Seed Curator

Seed Curator

Manages the full seed lifecycle — from planting raw patterns through harvesting session insights to elevating proven seeds into reusable skills.

Philosophy

Skills are the endpoint; seeds are the evidence path. A good skill is never invented — it's distilled from repeated real experience that earned a pattern status through validated use. Seeds are the intermediate form: a named insight with a confidence score and counter-examples, held in the library until the pattern is proven enough to justify a full SKILL.md with triggers and workflow. Without seed management, patterns get lost between sessions or promoted too early (becoming brittle skills) or too late (staying buried in transcripts forever).

When to Use

• plant — You just encountered an approach that worked in a non-obvious way and want to capture it before the session ends
• harvest — At session close, scan for corrections, wins, or friction that should be captured as candidate seeds
• search — Before starting a task, check whether the seed library has relevant patterns (especially for debugging or architectural decisions)
• elevate — A seed has accumulated high confidence across multiple real uses and is ready to become a SKILL.md
• audit — Periodic check of the seed library for stale references or low-quality entries

Do NOT use elevate immediately after planting a seed — seeds need confidence ≥ 0.8 and 3+ independent applications before elevation.

Workflow

1. MODE

Determine the operation from user intent:

• plant — Extract a new seed from a specific experience or insight
• harvest — Scan a recent session for candidate seeds
• search — Find seeds relevant to a current situation
• elevate — Promote a proven seed to a full SKILL.md
• audit — Check seed library health

2. PLANT (when mode = plant)

Extract a reusable pattern:

• Name: Short, memorable, hyphenated (e.g., compilation-as-contract)
• Source: Where this pattern was observed (session, project, incident)
• Confidence: 0.0-1.0 based on how many times it's been validated
• Tags: 3-5 categorization tags
• Pattern: What the seed captures — the reusable insight
• When to use: Situations where this seed applies
• Example: Concrete instance of the pattern in action
• Counter-example: When this pattern does NOT apply Write as YAML-fronted markdown file.

3. HARVEST (when mode = harvest)

Scan the current or recent session for seed candidates:

• Corrections received from the user (feedback → pattern)
• Approaches that worked unexpectedly well (success → pattern)
• Non-obvious decisions that required investigation (discovery → pattern)
• Friction points that were resolved (obstacle → pattern) Present candidates to user for confirmation before writing files.

4. SEARCH (when mode = search)

Given a situation description, score all seeds by relevance:

• Tag match (exact tag = 3 points, partial = 1)
• Description keyword overlap (each matching word = 1 point)
• Confidence weighting (score × confidence) Return ranked list with seed name, confidence, and one-line relevance explanation.

5. ELEVATE (when mode = elevate)

For seeds meeting promotion criteria:

• Confidence ≥ 0.8
• Applied 3+ times across different contexts
• No contradicting seeds in the library Generate a SKILL.md draft following the skill-creation skill's format:
• YAML frontmatter with name, description, triggers, version
• Structured workflow steps derived from the seed's pattern
• Anti-patterns derived from the seed's counter-examples

6. AUDIT (when mode = audit)

Check all seeds for:

• Stale references (files/functions that no longer exist)
• Low confidence + no recent use (candidates for archival)
• Contradictions between seeds (same situation, opposite advice)
• Missing fields (no counter-example, no tags, no confidence score) Produce audit report with recommendations.

7. OUTPUT

Produce the appropriate artifact based on mode:

• plant → seed file written to memory directory
• harvest → candidate list with user confirmation prompts
• search → ranked seed list
• elevate → SKILL.md draft for review
• audit → health report with action items

Best Practices

• Confirm before writing. In harvest mode, always present candidates to the user before writing any file — they know which patterns are worth keeping.
• Counter-examples are mandatory for plant. A seed without a counter-example is a rule, not a pattern. Every seed should state when it does NOT apply.
• Keep confidence honest. First observation = 0.3–0.4. Second independent validation = 0.6–0.7. Three or more contexts = 0.8+. Never start a seed at 0.8.
• Seeds are not memories. Memories capture facts about projects and preferences. Seeds capture reusable behavioral patterns. Overlap is a signal to choose — don't maintain the same insight in both stores.
• Elevation is a draft, not a deploy. The SKILL.md produced by elevate is a draft for review. The user decides when it goes into the skills directory.

Example

Session ends. User asks to harvest seeds.

The curator scans the session and identifies three candidates:

1. A correction: "always check config before debugging code" — seen twice now, confidence 0.6, tagged debugging, config
2. An approach that worked: using Grep before Read for large files — already exists as a seed at confidence 0.7 (update confidence, don't duplicate)
3. A friction point resolved: PowerShell commit messages need heredoc to avoid variable interpolation — new seed, confidence 0.4

Candidates presented to user. User confirms #1 and #3, declines #2 (already covered). Two seed files written; #2 gets a confidence bump in its existing file.

Later, #1 accumulates two more confirmations across different sessions (confidence now 0.85, 3+ uses). User runs elevate. A SKILL.md draft is produced with triggers, workflow steps, and anti-patterns derived from the seed's counter-example. Draft lands in a review location — not deployed automatically.

Quality Checklist

• Mode determined before any file reads or writes
• In harvest mode: candidates presented before any file is written
• Every new seed includes: name, source, confidence, tags, pattern, when-to-use, example, counter-example
• Confidence value is within the honest range for number of observations (new seed ≤ 0.5)
• No seed was elevated with confidence < 0.8 or fewer than 3 documented uses
• Elevated SKILL.md is labeled as a draft and placed for review, not written directly into the skills directory
• Audit report distinguishes stale/contradicting/missing-fields as separate categories
• No duplicate seeds created — search the library before planting

Related Skills

• memory-curator — manages the parallel memory store; coordinate when an insight could live in either system
• convergence-checker — seed audit can be triggered as part of a convergence run