Capture (Settled Thinking)

I keep a small set of source-of-truth files for my own thinking — settled positions, the state of my work-streams, ideas worth banking. The danger to that record is drift. A single conversation throws off a dozen half-formed ideas, and if they all get written down as settled, the files inflate into noise and stop being trustworthy. The discipline I want is a funnel that admits only what genuinely moved.

This skill is that funnel. At the end of a thinking session it reads the conversation and extracts what changed — positions that hardened, beliefs that shifted, work-stream developments, stray ideas — into one strictly-formatted file, then ships it to an inbox where a separate reconcile step folds it into the source-of-truth files. It is deliberately not a summary of what was said. It is a structured extraction whose every heading and field marker is shaped so the downstream parser can read it literally.

How it works

The file is named by date and a short slug, and written in exactly one structure. A header gives the date and a two-to-four-sentence note on how the thinking moved. Then five sections, in a fixed order with verbatim field markers, because the reconcile step parses them rather than reads them:

• Settled positions — the claim, the context behind it (derivation, scope, what it supersedes), and a source quote in my own words when a sharp one exists.
• Changes to standing thinking — the change, an explicit CONTRADICTS or REFINES line naming the existing position and the file it lives in, then the context.
• Updates to current state — what is now true, typed as top-of-mind, recent shift, recent decision, or parked, plus which existing entry it makes stale.
• Updates to work-streams — a development, tagged to one work-stream, typed as a new development or shifted positioning.
• Idea backlog — a flat, date-prefixed list, one bullet per idea, never classified.

The extraction rules enforce honesty over volume. Only what settled or shifted this conversation goes in; empty sections read “None this conversation” and are never padded. A refinement to an existing belief is forced into “changes to standing thinking”, not promoted to a new settled position. Contradictions are surfaced out loud rather than smoothed over. Context is asked to carry more than the conclusion — how settled it felt, what would reverse it, and anything explicitly ruled out — because the reconcile step already reads that field, so the detail is free.

Shipping is a fixed shell sequence: shallow-clone the courier repo, copy the file into the inbox, commit under a set identity, and push, with one pull --rebase retry if the push is rejected. Success is never reported without the actual push output quoted back. If transport fails twice, the capture is handed over as a downloadable file with a plain statement that the push failed — so no capture is ever lost.

The split that makes it work

This is one of a deliberately separate pair. A sibling skill records what happened in a chat — descriptive history plus artifacts. This one extracts what was decided, in a parse-ready format. They are not interchangeable, and the hard rules forbid imitating the history-capture summary here. The strict format is the contract: the reconcile step is a parser, not a reader, so a free-form summary would break the pipeline. Lock the section names and field markers and the merge stays automatic. The credential that authenticates the push must never appear in the file or anything shipped.