personal-wiki-mine

name	personal-wiki-mine
description	Mine mechanism-first candidate claims from a bounded source set across thoughts/ and Heptabase MCP. Writes candidate files to thoughts/global/fredrick/candidates/ and stops for human scoring. Trigger: "personal-wiki-mine", "zettel mine", "candidate claims", "卡片盒連結挖掘".
allowed-tools	Read, Write, Bash, Agent

personal-wiki-mine

Generate candidate claims for a personal knowledge mining workflow. The output is always a candidate, never adopted knowledge.

Scope

Use this skill when the user wants to:

• mine cross-source candidate claims;
• rerun a bounded Zettelkasten-style curation workflow;
• produce a reviewable candidate file for human scoring;
• turn a known source bundle into mechanism-first insights.

Do not use this skill for:

• background scanning;
• automatic wiki promotion;
• team wiki ingestion;
• unbounded discovery across every available source.

Required Input

The user must provide a bounded source brief, explicit file paths, or a known source bundle.

Examples:

• Mine candidate claims from AI Coding / personal wiki / item-label sources
• Use the successful v2 source set again
• Generate candidate claims from these 5 thoughts files and related Heptabase cards

If the source brief is missing, ask for it before proceeding.

Output Path

Write the result to:

thoughts/global/fredrick/candidates/YYYY-MM-DD-personal-wiki-mine-candidates.md

If thoughts/global/fredrick/candidates/ does not exist, create it first.

Workflow

1. Resolve Source Set

Turn the user request into a bounded source manifest:

• Heptabase whiteboards/cards to fetch
• card-level search queries to use if whiteboard coverage is partial
• thoughts files to read fully
• any previous candidate or decision files to use as eval traces

Record any assumptions in the output manifest.

2. Fetch Heptabase Sources

Whiteboards:

1. search_whiteboards to confirm the target
2. get_whiteboard_with_objects to fetch sections, connections, and cards
3. classify coverage for each requested whiteboard:
   • complete: sections/cards/connections are available
   • partial: whiteboard exists but object expansion is empty or clearly incomplete
   • not-found: requested whiteboard cannot be found
   • ambiguous: multiple plausible whiteboards exist

Cards:

1. semantic_search_objects to find candidates
2. get_object to read the selected cards fully

[concept] cards:

• use title/semantic search
• do not rely on get_tag_cards("concept")

Coverage rule:

• Do not treat a partial whiteboard fetch as complete source coverage.
• If a requested whiteboard is partial, not-found, or ambiguous, run compensating card search before mining claims.
• Compensating searches must use the whiteboard name, source-brief concepts, visible section/card titles when available, and known related terms from thoughts sources.
• Fetch selected compensating cards fully with get_object.
• Do not use a partial whiteboard shell as direct evidence for a candidate unless it is paired with fetched cards or thoughts files.
• Record the coverage classification, compensating queries, selected cards, and known missing signals in the output file.

3. Read thoughts Sources Fully

Read all selected markdown files fully.

Extract:

• strong claims
• failure modes
• design rules
• lifecycle states
• review/oracle positions

Do not treat raw facts as claims until they imply a stance or mechanism.

4. Extract GT Patterns

From whiteboards, extract the user's hand-authored relation shapes:

• problem -> technique
• technique -> workflow
• workflow -> best practice
• concept -> concrete action

Summarize 3-5 GT patterns in the candidate file before mining claims.

5. Extract Atomic Claims

From each source, produce 1-5 atomic claims.

Atomic claims should be:

• one sentence
• opinionated or explanatory
• grounded in the source
• reusable in multi-source synthesis

6. Mine Candidate Claims

Generate 5 candidate claims.

Each candidate must have:

• at least 3 evidence items
• at least 2 source types
• a mechanism-first hypothesis
• a falsifiable rebuttal path
• an operational next step

Prefer cross-context mechanisms over clever metaphors.

7. Run Calibrated Anti-Filter

Reject any candidate that is:

• pure structural analogy
• single-source summary disguised as synthesis
• built on similar, related, or both without mechanism
• unable to produce a checklist, schema, workflow, experiment, or implementation decision

Use the rules in references/filter-calibration.md.

8. Write Candidate File

Use references/output-template.md and references/candidate-schema.md.

The file must include:

• frontmatter with status: awaiting-user-scoring
• source manifest
• source coverage audit
• Heptabase fetch notes
• GT pattern summary
• 5 candidate claims
• self-audit table
• user scoring section
• overall reflection section

If scripts/validate_candidate_file.py is available, run it on the generated file and fix structural failures before stopping.

9. Stop For Human Scoring

Do not promote anything into thoughts/wiki/.

The workflow ends after writing the candidate file and telling the user where it is.

Quality Bar

The candidate file is only acceptable if:

• every candidate is mechanism-first;
• evidence is concrete and source-linked;
• next steps are operational;
• failed or partial MCP fetches are disclosed and compensated with card-level search where possible;
• no candidate relies only on a partial whiteboard shell as evidence;
• source coverage audit explains what was requested, fetched, compensated, and still missing;
• the output clearly remains in candidate state.

References

• references/candidate-schema.md
• references/filter-calibration.md
• references/source-protocol.md
• references/output-template.md
• scripts/validate_candidate_file.py
• evals/evals.json