Skip to main content

​
Nurture-First Development (NFD)

A proposed methodology for building domain-expert AI agents by growing them through sustained conversation, rather than building them upfront. Development and deployment run concurrently: the agent starts as minimal scaffolding and accumulates expertise from daily use, which is periodically consolidated into reusable knowledge assets. Framework-agnostic — instantiable on any harness with persistent memory + on-demand skills (the source paper cites Claude Code and “OpenClaw”; the pattern applies equally to Codex, Copilot, Pi, Gemini setups). Source: Nurture First Agent Development (Zhang 2026) — a position paper, so the framework is proposed, not empirically validated.

​
Three paradigms of agent development

Code-first and prompt-first both assume build → then → deploy. NFD’s claim: expertise is tacit, personal, and evolving, so any static upfront encoding starts decaying immediately. NFD dissolves the build/deploy boundary. Its defining move is who develops the agent — the domain expert, through daily dialogue, not an engineer. (Contrast the pure code-first pole in The Minimal Coding Agent: LLM + Loop + Tools, where an agent is a fixed loop + tools you write once.)

​
Three-Layer Cognitive Architecture

Organize the agent’s knowledge by volatility × personalization:
  • Constitutional — identity, principles, rules; loaded every session; low volatility; hold indices and pointers, not detail; keep it small (budget ~10–15% of context). (Cross-agent forms: MEMORY.md/AGENTS.md/SOUL.md/USER.md, CLAUDE.md.)
  • Skill — modular, single-responsibility task capabilities loaded on demand; medium volatility; the home for crystallized knowledge; skills coordinate through shared memory files, not direct invocation. (Cross-agent: SKILL.md + references/ + scripts/.) See Agent Skills.
  • Experiential — dated logs, case memories, error patterns from use; semantic-searched; high volatility; append-only. (Cross-agent: memory/YYYY-MM-DD.md.) Overlaps Memory Bank Pattern and Agentic Memory Tool.
Two flows: grounding (down — principles/skills interpret new experience) and crystallization (up — experience consolidated into skills/constitution).

​
The Knowledge Crystallization Cycle (the engine)

The mechanism that turns fragmented conversational knowledge into structured assets — an operationalization of Nonaka–Takeuchi externalization (tacit→explicit). An ascending spiral of four phases:
  1. Conversational Immersion — expertise transfers implicitly through operational dialogue; the agent captures the reasoning, not just conclusions.
  2. Experiential Accumulation — every interaction is logged and tagged. Six categories: operational records, reasoning traces, pattern observations, error records, contextual annotations, insight fragments (tags like [DECISION]/[INSIGHT]/[ERROR] make later extraction cheap).
  3. Deliberate Crystallization — a periodic, human-in-the-loop batch job: extract patterns → structure → de-contextualize (generalize) → validate against the full corpus → integrate with version tracking.
  4. Grounded Application — crystallized patterns re-enter service as hypotheses, tested against new experience; contradictions trigger re-crystallization.
Key safeguard: value is monotonic across cycles only because a human validates which patterns get promoted. Automated, unreviewed crystallization loses that guarantee — and the source names full self-directed crystallization as the paradigm’s main open problem (the “crystallization bottleneck”). This is the same human-approval discipline as Preference Feedback Loop and the mistakes→rules distillation this wiki already runs.

​
Operational patterns

  • Dual-Workspace — a surgical workspace (full filesystem access; scaffolding, bulk migration, crystallization, refactoring — treats the knowledge base as a data structure) separate from a nurturing workspace (the runtime conversational channel where immersion/accumulation happen). Both share file state. Maps cleanly onto: agentic coding tool (surgical) + conversational assistant (nurturing).
  • Spiral Development — Bootstrap (aim for bootability, not completeness) → Nurture → Crystallization Checkpoint → repeat, with each revolution raising the baseline. Triggers: scheduled, threshold, or event.

​
Why it matters here

NFD is essentially the theory behind this wiki: a Constitutional layer (CLAUDE.md, MEMORY.md), a Skill layer (wiki/ pages + skills), and an Experiential layer (mistakes/, log.md, dated captures), with synthesize-mistakes / ingest as crystallization operations promoting fragments into durable rules. It gives that practice a name and a formal shape. Closely related: Compounding Knowledge Base (making each unit of work make later units easier) and Context Compression Strategies (curating the experiential layer to preserve signal).

​
Caveats

The backing paper is conceptual with a single-user, no-control case study — adopt the structure (layering, tagged experiential logs, human-validated crystallization checkpoints) as a design vocabulary, not as evidence of measured performance. Its own limitations apply: cold-start low value, bias absorption (the agent can crystallize a bad habit as readily as a good one), no objective quality metric for nurtured knowledge, and org-scale sharing unsolved.