# Capture > Write a durable note into this project's book — a decision with the alternative that lost, a gotcha with the error text that identifies it, or a plain fact worth keeping. Use the moment something is learned that a future session would otherwise rediscover. - Skill: `erfanhabibipanah/capture` (Agent Skill) - Install (CLI): `npx skillmds@latest add erfanhabibipanah/capture` - Raw SKILL.md: https://api.skillmd.com/api/skills/erfanhabibipanah/capture/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - License: MIT - Author: erfanhabibipanah (https://skillmd.com/u/erfanhabibipanah) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/erfanhabibipanah/capture --- # Capture Write one note into the project's book. `commonbook caps` gives you the directory; if the repo is unbound, run `commonbook bind` first and say so. Pick the type honestly. The type is not decoration — each one has a requirement, and a note that cannot meet it is a different type. ## decision A choice between real options. **Required: the alternative that lost, and why.** This is the entire reason the note exists. In six months the chosen path is visible in the code; what is invisible is that you already considered and ruled out the other one, which is what stops it being relitigated. If nothing was rejected, this is not a decision. It is a preference or a default — capture it as a note instead. Also record what would make this wrong: the scale, version or price change that should trigger a revisit. ## gotcha Something that cost real time. **Required: the error text, verbatim, in a fenced block.** Retrieval is a literal string search, so a paraphrased or tidied error is an unfindable one. Copy it exactly, including the parts that look like noise. Then: the actual cause — not the first theory you abandoned — and the fix. Add what you tried that did *not* work; that section is usually worth more than the fix, because it is what stops the next person walking the same path. ## note Anything else worth keeping: a constraint, a fact about a service, where things stand. ## Checking your own work `commonbook lint` reads every note in the book and reports the ones that do not meet the contract for their declared type. Run it after writing a batch, and before trusting the book to answer anything: ``` commonbook lint ``` It ignores untyped notes entirely and exits non-zero when something is wrong, so it also works as a pre-commit check on a book kept in git. If it flags a note you just wrote, the honest fix is usually to change the *type*, not to pad the section — a fact with no rejected alternative is a note, not a decision. ## Rules for all three - Write nothing a reader could get by opening the code. No summaries of what functions do. - Never write credentials, tokens or keys into a note. - Prefer editing an existing note over adding a near-duplicate. - Keep `MEMORY.md` an index — one line per entry, detail in topic files. It is loaded into every session and is capped; long entries push later ones out. Show the finished note and where it was written.