prysai/prysai-llm-playbook/prysai-shift-handoff

Claude Code and agent skill by prysai/prysai-llm-playbook

Trust tier 3
  • community
Source check pending

Checking that the SKILL.md file is still in the repository.

Prepare one current-work brief for a recurring LLM collaboration by keeping reusable criteria, the changing item, authority, and acceptance evidence distinct. Use when yesterday's context or an earlier example could be mistaken for today's work. Do not use to create a product context, design a full task protocol, recover an interrupted task, audit a completed claim, or execute an action.

Install

npx vskill@latest install prysai/prysai-llm-playbook/prysai-shift-handoff

SKILL.md

Shift Handoff

Prepare the smallest brief that makes one changing work item visible without pretending that a prior chat turn, example, permission, or result still applies. This Skill prepares a handoff; it does not send a prompt, inspect a system, or perform the work.

Own only the recurring-work seam

Use this Skill when a repeated text workflow already has durable criteria but one item changes: for example, classify today's feedback note under an approved taxonomy, review this week's short update against stable house style, or turn one new source record into a fixed output shape.

Yield instead when:

  • the reusable product, audience, positioning, or measurement context itself

needs a versioned decision: use Product Context;

  • the outcome, scope, authority, or acceptance for this task is still unclear:

use Task Protocol;

  • a prior task stopped before its evidence was visible: use Interruption

Checkpoint;

  • a request, reply, and expected result already exist and need a controlled

repair: use Communication Failure Triage;

  • the changing item is a current fact that needs a source check: use Source

Investigator; or

  • the work includes a file, dataset, tool, account, network request, shared

system, or external action: hand off to Task Protocol before preparing a current-item brief.

Do not turn one repeated-chat pattern into a claim about memory, context-window behavior, cost, persistence, automation, or a named product configuration.

Require a stable card and a current card

Collect only the following visible inputs. Mark an absent field missing; do not retrieve it from another chat turn or infer it from an earlier example.

Stable card — reusable within a named work stream:

  1. work_stream — the recurring job in plain language;
  2. criteria_revision — a version, date, or immutable reference to the rules;
  3. allowed_inputs — material that may be used for every item;
  4. forbidden_assumptions — facts, sources, permissions, or previous outputs

that must not be inherited; and

  1. response_shape — the required form of the result.

Current card — true only for this one item:

  1. item_id — a non-sensitive local label;
  2. item_input — the supplied current text or a minimal safe summary;
  3. item_change — what is new or different today;
  4. task_request — the one result requested now;
  5. acceptance_evidence — the visible rule or artifact that will check it; and
  6. authority_and_risk — R0 text-only preparation, or handoff_required.

Reject a brief when the item contains a secret, private record, unlicensed source text, unsupported factual claim, or an unapproved action. Do not ask for unnecessary prior conversation history.

Compare before writing

  1. Identify every field that belongs to the stable card and every field that

belongs only to the current item.

  1. Preserve an earlier example only as a labelled reference. It is not a fact

about the current item and is never an acceptance result.

  1. Mark any current fact, permission, source, deadline, destination, or

acceptance check that was not supplied again as missing or not_authorized.

  1. Stop if the current item changes the stable criteria. Do not silently amend

the reusable card; hand off to its owner or to Product Context/Task Protocol as appropriate.

  1. Return a ready-to-copy brief only for R0, supplied-text work. The later

action still needs its own boundary and evidence.

Return one handoff receipt

Return exactly:

text
handoff_status: ready_for_text_only_current_item | blocked_on_<field> | handoff_required
work_stream:
criteria_revision:
stable_card:
current_item:
item_change:
allowed_inputs:
forbidden_inheritance:
requested_response_shape:
acceptance_evidence:
authority_and_risk:
unknowns_or_conflicts:
next_owner_or_action:
claim_limit:

Use ready_for_text_only_current_item only when the stable and current cards, request, response shape, acceptance evidence, and R0 boundary are all visible. The receipt is a context boundary, not evidence that a model retained rules, understood the item, generated a correct answer, or completed the next task.

Failure checks

Stop or hand off when:

  • a requester says "use the same rules as last time" but cannot name the

criteria revision or current acceptance check;

  • an old example silently becomes today's source or truth;
  • a current item includes files, credentials, private material, browsing,

publication, spending, account changes, or another external effect;

  • the current item changes the stable rubric, permission, destination, or

output contract; or

  • a response is already being treated as complete. Use Evidence Review rather

than relabeling it as a handoff.

Maintenance record

  • source: original Prysai Lab method derived from the source-bounded recurring-item research record, Task Protocol, Product Context, and Interruption Checkpoint boundaries
  • license: original rewrite; official guidance and public reports remain reference-only under docs/sources/asset-register.md
  • owner: workflow-systems maintainer
  • version: 0.1.0
  • review_date: 2026-11-14
  • content_status: candidate

Security

  • Security scanPassed, 95 of 100Pattern checks on the skill's files.
  • Model reviewPassed

Quality evaluation

  • Task evaluationNot runNo evaluation yet. Run one in Skill Studio.

Source

Popularity

Repository stars
437
Repository forks
8
7-day trend score
30

Stars and forks are for the whole GitHub repository, not this skill alone. Not refreshed yet.

Works with

Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.

  • Claude Code
  • Cursor
  • GitHub Copilot
  • Windsurf
  • Codex
  • Gemini CLI
  • Cline
  • Amp
  • Roo Code
  • Goose
  • Continue
  • Kilo Code
  • Junie
  • Trae
  • OpenHands
  • Qwen Code
  • AI Agent
  • Tabnine Chat
  • Sourcegraph Cody
  • Amazon Q
  • Firebase Genkit
  • Sweep
  • Devin
  • Phind
  • Mentat
  • Claude Desktop
  • Aider
  • GPT Pilot
  • Plandex
  • bolt.new
  • v0
  • Replit Agent
  • Anse
  • TypingMind
  • BetterChatGPT
  • LibreChat
  • Warp AI
  • Pieces
  • Fig AI
Embed badge
vskill VERIFIED badge
[![vskill VERIFIED](https://verified-skill.com/api/v1/skills/prysai/prysai-llm-playbook/prysai-shift-handoff/badge)](https://verified-skill.com/skills/prysai/prysai-llm-playbook/prysai-shift-handoff)

Something wrong with this skill or its page? Tell us.