skill-harvest

What this skill does

The video that birthed this (Chase AI, "Agentic OS Setup That Will 10x Claude Code", 2026-06-25) names three ways to find what to turn into a skill: do it manually, have Claude read your last N sessions and propose skills, or have Claude interview you. /level-up is the interview. This is the read-your-sessions one.

It runs a deterministic miner over the real session transcripts, then I (Claude) do the semantic clustering and judgment the script can't — so we go off what Gera actually does, not what he thinks he does. One run = a ranked candidate table + one skill teed up to build.

It pairs with [[level-up]] (the 3Ms interview) and feeds [[skill-builder]] (which actually builds the chosen skill). It is discovery, not construction — it never writes a skill itself; it hands the winner to /skill-builder.

When to use vs /level-up

• /skill-harvest when you want evidence: "what does the data say I repeat?" Best run every few weeks, or when Gera suspects he's doing something by hand a lot.
• /level-up when you want the 3Ms interview + the Mindset install. Best as the Friday ritual.
• They compose: harvest surfaces candidates → level-up/skill-builder ships one.

The process

Step 1 — Mine the sessions

Run the miner and read its report:

python3 scripts/skill-harvest.py --sessions 40

Flags: --sessions N (default 30; 0 = all), --days D (only sessions touched in the last D days), --min-repeat K (group threshold, default 2), --json OUT (also dump the structured payload). For a broad sweep use --sessions 80; for "what have I done lately" use --days 10.

The report gives you: the existing-skills roster (so you exclude what's covered), recurring intents grouped by repeat count (the top candidates), slash-command usage (hot skills + any ⚠ not a known skill commands), a distinct-prompt list (the raw material to cluster), slash-command args, and a bash-pattern histogram (recurring manual command workflows). The script already strips harness noise, routine/agent prompts, and Moltbook challenges — what's left is mostly real intent.

Step 2 — Cluster semantically (this is the judgment the script can't do)

Read the distinct-prompt list and group prompts that are the same underlying task even when worded differently. Example clusters seen in practice: "draft a status message for my SoftServe team chat", "prep + QA a research deliverable doc", "make an explainer video for person/team X". A cluster of 2+ distinct prompts that map to one repeated workflow is a candidate.

Step 3 — Classify every candidate (do NOT skip this)

Tag each cluster as exactly one of:

1. Already covered — an existing skill does this. Drop it (note which skill).
2. Already automated — a routine/agent already runs it unattended (the daily posts, pending-drain, moltbook, the messaging bridges). Drop it.
3. 🚧 SoftServe-bound — the work is SoftServe customer/Internal+ work (R&D datasets, RaaS/Jumpstart deliverables, customer code, team-chat about SoftServe work). This is the critical gate. Per references/cowork-bridge.md + references/softserve-ai-usage-policy.md, personal Servy must NEVER build a skill that operates on SoftServe data. Recommend a Cowork playbook instead: Servy drafts the instructions-only playbook into routines/cowork/ and Gera runs it in Claude Desktop (Enterprise). Surface it as "build this, but Enterprise-side, not here."
4. ⏰ Routine-shaped — the cluster fires on a predictable trigger, not on Gera deciding to run it: an event ("every time a lead lands / a specific type of email arrives / a ticket comes in") or a cadence ("every Monday / every Friday"). Propose a ROUTINE (self-firing), not a manual-fire skill — a skill Gera must trigger keeps him as the trigger: "if you're not around to trigger it, nothing happens." (Nate 6h course, 2026-07-11 [0:20:51] + [1:55:16]; the interview-side twin is /level-up Phase 1 Q6–Q7.)
5. ✅ Buildable here — a genuinely repeated, personal/public-data workflow with no skill yet. These are the real wins.

Also flag any ⚠ not a known skill slash command — that's Gera invoking a command that doesn't exist (a skill he wishes he had), which is a strong signal.

Step 4 — Present the candidate table

Output a scannable table, candidates ranked by (repeat count × leverage), buildable ones first:

Keep it honest: if the only repeated work is already covered or SoftServe-bound, say "nothing net-new to build here, and here's why" — that's a valid, useful outcome (same discipline as the wiki build-gate).

Step 5 — Offer to build the top pick

Recommend the single highest-leverage ✅ Buildable candidate and offer to hand it to /skill-builder now. A ⏰ Routine-shaped winner goes to /routines-builder / /servy-routine instead — the deliverable is a self-firing routine, not a skill. For 🚧 SoftServe candidates, offer to draft an instructions-only playbook into routines/cowork/ (zero SoftServe data) for Gera to run in Claude Desktop (Enterprise). One run should end with either a skill being built or a clear "nothing to build, here's the evidence."

Boundaries

• Discovery only. Never write a skill here — that's /skill-builder's job.
• SoftServe firewall is non-negotiable. Never propose building a SoftServe-data skill in personal Servy. Route it to Claude Desktop (Enterprise) as a routines/cowork/ playbook. When unsure whether a candidate is SoftServe-bound, treat it as SoftServe-bound.
• Read-only on history. The miner only reads transcripts; it writes nothing except the optional --json dump you point it at (use the job tmp dir, not the repo).
• No raw transcript dumping into chat. Read the miner's report, not the raw .jsonl files — they're huge and may carry SoftServe content.

Why a script underneath (WAT)

The deterministic extraction — parse 30+ transcripts, filter noise, count repeats, diff against the skill roster — is code's job (scripts/skill-harvest.py). The semantic clustering + classification + judgment is mine. Code extracts, AI reasons. See [[wat-framework]].