Model Migration

Personalized upskilling for a new model, grounded in your real history. Inspired by Mark Kashef's "dynamic workflows" model-migration pattern (YT 9_ExDZFlaNc). The point: skip the generic "this model is INSANE" tutorial and generate advice tailored to how you actually prompt and which tools you actually use.

When to run

A new model shipped and you want to know what to change. Run it per model bump (4.8 → 4.9 → 5 → ...). Not a daily tool — it burns tokens via agent fan-out. Re-run sporadically.

The recipe

1. Deterministic full-corpus mine (zero LLM tokens)

Run the bundled miner — it processes EVERY ~/.claude/projects/**/*.jsonl session with jq/awk and writes a digest. This is the "data analysis" layer; doing it deterministically keeps the 600MB+ of raw session noise out of agent context.

bash ~/.claude/skills/model-migration/mine_sessions.sh

Outputs to the directory set by $MODEL_MIGRATION_OUT (default: ~/model-migration-output/):

• 01-models.txt — model usage histogram (are you ignoring your own routing matrix?)
• 02-tools.txt — tool-use histogram (your real workflow shape)
• 03-tokens.txt — input/output/cache token totals
• 04-user-prompts.txt — your full prompt corpus (the signal for prompting-pattern analysis)
• 05-session-dist.txt — session length distribution

2. Fan-out analysis (Workflow tool)

Requires explicit opt-in (the user asking for this skill IS the opt-in). Author a Workflow that:

• Qualitative prompt-pattern agents — chunk 04-user-prompts.txt (it's a few MB, NOT the full 600MB) across N agents; each extracts: recurring prompt shapes, verbosity, where the user over/under-specifies, retry/frustration signals, task archetypes.
• claude-code-guide agent (agentType: 'claude-code-guide') — fetch the NEW model's release notes + Claude Code docs; produce a "what changed" brief grounded in Anthropic docs.
• Synthesis agent — merge the deterministic digest + qualitative findings + the docs brief into the report.

3. Report sections (match Mark's structure)

1. Executive summary / checklist
2. Usage profile analysis (from the digest)
3. Direct competency comparison (old model → new model, tied to your tasks)
4. Prompting patterns to change (be more literal with newer models; they handle direct prompts without hedging)
5. Skill & tool recommendations
6. Dynamic workflows you could build next
7. Advanced opportunities

Write the report as HTML to the output dir so it's shareable.

4. (Optional) Video

Mark converted the HTML to a 2-min explainer with Hyperframes (open-source, by Hay Jin). You have Remotion skills (remotion-studio, remotion-kickstart) — use those instead to render a walkthrough if you want the video artifact.

Notes

• "All sessions" is honored by the deterministic pass (it genuinely reads all 1,014). The qualitative agents work the distilled prompt corpus — reading raw tool-call/thinking noise wastes tokens without adding signal.
• Scope agent count to the model's importance. A point release ≠ 60 agents.

Workflow reliability (learned 2026-06-02)

Do NOT attach a schema to the prompt-pattern agents — they read a chunk and naturally end in prose, so StructuredOutput forcing fails (12/12 failed on the first run; synthesis compensated). Have pattern agents return MARKDOWN and let the synthesis agent read it, OR end their prompt with "Your FINAL action MUST be the StructuredOutput call; do not end with prose." Synthesis/writer agents with no schema are reliable. Pattern: have agents return MARKDOWN and let synthesis read it; schema forcing on prose-output agents fails reliably.