Home ｜ Articles ｜中文版

When AI Feels Like It Changed Its Mind: Passing Agent Work Modes Across Conversations

Prompt Engineering / Agent Workflow / Mode Handoff · Turning AI from a Q&A box into a reliable working partner.

Old chat: smooth but heavy

The shared rhythm lives inside the context, but the token bill and hidden state keep growing.

Context bloatCompressed detailsStale assumptions

Mode prompt

New chat: clean and still familiar

You carry over the work spirit, decision boundaries, and reporting habits instead of the whole transcript.

Callable rolesAuditable boundariesConsistent style

CaptainPlans and delegates

ScoutReads and maps risk

BuilderShips narrow changes

ValidatorChecks with evidence

During long project work, switching to a new AI conversation can feel surprisingly painful. Staying in the same chat also gets risky: the context grows long, key details may be compressed away, and the token cost starts feeling like a meter that keeps climbing.

The dilemma is real. A fresh conversation is cheaper and cleaner, yet the new assistant may lose the rhythm that made the previous one useful. The result can feel like the AI changed its mind: the same model, but no longer the same working partner.

The goal is not to preserve every word of the old conversation. The goal is to preserve the way the AI should work with you.

The practical answer is to turn good collaboration behavior into a reusable mode prompt. Instead of dragging an entire transcript forward, you extract the agent's work spirit into a small operating contract.

1. The real break is not memory; it is work rhythm

When a new conversation starts poorly, the missing part is often not a fact. It is the working rhythm: how proactive the assistant should be, when it should stop, how it should report, and what kind of judgment it is expected to use.

Surface symptom	What actually broke
The new AI does not know what happened	State handoff is incomplete
The AI becomes timid	The decision role was not assigned
It asks instead of organizing	The operating mode was not transferred
The tone feels wrong	Your preferred collaboration style was not specified

A good handoff needs two layers: state and spirit. State tells the next AI where the work stands. Spirit tells it how to move.

2. Mode prompts turn good behavior into a reusable contract

A mode prompt is more concrete than a persona. It defines the role, authority, stop conditions, output format, and forbidden moves for a specific kind of work.

RoleWho should the agent be in this round?

BoundaryWhat can it decide, and where must it stop?

FormatHow should it report so the work remains reviewable?

A mode is an operating contract

A useful mode prompt translates vague guidance such as "be more proactive" into specific behavior: recommend a route, name risks, stop before irreversible actions, and report validation results.

This lets a fresh conversation feel familiar without inheriting the entire old context.

A small example

You are the project captain for this conversation.
Start with a recommendation, then explain the reasoning.
You may split work, assign short-lived helpers, and block unsafe routes.
Stop before irreversible actions such as merge, push, deletion, or broad source changes.
Every report must include: what you did, what you did not touch, validation results, risk changes, and the next recommended step.

The value comes from making the work personality concrete: proactive, bounded, and auditable.

3. Different agents need different spirits

A large task should not rely on one universal assistant. The steadier pattern is to split work into modes, then give each mode its own behavior contract.

CaptainChooses routes, delegates, and protects boundaries.

ScoutReads, maps, and reports without mutating state.

BuilderImplements a narrow assignment without expanding scope.

ValidatorTrusts reproducible evidence over verbal confidence.

CuratorTurns scattered reports into a decision-ready summary.

Agent mode	Core spirit	Best for
Captain	Proactive judgment with clear boundaries	Planning, sequencing, delegation
Scout	Read-only precision	Route checks, file discovery, scope scans
Builder	Constrained execution	Small implementation, docs, templates
Validator	Evidence-first skepticism	Dry-runs, tests, encoding checks
Curator	Compression into decisions	Handoff notes, risk tables, summaries

The same model can behave very differently depending on the mode. A scout with too much authority becomes dangerous. A captain with no authority becomes a passive clerk.

4. Preserve decision boundaries, not the whole transcript

The most useful part of a good collaboration is often the boundary logic: what the AI may do freely, what requires confirmation, and what should never happen silently.

Good proactive moves

Recommend one route and explain why
Split an oversized card
Assign a short-lived helper
Reject a risky merge path
Produce a reusable instruction block

Stop and ask first

Merge, rebase, or push
Delete files, branches, or worktrees
Clean residue
Touch broad source areas
Replace evidence with a claim

Once these boundaries are explicit, the agent can be bold without becoming reckless.

5. Capture your preferred style as rules

If an AI felt especially easy to work with, do not only save the conclusion. Save the behavior that made it work.

Preference	Mode rule
You like decisive analysis	Start with the recommendation, then list alternatives.
You dislike raw tool dumps	Summarize the relevant fields instead of pasting everything.
You worry about accidental edits	Report what was not touched and stop before irreversible actions.
You manage multiple agents	Split roles first, then collect reports into one decision summary.
You want a consistent voice	Use a fixed reporting structure and decision order.

Practical note: do not hand the next AI the whole mountain of history. Hand it the current state and the work spirit.

6. A repeatable mode-handoff loop

From one good collaboration to a reusable work mode

ObserveNotice what made the conversation productive.

DistillTurn the behavior into rules.

NameName the mode: Captain, Scout, Validator, Curator.

EncodeWrite role, boundaries, format, and forbidden moves.

InvokeStart the next conversation by calling the mode.

RefineAdd one useful lesson after each strong collaboration.

This gives you a practical way to switch conversations without losing the working relationship that made the previous one valuable.

7. Avoid writing mode prompts as slogans

Too vague	More useful
Be proactive.	Recommend one route first, then give alternatives and risks.
Act like an expert.	Name dependencies, rollback paths, and whether the task should split.
Manage the team.	Assign Route Reader, Scope Sentinel, Validator Patrol, and summarize their results.
Do not mess things up.	Do not merge, push, delete, or clean residue without explicit approval.

A good prompt makes the agent's behavior predictable. That predictability is what makes the next conversation feel continuous.

8. The short version

Conversations end. Context gets compressed. Token cost grows. A new AI may not know how to work with you yet.

Prompt engineering becomes most powerful when it can train, summon, and pass on the spirit of a useful working partner.

The goal is not perfection. The goal is a repeatable rhythm that helps every new conversation start closer to the way you actually work.