PTN—16
Working Transcript
A live, inspectable log of the steps an AI takes — searches, tool calls, files touched — while it works.
When an AI works for minutes instead of seconds, a spinner is a trust vacuum. A working transcript replaces it with a narrated log: searching these three sources, reading this file, running these tests. The person can see what the system is doing, judge whether it is on track, and step in before a wrong turn compounds. This is the difference between delegating to a colleague and dropping a request into a black box.
The transcript is a summary channel, not a firehose. Each entry should be a legible headline — “Searched flight prices for March 14” — with the raw payload one click deeper. Products that dump full tool output into the transcript bury the one signal that matters: is this going the right way? Products that show nothing force people to wait out failures they could have caught in the first ten seconds.
Transcripts also do post-hoc work. After completion, the collapsed log becomes the audit trail: what was consulted, what was changed, what was skipped. That record is what lets someone verify an agent’s answer without re-doing the task — and it is what makes long-running autonomy tolerable at all. Design it to be skimmed twice: live at a glance, and afterward as evidence.
The pattern emerged when task duration outgrew the spinner — once agents ran for minutes and touched real systems, Claude Code and ChatGPT’s deep research converged on narrated logs because nothing else made the wait tolerable. The enduring tradeoff is narration depth: every step surfaced is a claim on attention, and a transcript tuned for engineers drowns everyone else. Maturity means designing the log as two products in one — a glanceable headline stream for supervision, and a durable, expandable record for audit — and treating it as compliance infrastructure, not chrome. When teams start citing transcripts in incident reviews, the pattern has graduated from UI treatment to system of record.
Anatomy
- Searched invoices for balances 30+ days past due
- Matched overdue customers to open jobs
- Drafted 12 reminder emails for review
An agent’s step log rendered as headlines, with the current step live and prior steps collapsed.
- 1Step headline. Each action is a one-line, human-readable summary, not raw tool output.
- 2Live step indicator. The current step is visually active, so progress reads at a glance.
- 3Expandable detail. Raw payloads — queries, diffs, command output — sit one click below each headline.
- 4Completed-state collapse. On finish, the transcript folds to a summary line and the answer takes the stage.
- 5Failure surfacing. Errors and retries appear inline in the log rather than being smoothed over.
When to use
- Agentic tasks that run longer than ~10 seconds or take multiple steps.
- When the AI touches external systems — files, browsers, APIs — and people need to know what.
- When mid-task intervention is possible and a wrong turn is cheaper to catch early.
- When the output alone cannot be verified without knowing how it was produced.
Design considerations
- 01Write step headlines in plain language — name the source, file, or query, not the tool ID.
- 02Show the current step live and collapse completed steps; the log should scan in two seconds.
- 03Put raw tool input and output behind a per-step disclosure, never in the default view.
- 04Surface errors and retries in the transcript instead of silently absorbing them.
- 05Collapse the whole transcript when the task completes, but keep it expandable as the audit trail.
- 06Timestamp or duration-stamp long steps so a stall is distinguishable from slow progress.
- 07Keep the stop and steer controls adjacent to the transcript — watching implies intervening.
Pitfalls
- ✕Dumping raw JSON tool calls into the transcript, which reads as noise and trains people to ignore it.
- ✕Vague step labels — “Thinking…”, “Working…” — that narrate nothing and rebuild the black box.
- ✕Hiding failures and retries, so the final answer arrives with no hint that step three silently broke.
- ✕Leaving the full transcript expanded after completion, forcing a scroll past scaffolding to find the answer.
- ✕A transcript you can watch but not act on — visibility without a stop button is just anxiety.
In the wild
- Claude
- Agentic coding sessions narrate each file read, edit, and command as a collapsible step, with raw output one level down.
- Cursor
- Agent mode logs codebase searches, file edits, and terminal runs as discrete inspectable steps in the chat pane.
- Perplexity
- Deep Research shows the sources being read and the reasoning steps in progress before the report lands.
- ChatGPT
- Deep research and agent tasks expose an activity log of browsing and analysis steps the user can open while the run continues.