| name | mine |
| description | Seed a brain from a LARGE existing source by TRIAGING it, then calibrating the agent's judgment on small batches, then bulk-mining. Use when the owner says "mine my [source]", "mine this export", "calibrate the mining", "idea mining", "process my archive", or "seed my brain from this export" — pointing at a sizable export (notes-app dump, chat history, voice-note archive, bookmarks, podcast snips). Phase 0 TRIAGE: categorize the source into keep/skip bins, sample the skip-bins, get owner approval per bin, build the target list. Phase 1 CALIBRATE: small batches → present COMPACT numbered proposals (verbatim quote + the agent's call + one-line reason, no tables) → owner corrects → write every correction back into the rules (CONTENT_TAXONOMY / KEYWORD_GUIDE / INTELLECTUAL_LANDSCAPE) → draft the full Template only after the calls are approved → repeat until a batch is clean. Phase 2 SET LOOSE: process the remainder autonomously, spot-check, re-calibrate on drift. Implements "invest in structure and calibrate the AI judgement early"; full procedure in Orientation_Docs/MINING.md. NOT for in-session pasted/dictated thoughts (use /ingest-brain-dump). Preserves the owner's exact language — no paraphrasing. |
| required_context_files | ["Orientation_Docs/MINING.md","Orientation_Docs/CONTENT_TAXONOMY.md","Orientation_Docs/KEYWORD_GUIDE.md","Orientation_Docs/INTELLECTUAL_LANDSCAPE.md"] |
Mine (Triage → Calibrate → Set Loose)
Turn a large existing source into clean Template A/B files. Triage it first so you only mine what's
worth mining; calibrate the agent's judgment on small batches; then set it loose on the bulk. Full
rationale: Orientation_Docs/MINING.md.
Quick Start
- Triage: categorize the source into keep/skip bins → sample each skip-bin (~10) → owner approves
per bin → build the target list.
- Calibrate: small batch → present compact numbered proposals → owner corrects → write corrections
back into the rules → draft full Templates only after the calls are approved → repeat until clean.
- Set loose: process the remainder autonomously with the converged rules; spot-check; re-calibrate
on drift.
Workflow Checklist
Copy and track:
- [ ] PHASE 0 — TRIAGE
- [ ] Budget it: estimate the source's token size (`scripts/estimate_tokens.py <path> --budget <plan>`) vs the owner's plan; if reading it all blows the session, stage it across quota resets
- [ ] Categorize the whole source into coarse keep/skip bins by kind
- [ ] Sample each skip-bin (~10 items) with a one-line observation
- [ ] STOP — owner approves each bin (propose, don't decide)
- [ ] Produce the target list, in an agreed order
- [ ] Load rules: CONTENT_TAXONOMY.md, KEYWORD_GUIDE.md, routing-rules.md, classification-rules.md, INTELLECTUAL_LANDSCAPE.md
- [ ] PHASE 1 — CALIBRATE (repeat):
- [ ] Take a small batch (~5–10 targets)
- [ ] Per item: semantic segment → /verify-idea dedup → classify + route
- [ ] Present COMPACT numbered proposals (verbatim quote + call + one-line reason; NO tables, NOT full files)
- [ ] Collect the owner's corrections to the calls
- [ ] Write EACH correction back: CONTENT_TAXONOMY.md / KEYWORD_GUIDE.md / INTELLECTUAL_LANDSCAPE.md (project boundaries)
- [ ] After the calls are approved, draft the FULL Template A/B and present before saving (Create? Y/N)
- [ ] Log corrections-this-batch; loop until a batch is clean (~0 corrections)
- [ ] PHASE 2 — SET LOOSE:
- [ ] Process remaining targets autonomously in larger batches with the converged rules
- [ ] Dedup each; log progress (processed, created, duplicates skipped)
- [ ] validate_template.py + verify-second-brain as you go (recommended)
- [ ] Spot-check periodically; on drift, drop to one calibration batch, fold the fix, resume
- [ ] Report: bins skipped (approved), items mined, files created, dedups skipped, rules updated
Critical Principles
- Triage before you calibrate. A raw export is mostly noise; sample the skip-bins and get the
owner's approval before processing anything in bulk. Agents propose, the owner decides.
- Calibrate before the bulk pass. A cold classifier mis-files at scale and you find out one wrong
file at a time. A few correction rounds up front save thousands of silent errors.
- Present compact proposals, not full files. During calibration show one numbered item per idea —
verbatim quote + the call + a one-line reason, no tables. Draft the full Template only once the
owner has approved the call (then present it before saving, per "Template A IS the artifact").
- Corrections are written back into the rules. A correction not folded into
CONTENT_TAXONOMY.md /
KEYWORD_GUIDE.md / INTELLECTUAL_LANDSCAPE.md will be made again. The rules are the memory of the
calibration — this is what makes the loop converge.
- Find a home for everything. Default to filing; skipping is rare and reserved for the obviously
content-free. When unsure, file it.
- An external reference is not an auto-skip. A note about someone else's work still earns a
reading-list / concept / tool note (Template B, attributed).
- Split multi-idea items, and verify the split. Note the expected number of ideas before splitting
and the extracted number after; if extracted < expected, review before approving.
- Semantic segmentation, never algorithmic chunking. A paragraph may hold several ideas; several
paragraphs may be one.
- Preserve exact language; mine the owner's words. Original Text is verbatim; keywords are the
owner's actual terms. In a multi-speaker source (chat export, interview), mine only the owner's side.
- Convergence is the stop signal — and it's not permanent. Stop calibrating when a batch comes
back clean; resume calibrating when a new source shape makes the bulk pass drift.
Phase 0 — Triage
- Budget it first. Estimate how much reading the source in full would cost —
python3 scripts/estimate_tokens.py <path> --budget <plan-tokens> — and weigh it against the owner's
plan (Claude Pro vs Max, or an API budget). If reading everything would blow one session window, plan
to stage the mine across quota resets (a Pro ~5-hour window, overnight, week's end) and use a
cheap/fast model for the bulk reading. Sizing up front is what stops a single import from burning the
whole plan.
- Categorize the whole source into coarse bins by kind (keep kinds vs skip kinds).
- For each skip-bin, pull ~10 items, state a one-line observation, and present the sample.
- Stop for the owner's approval per bin — "skip it" / "sample more" / "keep these." No bin is
skipped on the agent's judgment alone.
- Produce the mining target list in an order the owner agrees. For a mixed keep-bin, sample a few and
agree an approach first; don't auto-pilot.
Phase 1 — Calibrate (small batches)
For each batch:
- Pull a small batch (~5–10) of targets in the agreed order.
- Per item: segment →
/verify-idea (skip true duplicates, surface near-duplicates) → classify per
CONTENT_TAXONOMY.md + the classification rules → decide routing per the routing rules.
- Present compact numbered proposals: verbatim quote · the call (category / template / project) ·
one-line reason. No tables, no full files yet.
- The owner corrects the calls. Write back for every correction — a new boundary in
CONTENT_TAXONOMY.md, a term in KEYWORD_GUIDE.md, a project-boundary call in
INTELLECTUAL_LANDSCAPE.md. Note what changed.
- Once the calls are approved, draft the full Template A (original thinking) or Template B
(external, attributed) and present before saving.
- Track corrections per batch. When a batch comes back clean, calibration has converged.
Phase 2 — Set loose (bulk mine)
- Process the remaining targets autonomously in larger batches, applying the converged rules.
- Dedup-check each item; log progress (processed, created, duplicates skipped). Validating each created
file (
validate_template.py) and running verify-second-brain is recommended.
- Spot-check a batch periodically. On drift, drop to one calibration batch, fold the fix into the rules,
resume.
Integration
| Skill / script | When | Status |
|---|
scripts/estimate_tokens.py | Phase 0 — size the source vs the plan before reading | core |
/verify-idea | Per item — duplicate detection before creating a file | core |
/process-content | The per-item segment → classify → Template-create mechanic | core |
.claude/scripts/validate_template.py | Validate created files in the bulk pass | recommended |
verify-second-brain agent | Template/placement/keyword/language check | recommended |
/ingest-brain-dump | The in-session counterpart (pasted/dictated thoughts, not a bulk export) | sibling |
Verification Checklist
Before calling a mine run done, confirm:
- Every skip-bin was sampled and approved by the owner — none skipped on the agent's say-so.
- Every correction the owner made during calibration was written back into a rule file
(
CONTENT_TAXONOMY.md / KEYWORD_GUIDE.md / INTELLECTUAL_LANDSCAPE.md).
- Multi-idea items were split, and each split's extracted count met or exceeded the expected count.
- The bulk pass used the converged rules (not the shipped defaults).
- Original Text is verbatim; external content is Template B with attribution; in multi-speaker sources
only the owner's words were mined.
- You reported bins skipped, items mined, files created, duplicates skipped, and which rules were updated.