template.md. Fill in only what the conversation shows; mark the rest [STUB — fill in: ...].
3. Print a one-line reminder: "Skill stub written at <path>. Run the skill-creator plugin (or edit manually) to flesh it out before publishing."

project-doc

1. Decide scope: project-level (./CLAUDE.md and ./AGENTS.md — dual-write) by default; global (~/.claude/CLAUDE.md and ~/.claude/AGENTS.md) only if the candidate is repo-agnostic.
2. Extra confirmation for global writes: prompt the user one more time before touching ~/.claude/CLAUDE.md or ~/.claude/AGENTS.md. Global docs affect every session on this machine.
3. Read the target file. Find the best existing section (e.g. ## Commands, ## Conventions). Append under it.
4. If no good section exists, create a ## Learnings (added by /aprende) section at the end of the file.
5. Mark each appended line with an HTML comment for later audit: <!-- /aprende YYYY-MM-DD -->.
6. Dual-write to both CLAUDE.md and AGENTS.md. If one doesn't exist, create it with the entry alone. Content is identical between the two.

After all writes, emit one confirmation line per write:

Después de cada escritura, imprime una línea de confirmación:

✓ memory   → ~/.claude/projects/.../feedback_pnpm-over-npm.md  (+MEMORY.md)
✓ lesson   → ~/.claude/projects/.../lesson_webview-localstorage.md  (+MEMORY.md)
✓ skill    → ~/.claude/skills/verify-rls/SKILL.md  (stub — finish before use)
✓ project  → ./CLAUDE.md + ./AGENTS.md  (## Commands)

End with a one-paragraph summary of what was learned and a pointer to /aprende --review for later pruning.

Termina con un párrafo resumen de lo aprendido y un puntero a /aprende --review para podar después.

5. Output formats / Formatos de salida

The strict specs live in separate references so this file stays scannable:

Las especificaciones estrictas están en references separadas para que este archivo siga siendo fácil de escanear:

• references/memory-format.md — frontmatter + body for type: user|project|reference|feedback.
• references/lesson-format.md — frontmatter + 4 mandatory subheads (Reflexion).
• references/skill-stub-template.md — bare SKILL.md scaffold.
• references/signal-patterns.md — multi-language correction phrases and detection heuristics.
• references/review-workflow.md — semantics of /aprende --review.

The lesson format is the most important; here it is inline as well:

El formato lesson es el más importante; aquí va inline también:

---
name: lesson_<short-slug>
description: One-sentence summary of the mistake and its corrective rule.
metadata:
  type: lesson
  confidence: high | medium | low
  status: active   # active | retired | superseded-by-<name>
  createdAt: YYYY-MM-DD
  lastValidated: YYYY-MM-DD
  originSessionId: <session-id or omit>
---

**What happened / Qué pasó:** <factual paragraph, one to three sentences>

**Why it happened / Por qué pasó:** <root cause — not the symptom>

**How to avoid / Cómo evitar:** <concrete rule the next agent should follow>

**Detection signal / Señal de detección:** <how a future run knows it's about to repeat this>

Related: [[other-memory-name]]

6. --review mode / modo --review

When invoked as /aprende --review:

Cuando se invoca como /aprende --review:

1. Read every metadata.type: lesson file in the current project's memory folder, sorted by lastValidated ascending (oldest first).
2. For each, show: title, confidence, status, lastValidated, and the one-sentence description.
3. Prompt per-lesson: Still valid? [y/n/edit/skip]
   • y — bump lastValidated to today's date.
   • n — set status: retired. Do not delete the file.
   • edit — show the full body, let the user paste a new version inline, then overwrite (and bump lastValidated).
   • skip — leave untouched.
4. At the end, print a summary: N validated, M retired, K edited, L skipped.

Lessons with status: retired are skipped by future /aprende runs when reading memory at session start.

Las lecciones con status: retired se ignoran en futuras corridas de /aprende cuando lee memoria al inicio de la sesión.

Full procedure: references/review-workflow.md.

7. --portable mode / modo --portable

When invoked as /aprende --portable:

Cuando se invoca como /aprende --portable:

1. Run the standard 5-pass flow as usual.
2. After the normal writes for lesson candidates, also mirror each lesson file to ./.aprende/lessons/<filename>.md in the project root.
3. Update ./.aprende/index.md with a one-line entry per lesson — Codex and other tools can read this from AGENTS.md instructions.
4. The first time --portable is used in a project, append this line to ./AGENTS.md (creating it if missing):

   Before any non-trivial change, read `.aprende/index.md` for prior lessons.
   Antes de cualquier cambio no trivial, lee `.aprende/index.md` para ver lecciones previas.
   

This is off by default. Plain /aprende does not write ./.aprende/.

Esto está apagado por defecto. Un /aprende normal no escribe ./.aprende/.

8. Examples / Ejemplos

Example 1 — explicit correction / corrección explícita

User said earlier in the session: "No, in this repo we use spaces, not tabs. Always 2 spaces."

User invokes: /aprende

Expected output:

Found 1 candidate learning from this conversation.

 1. [project-doc] Indent with 2 spaces, not tabs — user stated repo convention explicitly  (high)

>

User: 1

Execution:

✓ project  → ./CLAUDE.md + ./AGENTS.md  (## Conventions)

./CLAUDE.md now contains, under ## Conventions:

- Indent with 2 spaces, not tabs. <!-- /aprende 2026-05-11 -->

./AGENTS.md mirrors the same line.

Example 2 — error→fix lesson / lección error→arreglo

Session contained: Bash call npm test failed with "no test command"; user said "oh, this project uses pnpm"; Bash call pnpm test succeeded.

User invokes: /aprende

Expected output:

Found 2 candidate learnings from this conversation.

 1. [lesson]      Don't default to npm in this repo — pnpm is the package manager  (high)
 2. [project-doc] Test command is `pnpm test`, not `npm test`  (high)

>

User: all

Execution: writes lesson file + project-doc append, MEMORY.md updated.

9. Safety / Guardrails

EN — Hard rules:

1. Never write before confirmation. No Write, Edit, or destructive Bash call may run between Pass D's output and the user's reply. Repeat: never.
2. Never overwrite an existing file. Always append, or create with a unique name (-2, -3 suffix).
3. Never delete a lesson. Use metadata.status: retired.
4. Extra confirmation for global writes. Touching ~/.claude/CLAUDE.md or ~/.claude/AGENTS.md requires a second user yes.
5. Dedup annotates, does not drop. Let the user decide on overlaps.
6. Cap at 15 candidates per run. Forces focus.
7. Confidence labels are mandatory. No item ships without (high|medium|low).
8. Hooks never write learnings. capture-signal.sh only appends signal records; stop-suggest.sh only emits text. Learnings are written exclusively by the user-confirmed /aprende flow.
9. If unsure, prefer the false-negative. Tell the user "I'm not confident this is a learning — re-run /aprende later if it repeats."

ES — Reglas duras:

1. Nunca escribas antes de la confirmación. Ninguna llamada Write, Edit, o Bash destructiva puede correr entre el output del Pase D y la respuesta del usuario. Repito: nunca.
2. Nunca sobrescribas un archivo existente. Siempre appendea, o crea con un nombre único (-2, -3).
3. Nunca elimines una lección. Usa metadata.status: retired.
4. Confirmación extra para escrituras globales. Tocar ~/.claude/CLAUDE.md o ~/.claude/AGENTS.md requiere un segundo sí del usuario.
5. El dedup anota, no dropea. Que el usuario decida sobre los overlaps.
6. Tope de 15 candidatos por corrida. Fuerza foco.
7. Las etiquetas de confidence son obligatorias. Ningún item sale sin (high|medium|low).
8. Los hooks nunca escriben aprendizajes. capture-signal.sh solo appendea registros de señales; stop-suggest.sh solo emite texto. Los aprendizajes se escriben exclusivamente por el flujo /aprende confirmado.
9. Si tienes dudas, prefiere el falso negativo. Dile al usuario "no estoy seguro de que esto sea un aprendizaje — vuelve a correr /aprende si se repite."

10. Project-slug derivation / Derivación del slug del proyecto

The memory folder is ~/.claude/projects/<slug>/memory/. To get <slug>:

La carpeta de memoria es ~/.claude/projects/<slug>/memory/. Para obtener <slug>:

1. Read pwd (the current working directory absolute path).
2. Replace every character that is not a letter or digit (A–Z, a–z, 0–9) with -. This is Claude Code's own project-folder convention, so the slug matches for every path — including Windows paths like D:/proyectos_python/... and any path containing _, ., :, \ etc. (Replacing only / would mis-slug and read from the wrong folder.)
3. If the result starts with -, that leading dash stays (matches the existing convention — ~/.claude/projects/-Users-soyenrique-Desktop-aprende-skill/).

Example (macOS/Linux): pwd = /Users/soyenrique/Desktop/myproject → slug = -Users-soyenrique-Desktop-myproject → folder = ~/.claude/projects/-Users-soyenrique-Desktop-myproject/memory/

Example (Windows): pwd = D:/proyectos_python/myproject → slug = D--proyectos-python-myproject → folder = ~/.claude/projects/D--proyectos-python-myproject/memory/

If the folder does not exist, create it with mkdir -p. Also create MEMORY.md with a single-line header (# Memory Index) if it's missing.

Si la carpeta no existe, créala con mkdir -p. También crea MEMORY.md con un header de una línea (# Memory Index) si no existe.

11. Integration with the optional hooks / Integración con los hooks opcionales

If the plugin's hooks are enabled (they ship ON by default), this skill benefits from them but never depends on them:

Si los hooks del plugin están activos (vienen ON por defecto), este skill se beneficia de ellos pero nunca depende de ellos:

• PostToolUse writes per-session signal records to ~/.claude/projects/<slug>/.aprende-signals.md. Pass A reads this file if it exists and prepends its content to the scratch list.
• Stop reads the same file at session end and emits an additionalContext reminder if signals exist and /aprende was not run.

After a successful /aprende run (one that produced at least one confirmed write), delete .aprende-signals.md to mark the signals as consumed.

Después de una corrida exitosa de /aprende (que produjo al menos una escritura confirmada), borra .aprende-signals.md para marcar las señales como consumidas.

12. Cross-tool note / Nota cross-tool

This skill is designed to work in both Claude Code and Codex:

Este skill está diseñado para funcionar tanto en Claude Code como Codex:

• project-doc writes always dual-write to CLAUDE.md and AGENTS.md. Codex reads AGENTS.md; Claude Code reads CLAUDE.md. Same content in both. No symlinks, no imports.
• For memory and lesson, use /aprende --portable to mirror to ./.aprende/ in the project root. Codex can then read .aprende/index.md per the line auto-added to AGENTS.md.
• The skill itself (skills/aprende/SKILL.md) is plain Markdown with YAML frontmatter — readable by any agent that supports the agentskills.io convention.

13. Quick reference / Referencia rápida

Made with discipline against false positives. / Hecho con disciplina contra los falsos positivos.