handoff-discipline

A summary loses what the next agent must not re-decide. This skill turns a session's working state into a handoff document the next agent can execute against without reopening closed questions.

Stance (load-bearing biases)

1. A handoff is a transfer of decision boundaries, not a transfer of information. Mechanism: a summary preserves information and discards decision status. The receiving agent needs the status — which questions are closed — more than the content, because an agent that cannot tell closed from open will re-litigate closed questions with full confidence. Field observation: in one cross-instance handoff I worked from, an account handle had been mistyped into a transposed variant. The handoff carried no pinned-decisions list and demanded no item-by-item echo, so the receiving instance built on the wrong handle and nothing in the pipeline stopped it. The error propagated silently into the output. Boundary: verified on handoffs between two AI projects owned by the same person, same long-running project. Cross-person and cross-organization handoffs: untested.

2. Three layers must be explicitly separated: pinned / status / rationale. Mechanism: when the layers are interleaved, the receiving agent cannot tell which sentence is a fixed constraint and which is updatable state. Tactical status gets read as a strategic invitation; pinned decisions get read as context it may improve on. Field observation: same handoff practice as above — the documents that worked were the ones that listed pinned decisions as a numbered block at the top, separate from progress notes. The ones that mixed them produced receiving agents that "helpfully" redesigned settled structure. Boundary: layer separation has been exercised on project-state handoffs (naming, architecture, scope decisions). Handoffs of pure creative drafts: untested — the pinned layer may be near-empty there.

3. The receiving agent gets a written role boundary, including a yield rule. Mechanism: without an explicit "execute, don't re-decide" clause plus a conflict priority (user's live word > handoff doc > older docs), the receiving agent defaults to its general helpfulness prior — which includes re-deciding. Field observation: handoffs in the source practice that included a receipt requirement (echo back pinned list + open items + next step) caught transcription drift at intake instead of at delivery. Boundary: yield rule tested only where the same user is on both ends and can adjudicate live.

Core actions

1. Classify before writing. Sort the source material into three layers — pinned / current status / strategic rationale — and quarantine the danger class: items pending review, which belong to neither pinned nor status. Rules: read reference/layers.md.
2. Fill the template. Use the handoff skeleton with its fixed field order. Fields: read reference/template.md.
3. Append the receiving protocol. Role boundary, yield rule, receipt requirement. Text: read reference/receiving.md.
4. Gate before output. Echo the pinned-decisions list to the user and get confirmation before producing the document. This is the only step that prevents a pending item from being frozen as pinned.

Skill fields

• Version: 0.1 (2026-06). Changelog: corrections are added, not erased; superseded guidance gets a reverse mark, old handoff documents are archived, never overwritten.
• Retirement condition: when AI systems hold persistent shared memory across sessions and instances, so that handoff no longer requires a human-mediated document, this skill retires.
• Failure behavior: if the source material does not let you tell pinned from pending, stop and have the user classify item by item. Do not guess which decisions are pinned — a wrong guess either freezes a judgment that should update or reopens one that should not, and both costs are silent.
• Composition: composes with any project-management or memory skill; this skill only governs the structure of a cross-context handoff. Yields to explicit user instruction.
• Time: judgments fixed 2026-06. The premise "AI lacks persistent cross-session memory" is moving; re-review expected within 12–24 months.
• Boundary: verified for one person handing one long-running project between two of their own AI projects. Cross-person, cross-organization, and adversarial handoffs: untested, unclaimed.