Back to Skills

learn

pjt222
Updated Yesterday
3 views
17
2
17
View on GitHub
Metaaidesign

About

The `learn` skill enables Claude to systematically acquire knowledge from unfamiliar domains using a structured reasoning process of surveying, hypothesizing, probing, and verifying. It is designed for situations like exploring an unfamiliar codebase, investigating topics beyond simple recall, or resolving conflicting information into a coherent model. Key capabilities include deliberate model-building with feedback loops and the use of tools like Read, Grep, and WebSearch for exploration.

Quick Install

Claude Code

Recommended
Primary
npx skills add pjt222/agent-almanac -a claude-code
Plugin CommandAlternative
/plugin add https://github.com/pjt222/agent-almanac
Git CloneAlternative
git clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/learn

Copy and paste this command in Claude Code to install this skill

Documentation

Learn

Structured knowledge acquisition session — survey unfamiliar, build initial models, test via deliberate exploration, integrate into coherent understanding, consolidate for durable retrieval.

Use When

  • Unfamiliar codebase / framework / domain, no prior ctx
  • User asks topic outside working knowledge, answer needs investigation not recall
  • Conflicting sources / patterns → coherent mental model from scratch
  • After remote-viewing surfaces intuitive leads → systematic validation
  • Prep to teach — must understand deeply enough to explain

In

  • Req: Learning target — topic, codebase area, API, concept, tech
  • Opt: Scope boundary — surface survey vs deep expertise
  • Opt: User's purpose — why this matters (prioritization)
  • Opt: Known starting points — files, docs, concepts familiar

Do

Step 1: Survey — Map Territory

Before understanding anything, map landscape → ID what exists.

Learning Modality Selection:
┌──────────────────┬──────────────────────────┬──────────────────────────┐
│ Territory Type   │ Primary Modality         │ Tool Pattern             │
├──────────────────┼──────────────────────────┼──────────────────────────┤
│ Codebase         │ Structural mapping —     │ Glob for file tree,      │
│                  │ find entry points, core  │ Grep for exports/imports,│
│                  │ modules, boundaries      │ Read for key files       │
├──────────────────┼──────────────────────────┼──────────────────────────┤
│ API / Library    │ Interface mapping —      │ WebFetch for docs,       │
│                  │ find public surface,     │ Read for examples,       │
│                  │ types, configuration     │ Grep for usage patterns  │
├──────────────────┼──────────────────────────┼──────────────────────────┤
│ Domain concept   │ Ontology mapping —       │ WebSearch for overviews,  │
│                  │ find core terms,         │ WebFetch for definitions,│
│                  │ relationships, debates   │ Read for local notes     │
├──────────────────┼──────────────────────────┼──────────────────────────┤
│ User's context   │ Conversational mapping   │ Read conversation,       │
│                  │ — find stated goals,     │ Read MEMORY.md,          │
│                  │ preferences, constraints │ Read CLAUDE.md           │
└──────────────────┴──────────────────────────┴──────────────────────────┘
  1. ID territory type + select primary modality
  2. Broad scan — not reading deeply, ID landmarks (key files, entry points, core concepts)
  3. Note boundaries: in scope, adjacent, out of scope
  4. ID gaps: important-looking but opaque from surface
  5. Rough map: list major components + apparent relationships

→ Skeletal map w/ 5-15 landmarks. Sense of clear surface vs deeper investigation needed. No understanding yet — just map.

If err: Territory too large → narrow scope. Ask: "Min to understand → serve user's purpose?" No clear entry → start from output (what produces?) + trace backward.

Step 2: Hypothesize — Initial Models

From survey → construct hypotheses.

  1. Formulate 2-3 hypotheses about structure / behavior
  2. State clearly: "I believe X because I observed Y"
  3. Per hypothesis → what evidence confirms, what refutes
  4. Rank by confidence: most supported, shakiest
  5. ID highest-value to test first (unlocks most understanding if confirmed)

→ Concrete falsifiable hypotheses — not vague impressions. Each has test. Collectively cover most important aspects.

If err: No hypotheses → survey too shallow → back to Step 1, read 2-3 landmarks in depth. All equally uncertain → simplest (Occam's) + build from there.

Step 3: Explore — Probe + Test

Systematically test each hypothesis via targeted investigation.

  1. Select highest-priority
  2. Design minimal probe: smallest investigation confirming/refuting
  3. Execute (read file, search pattern, test assumption)
  4. Record: confirmed, refuted, modified
  5. If refuted → update hypothesis w/ new evidence
  6. If confirmed → probe deeper: holds at edges or only center?
  7. Next hypothesis, repeat

→ ≥1 hypothesis tested to conclusion. Model taking shape — some confirmed, some revised. Surprises noted as valuable data.

If err: Probes consistently ambiguous → testing wrong things. Step back: "What would an expert consider most important fact?" Probe for that.

Step 4: Integrate — Mental Model

Synthesize findings → coherent model connecting pieces.

  1. Review confirmed hypotheses + revised models
  2. ID central organizing principle: "spine" everything connects to
  3. Map relationships: which components depend on which? What flows where?
  4. ID surprising findings — often deepest insight
  5. Look for patterns repeating across territory
  6. Build model predicting behavior: "Given input X, expect Y because Z"

→ Coherent model explaining structure + predicting behavior. Expressible in 3-5 sentences, specific claims not vague.

If err: Pieces don't integrate → fundamental misunderstanding in earlier hypothesis. ID piece that doesn't fit → re-test. Or territory genuinely incoherent (poorly designed exist) → note as finding rather than forcing.

Step 5: Verify — Challenge Understanding

Test model via predictions + check.

  1. Use model → 3 specific predictions
  2. Test each via investigation (not assuming true)
  3. Per confirmed → confidence increases
  4. Per refuted → ID where model wrong + correct
  5. Edge cases: hold at boundaries or break?
  6. Ask: "What would surprise me?" → check if possible

→ Model survives ≥2 of 3 prediction tests. Failures understood, model corrected. Now has confirmed strengths + known limitations.

If err: Most predictions fail → model fundamental flaw. Valuable info — territory works differently than expected. Return Step 2 w/ new evidence, rebuild. 2nd attempt much faster (wrong models eliminated).

Step 6: Consolidate — Store for Retrieval

Capture learning in form supporting future retrieval + application.

  1. Summarize model in 3-5 sentences
  2. Note key landmarks — 3-5 most important to remember
  3. Record counterintuitive findings (might be forgotten)
  4. ID related topics this connects to
  5. Durable learning (needed across sessions) → update MEMORY.md
  6. Session-specific → note as ctx for current conv
  7. State what remains unknown — honest gaps > false confidence

→ Concise retrievable summary capturing essential understanding. Future references start from summary, not re-learning.

If err: Learning resists summarization → not fully integrated → return Step 4. Learning too obvious to store → what feels obvious now may not in fresh ctx. Store non-obvious.

Check

  • Survey before deep investigation (map before dive)
  • Hypotheses explicit + tested, not assumed
  • ≥1 hypothesis revised based on evidence (= genuine learning)
  • Model makes specific testable predictions
  • Known unknowns ID'd alongside known knowns
  • Consolidated summary concise for future retrieval

Traps

  • Skip survey: Diving into detail before landscape → wastes time on unimportant + misses big picture.
  • Unfalsifiable hypotheses: "This is probably complex" can't be tested. "This module handles auth because it imports crypto" can.
  • Confirmation bias: Seeking only supporting evidence, ignoring contradictions.
  • Premature consolidation: Store model before tested → confidently wrong future predictions.
  • Perfectionism: Learn everything before applying anything. Iterative — use partial, then refine.
  • Learning w/o purpose: Knowledge w/o application → unfocused shallow understanding.

  • learn-guidance — human-guidance variant → coach person thru structured learning
  • teach — knowledge transfer calibrated to learner; builds on model constructed here
  • remote-viewing — intuitive exploration surfaces leads for systematic learning to validate
  • meditate — clear prior ctx noise before new learning territory
  • observe — sustained neutral pattern recognition feeding learning w/ raw data

GitHub Repository

pjt222/agent-almanac
Path: i18n/caveman-ultra/skills/learn
0
agentsagentskillsai-assisted-developmentclaude-codeskillsteams

Related Skills

content-collections

Meta

This skill provides a production-tested setup for Content Collections, a TypeScript-first tool that transforms Markdown/MDX files into type-safe data collections with Zod validation. Use it when building blogs, documentation sites, or content-heavy Vite + React applications to ensure type safety and automatic content validation. It covers everything from Vite plugin configuration and MDX compilation to deployment optimization and schema validation.

View skill

polymarket

Meta

This skill enables developers to build applications with the Polymarket prediction markets platform, including API integration for trading and market data. It also provides real-time data streaming via WebSocket to monitor live trades and market activity. Use it for implementing trading strategies or creating tools that process live market updates.

View skill

creating-opencode-plugins

Meta

This skill helps developers create OpenCode plugins that hook into 25+ event types like commands, files, and LSP operations. It provides the plugin structure, event API specifications, and implementation patterns for JavaScript/TypeScript modules. Use it when you need to intercept, monitor, or extend the OpenCode AI assistant's lifecycle with custom event-driven logic.

View skill

sglang

Meta

SGLang is a high-performance LLM serving framework that specializes in fast, structured generation for JSON, regex, and agentic workflows using its RadixAttention prefix caching. It delivers significantly faster inference, especially for tasks with repeated prefixes, making it ideal for complex, structured outputs and multi-turn conversations. Choose SGLang over alternatives like vLLM when you need constrained decoding or are building applications with extensive prefix sharing.

View skill