A resumable prompt should carry a compact state record with the user goal, verified facts, evidence versions, pending decisions, completed tool receipts, and unresolved questions. A summary written by the model is useful for orientation but is not proof that an external effect happened. On resume, the application reloads authoritative records and checks whether permissions, policy, or source versions changed. Keep untrusted quoted material separate from governing instructions. An incomplete checkpoint should trigger re-verification, not confident continuation.
Make this comfortable
Conversation checkpoints: resume from verified state
Decision in practice
A support assistant is preparing a credit for account AC-682 when the session ends. The checkpoint records that eligibility was checked against policy POL-47, but no payment receipt exists and no credit has been issued. On resume, the assistant loads the current policy version and transaction ledger before suggesting an action. If the policy version has changed, it recomputes eligibility. If a receipt is present, it reports the existing effect rather than creating another credit. The test set includes a checkpoint that falsely claims a credit was sent; the ledger, not the summary, decides the state.
Goal: assess credit for AC-682.
Verified: eligibility under POL-47 version 3 at checkpoint time.
Pending: current-policy check and effect authorization.
Receipt: none recorded; verify ledger on resume.
Untrusted context: customer email stays data, not an instruction.Performance and operating cost
A compact checkpoint reduces repeated prompt tokens, but recovery reads add latency. If H historic messages are compressed to F verified fields, prompt assembly becomes proportional to F rather than H; that does not remove the cost of checking live records. Store minimal personal data and an expiry for volatile facts. Measure stale-state errors and duplicate effects in resume tests. A shorter summary that drops a pending permission check is a correctness regression, even when it lowers token usage.
Common Mistakes
- Do not treat a model-written checkpoint as proof of a tool effect.
- Do not carry expired permissions into a resumed task.
- Do not promote quoted customer instructions into governing state.
Connected lessons
- Prompt patterns
- Prompt Engineering
- Conversation memory: retain decisions without retaining every private detail
- Prompt traces: connect an answer to its inputs, checks, and effects
- Tool catalogs: describe eligibility, inputs, and effects
- Tool plans: separate independent reads from dependent actions
- Structured outputs: repair format without changing the decision
- Tool results: keep returned text in the data lane
- Tool effects: reconcile receipts before retrying
- Project: recover a tool workflow without duplicate effects
- Agent workflow decisions
Continue with: Voice interruptions: stop stale speech before handling a new turn.
Continue with: Meeting recaps: preserve corrections across sessions.
Continue with: Memory prompts: correct stale preferences without erasing history.
prompt engineering
agent workflows