Session State

One rule: state lives in Linear and is pulled by reference. It is never relayed by paste.

Why this exists

An audit of ten consecutive sessions found four taxed by the handoff-doc loop (write doc → paste into new chat → repeat), and two killed on the first turn by policy-error rejections of the pasted blob. Pasted context also arrives as one undifferentiated block, which bypasses the skill triggers that would otherwise fire on a normal request. The handoff doc feels like diligence; it is actually the single largest workflow tax.

The three moments

1. Closing a session (intercepts "make me a handoff doc")

When Nick asks for a handoff doc, or the session is wrapping with known next steps:

1. Create or update the Linear state doc for the campaign/project (format below). Attach it to the relevant Linear issue or project. Run linear-write-cleanup before writing.
2. Reply with a one-line resume command, e.g.: Next session: "continue the paid ads campaign" — state is in Linear.
3. Offer to launch the next phase now — parallel subagents for research or production work, a scheduled task for recurring work — instead of leaving it documented-but-unstarted. Respect existing approval gates (Slack posts, HubSpot writes, Linear writes Nick reviews).

Do not produce a handoff document in chat. If Nick needs a portable copy for a person, write the Linear doc first and export from there — Linear stays the source of truth.

2. Resuming a session ("continue X")

1. Find the state doc: list_documents / get_document on the campaign's project or issue.
2. Load it, note its last-updated date, and flag what's likely stale (deal stages, quotes, metrics).
3. Re-verify volatile facts from their sources of truth (gtm-rules governs which) rather than trusting the doc's copies.
4. Start working. Never ask Nick to paste context he has already paid to produce once.

3. Intercepting a paste

If Nick pastes a handoff blob anyway: extract what's new, merge it into the Linear state doc, treat embedded quotes and pipeline facts as unverified until re-pulled, then proceed with the actual work. One line of redirect is enough — "merged into the Linear state doc; next time 'continue X' is all I need." Fix the plumbing without lecturing.

State doc format

Keep it under ~60 lines. It is a pointer index plus decision log, not prose:

# State: [Campaign / project name]
Updated: [date] · Status: [active / paused / blocked on X]

## Decisions made
- [decision] — [one-line why]

## Open items
- [item] — [owner]

## Next actions
- [action] — [launchable now? Y/N]

## Links
- Issue/project: [Linear ref]
- Assets: [files, designs, docs]
- Sources: [calls and docs the claims came from]

The cap matters: a long state doc becomes the new handoff blob. Anything wanting more than a line belongs in a linked asset, and verbatim customer quotes belong in transcripts (re-pull via messaging-from-calls), not in state.

What this skill does not own

• Sentence quality of the doc → linear-write-cleanup
• Campaign brief structure → idea-to-brief (its briefs persist to Linear under the same rule)
• External deliverables → content-proportion-check and the content stack

Precedence note

Where this skill and portable-handoff disagree about producing a handoff document, portable-handoff wins, by decision on 2026-08-06. The disagreement is kept visible on purpose; see the library README.