## Objective

Produce a compact but complete handoff that lets a fresh agent continue the actual work with minimal context loss, without replaying this entire conversation and without carrying forward stale conclusions.

Do not execute `/deep-save` yourself until after the continuation brief is completed. Prepare the conversation so `/deep-save` can save the right durable information.

## First, diagnose the conversation

Determine silently:

1. What outcome was the conversation actually pursuing?
2. What has been verified in the real world?
3. What was only proposed, inferred, attempted, or reported by another agent?
4. What changed during the conversation?
5. What remains unresolved?
6. Which decisions are durable?
7. Which earlier conclusions were superseded, corrected, or contradicted?
8. What is the narrowest next substantive action?
9. What would make the next agent waste time, repeat a failure, or take the wrong action?

The fluent failure to avoid is a polished summary that lists activity but does not tell the next agent what is true, what is stale, what remains, and what to do first.

## Evidence rules

Use current evidence over memory or conversational confidence.

Separate these states explicitly:

- `VERIFIED`: directly confirmed by a current read, command, test, live response, or user statement.
- `INFERRED`: strongly suggested but not directly confirmed.
- `ATTEMPTED`: action was taken, but the outcome was not proven.
- `UNVERIFIED`: relevant fact has not been checked.
- `BLOCKED`: a real permission, capability, or wrong-objective boundary exists.
- `SUPERSEDED`: an older claim or plan no longer governs.

For every blocked item, identify the exact attempt, specific failure evidence, changed approaches already tried, boundary type, owner, and exact action that would unblock it. A first error is not a boundary.

Do not repeat or persist secrets. Exclude passwords, API keys, bearer tokens, OAuth tokens, session cookies, private URLs, tokenized links, private personal data, and credential-bearing command output. Replace them with labels such as `[credential configured]` or `[token omitted]`.

## Required handoff contents

Write the handoff in the most useful durable location available. If a project-local continuation-brief location exists, use it. If no destination exists, provide the handoff directly in the response and state that no local artifact was created.

Use a new dated file when creating a local artifact. Do not overwrite another agent's work. Preserve the local naming convention if one is discoverable. If no convention is discoverable, use:

`continuation-brief-<topic>-<agent>_MMDDYY.md`

The handoff must contain these headings:

1. `Active Objective`
2. `Current Verified State`
3. `What Changed`
4. `Decisions and Durable Rules`
5. `Superseded or Stale Claims`
6. `Completed Work`
7. `Open Work in Dependency Order`
8. `Boundaries and Owners`
9. `Sources of Truth`
10. `Verification Evidence`
11. `Anti-Drift Guardrails`
12. `Immediate Next Action`
13. `Fresh Chat Resume Prompt`

## How to write each section

### Active Objective
State the desired state of the world in one sentence. Do not state only the task or the last command.

### Current Verified State
List the important facts that are true now. Include exact paths, URLs, identifiers, versions, branches, revisions, counts, dates, or statuses when they affect execution. Mark each fact with its evidence source.

### What Changed
List only changes that affect future decisions or execution. Distinguish edits, deployments, external writes, read-only checks, and decisions.

### Decisions and Durable Rules
Capture accepted decisions, constraints, user preferences, architecture choices, and corrections that must survive into the next chat. Include alternatives rejected and why when that prevents drift.

### Superseded or Stale Claims
Name old reports, prompts, plans, or assumptions that must not be followed. For each, state the current replacement evidence.

### Completed Work
Include only work whose acceptance condition was verified. Do not call a file created, command run, or deployment attempted complete unless the underlying outcome was checked.

### Open Work in Dependency Order
For every remaining item include:

- item and objective
- exact path, URL, record, or system
- status: broken, missing, required, unverified, or blocked
- dependency
- acceptance test
- owner
- whether it is safe to execute now

Start with the smallest next move that reduces uncertainty or advances the objective. Keep unrelated backlog out of this section.

### Boundaries and Owners
Separate:

- safe work the next agent can do without asking
- work requiring user approval
- work requiring credentials or dashboard authority
- destructive, irreversible, public, financial, privacy-sensitive, or production-changing work
- capability limits that remain after alternate approaches were tried

### Sources of Truth
List the current source files, live systems, trackers, documents, and commands that govern the work. Mark historical material as reference only. Do not provide tokenized URLs.

### Verification Evidence
Give a concise evidence table with:

- claim
- source or exact check
- result
- date or timestamp when relevant
- state
- remaining uncertainty

### Anti-Drift Guardrails
State what the next agent must not assume, repeat, overwrite, delete, publish, send, or redesign. Include the known fluent failure for this work.

### Immediate Next Action
Give one action only. It must be concrete, narrow, executable, and owned by a specific party.

### Fresh Chat Resume Prompt
Write a copy-ready prompt for a fresh agent. It must:

- point to the handoff artifact or current conversation source
- state the active objective
- require a short current-state verification before edits
- identify the immediate next action
- preserve the approval and safety boundaries
- require execution rather than another planning summary
- require verification of the underlying outcome
- require a final report separating complete, unverified, and blocked work

Do not ask the fresh agent to reread the full conversation unless a specific missing fact cannot be recovered elsewhere.

## Final quality gate

Before returning the handoff:

1. Read the created artifact back if one was written.
2. Confirm every required heading is present.
3. Confirm the active objective and immediate next action are unambiguous.
4. Confirm every completion claim has evidence.
5. Confirm stale claims are labeled and corrected.
6. Confirm every open item has an owner and acceptance test.
7. Confirm no secret, bearer token, tokenized URL, or private credential appears.
8. Confirm the handoff is sufficient for a fresh agent without requiring a planning question.
9. State what was verified and what remains unverified.

## Response contract

Return:

- the handoff location, or state that it exists only in the response
- the active objective
- the immediate next action
- the number of completed, open, unverified, and blocked items
- the verification performed
- any real boundary and its owner

Do not end with a vague promise. Do not claim `/deep-save` was completed unless you actually executed it and verified its durable writes.

Once the continuation brief is completed, run /deep-save.