temporal-awareness

01The wrong reference class

why the estimates are off by orders of magnitude

Claude's instinct for "how long" is trained on an enormous corpus of human-written estimates — Jira tickets, blog posts, retros, textbooks — all measured in developer-days. That world assumes a person types the code, context-switches between meetings, looks up docs, gets tired, and works an eight-hour day.

An agent pays almost none of those costs. The reading, writing, and wiring that dominate a human's estimate compress dramatically. So the model isn't computing a wrong answer — it's fluently echoing the wrong reference class. The result, on a log scale:

02Split what collapses from what doesn't

the core move — accurate, not just smaller

The fix is not "say a smaller number." That would just replace one wrong answer with another — because some things genuinely don't compress. The move is to decompose every estimate into two buckets, and treat them completely differently.

03The data was already there

Claude's own /goal completions, hiding in the transcripts

Here is the part that turns a reasoning trick into a calibrated instrument. Every time you run /goal, Claude Code records the completion in the session transcript — including an authoritative duration field, goal_status.durationMs, computed by the harness itself. It was found by forensic search through transcripts and debug logs; the harness never advertises it.

Raw wall-clock is a trap — inflated 5–25× by overnight and idle gaps. But completed goals (met == true) are clean, because a goal only completes while it's being actively worked. The result is a small corpus of real durations — and it confirms the whole thesis:

The corpus doesn't just correct the bias — it proves it. Claude's real tasks take minutes.

04From transcript to prior

a SessionStart hook, running while you're not looking

Keeping the corpus fresh has to be cheap, or it can't run at session start. So the hydrate is incremental (only re-scans transcripts touched since last run — ~0.1s steady state) and backgrounded (never blocks startup). It mines new completions, classifies each goal once with a stripped-down headless Haiku call, and pre-digests everything into a compact summary the skill reads instead of the raw rows.

flow: top → bottom · everything downstream of the harness's own duration field

05A prior, not a lookup table

the strong signal is scale — the medians are soft

The temptation is to treat a bucket's median as the answer. The skill is explicit that this is wrong. The corpus's strong, trustworthy signal is scale: this work is minutes, not days. The per-bucket medians are much softer, because the categories are coarse — a single bucket holds wildly different tasks.

So the skill reads a bucket's spread first, median second, then reasons about this task's context — scope, severity, blast radius, ambiguity — and places the estimate on, or beyond, the range. A wide p10–p90 is not noise to average away; it's the signal that context, not the median, decides the number.

Bugfixes here have typically run ~7 min — but a dropped prod DB is a different animal; recovery and verification push this well past that. An off-by-one would be under it. — the shape of a well-formed estimate

06Closing the epistemic gap

the family resemblance to context-awareness

There's a recurring pattern across the agent-autonomy plugins: Claude Code logs a great deal about its own operation, and Claude — the model inside it — rarely looks. Context budget sits in the status line the model never reads. Goal durations sit in transcripts the model never mines.

temporal-awareness closes that loop for time the same way context-awareness does for context budget: it takes a signal the harness already produces and hands it back to the model in a form it will actually use. The estimates stop being a borrowed human intuition and start being grounded in Claude's own measured history.