diff --git a/src/crispr_analysis_utils/skills/cau-workflow/SKILL.md b/src/crispr_analysis_utils/skills/cau-workflow/SKILL.md
new file mode 100644
index 0000000..3204d63
--- /dev/null
+++ b/src/crispr_analysis_utils/skills/cau-workflow/SKILL.md
@@ -0,0 +1,102 @@
+---
+name: cau-workflow
+description: >-
+  Route a CRISPR guide library through crispr-analysis-utils from designed
+  spacers to an IGVF submission, in the right order. Use when a user names a
+  goal rather than a step, such as submitting a library to the IGVF portal,
+  preparing the guide_metadata.tsv the CRISPR Perturb-seq pipeline reads,
+  finding out what their guides target, or checking a design's own labels.
+  Also when a step's inputs are not on disk yet and the step that produces
+  them has to run first, or when a user asks which of the cau commands they
+  need. Covers the order, what each step consumes and produces, the keys the
+  tables join on, and what to settle with the user before anything runs.
+---
+
+# The crispr-analysis-utils workflow
+
+Four commands, run in order. Each reads the previous one's output, so the
+order is a data dependency, not a convention. The per-step skills hold the
+detail; this one decides which step the user is actually asking for.
+
+| Step | Command | Skill |
+| --- | --- | --- |
+| 1. Where does each guide match the genome? | `cau guide-alignment run` | `cau-guide-alignment` |
+| 2. What gene is each site in? | `cau guide-annotation run` | `cau-guide-annotation` |
+| 3. Which element of the screen does it test? | `cau element-mapping run` | `cau-element-mapping` |
+| 4. Write the IGVF table | `cau igvf-export run` | `cau-igvf-export` |
+
+Two modules sit outside the chain and need no predecessor:
+`cau-accession` (a lab-independent id encoded from the protospacer) and
+`cau-normalization` (counts per million, for a screen's count matrix).
+
+## The chain
+
+```bash
+cau guide-alignment run \
+    --guides guides.tsv --reference hg38.fa --outdir results/
+
+cau guide-annotation build-db gencode.v43.annotation.gtf.gz \
+    --out gencode.v43.protein_coding.parquet
+
+cau guide-annotation run \
+    --alignment results/alignments/guide_genome_annotation.tsv \
+    --store gencode.v43.protein_coding.parquet \
+    --promoters gencode.v43.protein_coding.TSS1bp.bed6 \
+    --outdir results/
+
+cau element-mapping run \
+    --alignment results/alignments/guide_genome_annotation.tsv \
+    --annotation results/annotations/guide_gene_annotation.tsv \
+    --elements elements.bed \
+    --outdir results/
+
+cau igvf-export run \
+    --alignment results/alignments/guide_genome_annotation.tsv \
+    --mapping results/elements/guide_element_mapping.tsv \
+    --elements elements.bed \
+    --store gencode.v43.protein_coding.parquet \
+    --promoters gencode.v43.protein_coding.TSS1bp.bed6 \
+    --genes measured_element_gene_pairs.tsv \
+    --outdir results/
+```
+
+`build-db` is a one-off per GENCODE release, not per library. Reuse the
+parquet store and the TSS file across runs and across steps 2 and 4.
+
+## What flows between the steps
+
+`guide_genome_annotation.tsv` is the spine: one row per library entry per
+site, and the only table with the whole picture. Everything downstream reads
+it, and steps 3 and 4 read it directly rather than through step 2's output.
+
+A guide with two sites has two rows, so joining on `guide_id` alone is
+ambiguous. The annotation table keys back on `guide_id` plus the four site
+columns; use all five.
+
+## Settle these before running anything
+
+Ask, do not guess:
+
+- **The reference and the annotation must be the same assembly**, and the
+  GENCODE release decides the ENSEMBL ids the IGVF table carries.
+- **Does the library have an added 5' G?** It changes what is aligned and
+  how a duplicate is judged. `cau-guide-alignment` has the rule.
+- **Is the element BED the screen's real element set?** An incomplete one
+  inflates `safe_targeting` silently rather than failing, so the count is
+  checked against the design before it is believed.
+- **Are there validated enhancers?** They are the one case where a design's
+  own label for an element outranks what the genome says.
+
+## Picking the step from what the user has
+
+Work back from the output they want, then forward from the files that exist.
+A step whose inputs are missing is a request for the step before it.
+
+- Only a guide table: start at 1.
+- Alignment output but no gene calls: 2.
+- Alignment plus an element BED, and the question is what each guide tests
+  or which guides are controls: 3.
+- All three, and the goal is a submission or the pipeline's input: 4.
+- A question about one guide's identity across libraries: `cau-accession`,
+  no chain needed.
+- A count matrix to normalize: `cau-normalization`, no chain needed.
diff --git a/tests/test_skills.py b/tests/test_skills.py
index 0de6991..633cec0 100644
--- a/tests/test_skills.py
+++ b/tests/test_skills.py
@@ -23,6 +23,9 @@
 
 # Which analysis modules each skill or agent covers.
 COVERAGE = {
+    # The workflow skill routes between the others: it covers no module of its
+    # own, and an empty tuple is how a skill says so.
+    "cau-workflow": (),
     "cau-accession": ("accession",),
     "cau-normalization": ("normalization",),
     "cau-guide-alignment": ("guide_alignment",),