extract

/extract

Scan the entire conversation of the current session, extract signals worth long-term recording, and append them to $AI_INFRA_DIR/extract-buffer.md.

Path configuration: $AI_INFRA_DIR must be set before use. This is the directory where all Contour data files are stored.

• Default recommended path: ~/.claude/contour/
• Set via environment variable, or replace this placeholder with your absolute path.

---

Reference Files

Before executing, read the following reference files for format specifications:

• references/signal-formats.md — recording formats for all three signal types
• references/extract-output.md — Extract Buffer write format, example output, and report template

---

File References

File	Term	Role	Access
extract-buffer.md	Extract Buffer	Temporary buffer: write extracted signals	Append
{user}-{domain}.md	Domain State	Current cognitive state for the domain	Read only
{user}-core.md	Core Profile	Personality-level traits, thinking patterns, core preferences	(not accessed)

---

Constraints

• Read the conversation content of the current session
• Read the corresponding Domain State file (e.g., {user}-coder.md) for cognitive state comparison
• Do not load Core Profile files
• Only write to extract-buffer.md (append, do not overwrite existing content)
• Do not judge deduplication, do not make target file routing decisions — those are /sync's responsibilities
• Do not run in a Contour-operational session — If the current session's primary activity was /contour:sync, /contour:setup, or other Contour commands, decline to execute and inform the user: "Extract should run in a work session, not a Contour-operational session. The sync conversation would produce echo signals."
• If the current session has no signals worth extracting, inform the user "No extractable signals in this session" and write nothing

---

Signal Type Definitions

[cognition] Cognitive Signals

Record changes or exposures in the user's cognitive state regarding a specific knowledge point. Compare against Domain State to determine whether observed behavior represents a state change.

Extraction conditions (any one is sufficient):

• The user asked a conceptual question about X ("What is X?", "How does X work?", "Explain X", "I don't understand X") → partial (the question itself signals a knowledge gap; no confirmation needed)
• The user indicated they already know something AI was explaining ("I know this", "skip the explanation", "no need to explain") → mastered
• The user demonstrated behavioral mastery: correctly used a concept/tool/process in a practical context (e.g., giving the instruction "submit this as a PR" when PR submission was previously marked as partial in Domain State) → mastered
• The user asked a basic question about a knowledge point that is marked as mastered in Domain State → partial (cognitive regression)

Do not extract:

• The user asked a purely operational question ("how do I spell this command") with no cognitive significance
• AI proactively explained a concept but the user had no reaction or interaction — cannot determine whether the user absorbed it
• Routine tool usage of concepts not tracked in Domain State — no baseline to compare against
• The user used a concept that is already marked as mastered in Domain State with no change — not a signal, just normal usage

Record each signal using the format defined in signal-formats.md.

[thinking] Thinking Signals

Record patterns or changes in the user's thinking patterns, decision logic, and problem-solving approaches.

Extraction conditions (any one is sufficient):

• The user demonstrated a clear priority ranking when making a decision (e.g., chose a lower-cost but less feature-rich solution)
• The user's argumentation exhibited a reproducible thinking pattern (e.g., always asks "what's the worst case" first)
• The user demonstrated a shift in thinking pattern within the current session (e.g., initially used one approach, then switched to another)
• The user explicitly expressed a meta-cognitive statement about "how to think about problems" (e.g., "I think we should validate before expanding")

Do not extract:

• Pure emotional expressions ("this is so annoying") — unless it reveals a value judgment
• Routine task progression without decision points
• AI-led reasoning that the user passively followed

[preference] Preference Signals

Record the user's explicit preference expressions regarding AI communication style, workflow, or tool usage.

Extraction conditions (any one is sufficient):

• The user explicitly asked AI to change communication style ("don't use metaphors", "just give the conclusion", "use Chinese")
• The user expressed satisfaction or dissatisfaction with a specific AI behavior ("this explanation is great", "stop auto-embellishing")
• The user demonstrated a work habit or workflow preference ("I prefer to get it working first, then optimize")

Do not extract:

• One-time instructions specific to the current task ("put this file in that directory") — not reusable across sessions
• General evaluations of AI capability ("you're very useful") — no actionable signal

---

Execution Steps

1. Read the reference files listed above for format specifications
2. Read the corresponding Domain State file to establish the user's current cognitive baseline
3. Review the entire conversation of the current session
4. Screen for each of the three signal types according to the definitions above, strictly following extraction and non-extraction conditions
5. If signals exist, append to extract-buffer.md following the write format in extract-output.md
6. Report extraction results to the user following the report template in extract-output.md
7. If no signals, inform the user and write nothing