SkillMind Learner

Role

Turn project-local Peec outputs into cross-project patterns. Each run does two things:

1. Write — extract 1–3 patterns from a just-produced artifact (decision, brief, zone map, outreach log, learnings.json) and store them in SkillMind with tags so they can be retrieved later.
2. Read — on request, recall patterns matching a project / skill / gap type and hand them back as priors for the next orchestrator cycle.

This is the memory layer beneath peec-report: that skill persists learnings for the project, this skill promotes them across projects.

Input

For write mode:

• project_id — Peec project the artifact came from
• source_skill — which skill produced the artifact (peec-agent, peec-cluster, peec-outreach, peec-content-intel, peec-report)
• artifact_path or artifact_content — the file or inline content to extract from
• optional max_patterns — default 3

For read mode:

• query — what the caller wants to recall (e.g. "editorial outreach DACH high citation rate")
• optional project_id — narrow to patterns originally written for this project
• optional source_skill — narrow to patterns originally written by this skill
• optional k — default 5

Output

Write mode: JSON list of {pattern_id, title, tags, summary} for each persisted pattern, plus a one-line confirmation ("added 3 patterns · skipped 1 dupe").

Read mode: ranked list of {pattern_id, title, summary, provenance: {project_id, source_skill, date}, score}. Empty list is a valid result — say so plainly.

Neither mode produces dashboards.

When to use

Write:

• Right after peec-report emits learnings.json
• After a peec-outreach batch closes (week-end ritual)
• After peec-cluster ships a zone map (zones become reusable taxonomy patterns)
• After peec-agent logs a decision whose 4-week metric came in (attribution is known)
• After peec-content-intel ships a brief that later won its prompt (write the retrospective pattern, not the brief itself)

Read:

• At the top of peec-agent Phase 1 (state read) — recall patterns tagged with the current gap type
• At the start of peec-outreach — recall domain-class patterns with high historical citation gain
• At the start of peec-cluster — recall zone-shape patterns that worked in adjacent projects

Do not use when:

• The artifact is <24h old and no outcome is measured yet (you'd persist speculation, not a pattern)
• The artifact is a raw data dump (list_chats output) — needs to be interpreted first
• SkillMind MCP is unavailable — fall back to appending a line to <project>/growth_loop/patterns.md and flag the skip in the output

Pipeline — write mode

1. Read artifact

Read(artifact_path)
# or accept inline artifact_content

Supported artifact shapes:

• decisions_log.md entry (single decision block)
• learnings.json (winners / losers / surprises)
• brief.md with a later-known outcome (prompt visibility moved from X → Y)
• outreach_log.md row with status=citation_live and a measured lift
• zones.md with ≥4 weeks of tag-level visibility data

2. Extract candidate patterns

Ask: what would transfer to another project? Good patterns are:

• Causal — "<input pattern> → <measurable outcome>", not "<thing happened>"
• Transferable — not tied to a single brand / client
• Falsifiable — someone else applying this could confirm or refute it

Anti-patterns (reject):

• Project-specific trivia ("antonioblago.de's homepage")
• Restatements of Peec docs ("get_actions has a scope parameter")
• Generic SEO wisdom ("write good content")

Target: 1–3 patterns per artifact. If you can only find 1, persist 1. Zero is a valid result.

3. Check for duplicates

mcp__skillmind__recall(query=<pattern_title>, k=5)

If any hit has ≥0.85 semantic similarity to the new candidate:

• Same claim + stronger evidence → mcp__skillmind__update_memory (don't re-add)
• Same claim + weaker evidence → skip
• Contradicting claim → persist anyway, tag contradicts:<existing_pattern_id>

4. Persist

mcp__skillmind__add_pattern(
  title="<≤80 chars, causal phrasing>",
  body="<structured pattern, schema below>",
  tags=["peec", "<source_skill>", "<gap_type>", "<funnel_stage>", "<market>"]
)

Required tags every pattern carries:

• peec (project family)
• source:<skill-name> (which skill observed it)
• project:<slug> (anonymized if needed)
• date:<YYYY-MM-DD> (observation date)
• ≥1 semantic tag (gap:taxonomy / funnel:decision / channel:reddit / lever:editorial / ...)

5. Confirm

Return the list of persisted patterns. If a pattern was skipped as a duplicate, say which existing pattern it merged into.

Pipeline — read mode

1. Query

mcp__skillmind__recall(
  query=<query>,
  k=<k, default 5>,
  filter_tags=[<optional narrowing tags>]
)

2. Filter by provenance (optional)

Drop hits whose project: tag matches project_id if the caller wants cross-project priors only (supplied via a exclude_own=true flag). Default: include own project's patterns.

3. Rank

Score = semantic_similarity × recency_decay × evidence_weight

• recency_decay = 0.5 ^ (months_since / 6) — a 6-month-old pattern is worth half
• evidence_weight = 1.0 for single-project patterns, 1.5 for patterns with ≥2 projects of evidence (consolidated)

4. Return

Return top-k as structured list. The caller (usually peec-agent) uses them as priors in its Decision Framework.

Pattern schema

## <causal title — ≤80 chars>

**Claim:** <one sentence, causal, falsifiable>

**Evidence:**
- <project, date>: <observation with a number>
- <project, date>: <observation with a number>
- ...

**Transfer conditions:**
- Works when: <market / funnel stage / offer type>
- Does NOT transfer when: <named conditions>

**Counter-evidence (if any):**
- <project, date>: <what contradicted it>

**Related patterns:** <pattern_id>, <pattern_id>

Example pattern (concrete)

## Editorial citations on DACH micro-publications gain 3× faster than Reddit

**Claim:** For DACH service-business projects, editorial-pitch-wins at t3n / OMR /
fachportal-niveau produce citation lift inside 10–14 days; reddit-thread-answers
typically need 3–6 weeks to surface in LLM training signal — if they surface at all.

**Evidence:**
- project:antonioblago, 2026-04: 2 editorial wins → 5 citations in 10d, 1 subreddit
  answer → 0 citations in 30d
- project:paroc, 2026-03: 1 t3n contribution → 4 citations in 8d

**Transfer conditions:**
- Works when: DACH market, B2B service offer, target domain DR 40–60
- Does NOT transfer when: consumer B2C (reddit is faster there)

**Counter-evidence:** none yet.

**Related patterns:** `pat_b12a` (reddit threads need authoritative first-5-sentences)

Tags: peec, source:peec-outreach, project:antonioblago, project:paroc, date:2026-04-22, gap:citation, channel:editorial, market:dach, funnel:decision.

Quick reference

Handoff points (where other skills call this one)

• peec-agent Phase 1 → read mode with query=<current gap type> to load priors before deciding
• peec-report Phase 7 → write mode for each entry in learnings.json.winners[] and .losers[] with source_skill="peec-report"
• peec-outreach Phase 7 (post 4-week measurement) → write mode for each status=citation_live + measured lift row
• peec-cluster Phase 7 (after zone tags exist 4+ weeks) → write mode for zones whose tag:zone:* visibility moved >10pp

Done criteria (self-check before returning)

Write mode is complete when:

1. 0–3 patterns persisted (zero is valid — don't force-fill)
2. Every persisted pattern has all required tags (peec, source, project, date, ≥1 semantic)
3. Duplicates are either skipped or consolidated, never silently doubled
4. Evidence references a measured number — no pattern is persisted on vibes

Read mode is complete when:

1. Top-k returned with scores (semantic × recency × evidence)
2. Empty result is announced plainly ("no matching patterns") — do not fabricate
3. Each result carries provenance (project + source_skill + date)

Guardrails (do not do these)

• Do not persist patterns from artifacts without a measured outcome — a brief that hasn't won its prompt yet is speculation, not a pattern
• Do not persist generic SEO advice — only claims grounded in a specific Peec-measured observation
• Do not persist more than 3 patterns per artifact — if a run produces 10 "learnings," most are noise
• Do not silently merge contradicting patterns — contradictions are information; tag them and keep both
• Do not recall without provenance — every returned pattern must say which project + skill + date it came from, or the caller can't judge transfer fit
• Do not run in write mode on <24h-old artifacts — measurement window hasn't closed
• Do not run if SkillMind MCP is unavailable — fall back to appending to <project>/growth_loop/patterns.md and flag the skip; never fabricate persistence