diff --git a/session-scripts/layout.py b/session-scripts/layout.py
index 7577220..99bdeb9 100644
--- a/session-scripts/layout.py
+++ b/session-scripts/layout.py
@@ -85,6 +85,13 @@ def path(*parts, make=True):
     correctly. The last component is the one that decides.
     """
     name = parts[-1]
+    if name in _DIRS:
+        # Asking for a directory by name returns the directory itself, not a
+        # file of that name nested inside it.
+        full = os.path.join(SESSION, _DIRS[name])
+        if make:
+            os.makedirs(full, exist_ok=True)
+        return full
     full = os.path.join(SESSION, subdir(name), name)
     if make:
         os.makedirs(os.path.dirname(full), exist_ok=True)
diff --git a/session-scripts/subdir_readmes.py b/session-scripts/subdir_readmes.py
new file mode 100644
index 0000000..08ab53d
--- /dev/null
+++ b/session-scripts/subdir_readmes.py
@@ -0,0 +1,117 @@
+"""Drop a short README in each session subdirectory.
+
+Each one says what the directory holds, where it came from, and whether it is
+safe to delete - the three things someone opening a folder cold needs.
+"""
+import os
+
+ROOT = r"C:\Users\lhorrocks-barlow\Downloads\NGC5128\20260721"
+
+READMES = {
+    "raw": """# raw/
+
+Exactly what iTelescope delivered, untouched.
+
+- `calibrated-*.zip` - the calibrated frames, as archives (FITS and TIFF)
+- `raw-*.zip` - the uncalibrated frames, kept for completeness
+- `jpeg-*.jpg` - iTelescope's own preview of each sub
+
+**Do not edit anything here.** This is the archive copy; everything else in the
+session can be rebuilt from it. `calibrated/` holds these same frames
+uncompressed.
+""",
+    "calibrated": """# calibrated/
+
+The 24 calibrated subs, uncompressed from `raw/`. 12 Luminance, 4 Red,
+4 Green, 4 Blue, all 300 s, all BIN2.
+
+`.fit` are the data the processing uses; `.tif` are the same frames in a form
+other software can open.
+
+Already bias, dark and flat corrected by iTelescope before delivery
+(`CALSTAT = 'BDF'` in the headers) - that cannot be undone here.
+
+Safe to delete: regenerable from `raw/` by re-running `unzip.py`.
+""",
+    "stacks": """# stacks/
+
+The combined frames, at two levels of processing.
+
+- `masters/` - the real stacks: registered, transparency-normalised, combined
+  with outlier rejection, and plate-solved. Everything downstream uses these.
+- `original/` - the baseline: aligned and averaged, and **nothing else**. No
+  rejection, no sky subtraction, no normalisation. Satellite trails and the
+  moon's gradient are still in it, deliberately. Useful when a processed result
+  looks odd and you need to know whether the data or the processing caused it.
+
+Both sets are pixel-aligned to the same reference frame, so they can be
+compared or subtracted directly.
+""",
+    "final": """# final/
+
+**The finished images.** Four renderings of the same data, differing only in
+how much processing was applied, plus a side-by-side comparison.
+
+| File | What it is |
+|---|---|
+| `NGC5128-final-1-stacked` | stacked and stretched, nothing else |
+| `NGC5128-final-2-processed` | conventional processing |
+| `NGC5128-final-3-best` | corrected by what the analyses established |
+| `NGC5128-final-4-closeup` | image 3 cropped to the galaxy, 23.4' square |
+| `NGC5128-final-comparison.jpg` | all three side by side, with a 1:1 crop |
+
+Each comes as `.png` (viewing), `.tif` (16-bit, for editing or printing) and
+`-preview.jpg` (small).
+
+The reasoning behind each is in `../METHODS.md`.
+""",
+    "renderings": """# renderings/
+
+Finished images that are not the four deliverables in `final/`.
+
+- `NGC5128-LRGB.*` - the first full composite
+- `NGC5128-LRGB-hdr.*` - the same with the core recovered by a second tone curve
+- `NGC5128-img-deconvolved.*` - sharpened luminance
+- `NGC5128-img-annotated.jpg` - coordinate grid and catalogued objects, placed
+  by the plate solution. Includes SN 1986G and SN 2016adj, both real supernovae
+  in this galaxy.
+- `NGC5128-img-starless.png` / `NGC5128-img-stars.png` - the field split into
+  nebulosity and stars
+- `NGC5128-img-core-print.jpg` - a print crop
+
+Kept because they show intermediate stages and because some are useful in their
+own right. `final/` is what to look at first.
+""",
+    "science": """# science/
+
+The measurements, as opposed to the pictures.
+
+- `notes/` - **start here.** Three write-ups covering the globular cluster
+  survey, the surface photometry, and the transient and moving-object search.
+  Each states its method, its numbers, and its limits.
+- `figures/` - the plots those notes refer to
+- `catalogues/` - the measured tables (CSV): cluster candidates, isophote
+  profiles, flagged objects
+- `data/` - the galaxy model, star and dust masks, extinction map, and a plain
+  text summary of derived quantities
+
+Headline results are summarised in `../METHODS.md`; the notes carry the detail,
+the caveats and the things that did not work.
+""",
+    "intermediates": """# intermediates/
+
+Caches and scratch files: downloaded star catalogues, detection lists, partial
+results, run logs.
+
+**Safe to delete.** Everything here is regenerated by re-running the scripts,
+at the cost of re-querying Gaia, SIMBAD and VizieR. Kept so a re-run is fast
+and so the exact catalogue data used is preserved rather than silently changing
+under a later query.
+""",
+}
+
+for name, text in READMES.items():
+    path = os.path.join(ROOT, name, "README.md")
+    os.makedirs(os.path.dirname(path), exist_ok=True)
+    open(path, "w", encoding="utf-8").write(text)
+    print(f"wrote {name}/README.md")