anchor

Re-deriving a mental model that was already built once is expensive; skimming ten one-line bullets is cheap. Your job: capture the identified model (argument) or models built in this conversation so the next session — or the user walking into a meeting cold — can "reload" the concept or concepts in under a minute.

This is NOT a summary of the conversation. It is a distillation of the understanding of one or more concepts — the things that were confusing until they weren't.

What to capture

INVARIANT — Argument sets the scope

Arguments set the scope. If the user named a topic or pointed at a thing ("that diagram", "the yes/no path"), anchor that argument only — the unit is the argument, not the session. Do not widen to cover the session: other topics, today's facts (ids, dates, numbers, incidents), and who-said-what stay out even when they feel load-bearing — they belong in the project's fact/doc files, not here. The scan list below applies only to grounding the named argument.

Argument-scoped sheets are SHORT. A named argument typically needs 1 section and 2–5 bullets — if your draft exceeds that, you are smuggling in session material; cut it.

No argument given

Only when NO argument is given: scan the recent conversation for moments where clarity was expensive to reach:

• Distinctions that resolved confusion — "X and Y are two different moments, don't blend them". These are the highest-value anchors; if one distinction organizes everything else, lead the whole sheet with it.
• Cardinalities and mappings — 1:1:1 relationships, what scales with what, which object multiplies. Confusion here is quiet and expensive.
• Anatomy — the shape of a key/token/payload/config the user had to have explained.
• Decided vs open — one-liners for what's settled (so it isn't relitigated) and what's genuinely open (so it isn't assumed settled).
• Guardrails — "settled, don't reopen", priority ceilings, constraints from other teams. Collect these in a closing Standing guardrails section.
• Corrections of prior belief — where the user (or the docs) believed X and it turned out to be Y. State Y plainly; don't re-narrate the correction.

Leave out: narrative, process, who-said-what, anything the user never stumbled on. If it was never confusing, it doesn't need an anchor.

Fresh-session test for every bullet: would this line still be true — and still useful — in a new session on this same project, without today's context? A bullet naming session specifics (a teammate, an environment id, a probe result, a latency number) fails: that is a fact, not a model. Cut it, or move it to the project's fact file.

With arguments: scope as above. Without arguments (only then): judge from the list.

Where it goes

One sheet, merged over time — never a new file per invocation.

First resolve the root, in this order:

1. If the user's instructions (including their local context file) say where generated, non-committed files go — use that.
2. Else, if .context/ exists at the repo root — use it, following .context/AGENTS.md if present.
3. Else — ask once where to save (use a structured question dialog if your tool has one); if you cannot ask, use the repo root and state where you wrote.

Then place the sheet under that root. Two cases:

• An active project directory exists — a project folder (under projects/ or research/) that this session has already been writing into, or one the user named. The sheet is anchor.md inside it, alongside the effort it describes. A project path merely mentioned or read once does not count; when in doubt, use the bucket.
• Otherwise — the sheet goes in an anchor/ bucket under the root, named YYYY-MM-DD-<slug>.md, where the slug names the topic.

Merge — never duplicate. Read the file before writing. Update an existing section in place when a topic recurs; replace bullets the conversation has contradicted (the sheet states current truth, not history); append new sections for genuinely new topics.

In a project directory there is exactly one anchor.md, so the merge target is obvious. In an anchor/ bucket, glob anchor/*-<slug>.md first: if one matches, merge into it and leave the filename alone — the date records when the sheet was started, not when it was last touched. Create a new dated file only when no slug matches. A bucket full of near-duplicate sheets defeats the whole point, because then the user has to read all of them to find out what's currently true.

Style contract (the whole point — hold it strictly)

• One idea per bullet, one line each. If a bullet needs two lines, it's two bullets or too much detail for a cheat sheet.
• Bold the key term in a bullet so a skimming eye catches it.
• Lead with the most organizing distinction — the one bullet that makes the rest make sense.
• Short ## sections with plain names. A section is 2–6 bullets; more means split or prune.
• Tiny ASCII fragments only when a shape genuinely beats words (a 3-line mapping, an arrow chain). No elaborate diagrams — those belong in the full docs, which the sheet can point to. The sheet may contain at most one ASCII fragment, at most 8 rendered lines.
• Plain words. No unexplained jargon: the sheet must work when the reader has ZERO context loaded — that's its entire job.
• When the argument names a diagram (the yes/no path, a flow): the diagram itself goes on the sheet as the section's body — surrounding bullets stay ≤ 4.
• Standing guardrails are FORBIDDEN in argument-scoped sheets (they belong only to no-argument session scans). Never invent them for a named topic — the named topic absorbed no session guardrails.

Example (shape, not content)

The subject below is invented. Read it for the shapes — an organizing distinction first, then cardinalities, anatomy, and guardrails last.

# Cheat sheet — mental models (skim to reload context)

## The two delivery moments (don't blend them)
1. **Accept** ("did we take the event?", once) → 202 and a queue write. Decided.
2. **Deliver** ("did the subscriber get it?", retried) → the backoff debate lives here.

## Cardinalities (1:N:N)
- 1 event : N subscriptions : N attempts. Attempts scale with **retries**, not subscribers.
- Pausing a subscription stops **new attempts only** — in-flight ones still land.

## Payload anatomy
- `id` is the **attempt** id, not the event id. Dedupe on `event_id`.

## Standing guardrails
- Ordering: not guaranteed, by design. Settled — don't reopen.
- Retry ceiling is 5. Raising it needs a capacity review first.

After writing

Reply with two or three sentences: where the sheet lives, which sections were added or updated, and — if you merged — what got replaced. Don't paste the sheet back into the chat; the user just wrote it with you.