diff --git a/skills/observe/SKILL.md b/skills/observe/SKILL.md
new file mode 100644
index 0000000..f16ed32
--- /dev/null
+++ b/skills/observe/SKILL.md
@@ -0,0 +1,60 @@
+---
+name: observe
+description: >
+  Watch a named target (a CI run, a pipeline, a build, a log) on a fixed
+  cadence and apply the triage skill whenever new failures appear. Trigger on
+  /observe <target> or when the operator asks to keep monitoring something and
+  fix regressions as they surface. Portable across projects.
+---
+
+Keep a target under continuous watch and act only when it degrades. The loop is
+the watcher; triage is the response. Do not re-investigate failures already
+resolved this session, and do not idle-poll work the harness will notify you
+about.
+
+## Before you start: fix the target
+
+The operator names one concrete target: a CI run id or url, a pipeline, a
+branch's checks, a build, or a log file. Restate it in one line and record the
+current failure set as the baseline (the failures already known or already
+fixed this session). Everything the watch reports is diffed against that
+baseline, so a stale baseline re-triages solved failures.
+
+## The loop
+
+Drive this with `/loop` in dynamic mode (no fixed interval unless the operator
+gave one): after each observation, schedule the next wake-up yourself and pass
+the same `/observe` prompt back so the next firing repeats the watch.
+
+1. **Poll the target.** Pull its current state (job list, check conclusions, the
+   log's tail). Enumerate the failures exactly as the `triage` skill's step 1
+   does: failure, cancelled, timed-out; a downstream aggregate that fails only
+   because an upstream did is not its own failure.
+2. **Diff against the baseline.** A failure is *new* only if it is not in the
+   baseline and was not already triaged this session. If nothing is new, log one
+   short line ("still green" / "N known failures, no change") and reschedule.
+3. **Triage each new failure.** Invoke the `triage` skill on the new set only —
+   which applies `dialectic` per failing job to a ~99% root cause and a real
+   fix. Never mask (no retry-until-pass, no disabled check, no soft-skip). Fold
+   each newly fixed failure into the baseline so it is not re-triaged.
+4. **Reschedule.** Pick the next delay from what you are waiting on: an active
+   run that changes every few minutes gets a check matched to its cadence; a
+   quiet or long-running target gets a long fallback (1200s+). Keep watching
+   until the target reaches a terminal state (run completed, pipeline finished)
+   AND every failure it produced has a verified fix, then stop the loop.
+
+## Rules
+
+- The loop is a fallback timer, not a busy-wait: if the harness will notify you
+  when tracked work finishes (a background task, a spawned agent), wait for that
+  notification instead of polling for it; reserve the timed wake-up for external
+  state the harness cannot see (a remote CI run, a deploy queue).
+- Executed evidence beats reasoning: read the actual failing log line and the
+  artifact, not the job title — triage and dialectic carry this through.
+- One baseline, updated in place: dedup every observation against all failures
+  seen so far, or a flaky-then-fixed failure reappears every cycle and the watch
+  never converges.
+- Report only on change: a new failure and its triage verdict, or a terminal
+  "target is green, watch closed". Silence between changes is correct.
+- Never invent results for a run still in progress; report what the poll
+  actually returned.