| name | anti-legacy:prd |
| description | Render a detailed, stakeholder-facing Product Requirements Document from the requirements graph (the resolved-or-flagged spine) plus the coverage report — domains, per-requirement business rules / validations / error paths, entity tables, the explicit dropped-scope list, a coverage-&-gaps section, and a req_id → legacy_components → rule-id traceability appendix. Deterministically derived (no LLM prose), registered as a manifest artifact, never advances the phase. Use when: "generate the PRD", "product requirements document", "detailed requirements deliverable".
|
anti-legacy:prd
The PRD is a deliverable: a polished, product-/business-facing rendering of
what the pipeline already decided. It is the readable face of
requirements_graph.json — one document a product owner, BA, or sponsor can
read end-to-end to see every capability the target system will implement, the
business rules behind it, what was deliberately cut, and where the analysis is
still soft.
Mental model
This skill does not think. It renders. The requirements graph (produced
by anti-legacy:graph-translator) is the source of truth; the PRD is a
deterministic projection of it for humans. Every requirement-bearing line traces
back the same thread the rest of the pipeline uses — req_id → legacy_components
→ business-rule ids (AGENTS.md §2). There is no free-written prose: edit the
upstream artifacts and re-render, never hand-edit the PRD.
It is a complement to anti-legacy:review-packet (which feeds GATE_1), not a
replacement — the review packet is the gate's working document; the PRD is the
standing stakeholder deliverable. It lands under .anti-legacy/deliverables/ and
registers itself; it never calls manifest advance (phase advancement is
owned by the phase skills).
The voice is factual (AGENTS.md §6): the Out of scope (dropped) section names
every dropped requirement WITH its reason, and the Coverage & gaps section
names requirements with no business rules, anything flagged review /
unresolvable, and every low-confidence rule. A clean-looking PRD that hides
holes is wrong.
Cross-Platform Notes
Pure standard-library Python via from antilegacy_core import deliverables as D.
Every path is built with os.path; no shell-isms. Runs identically on macOS,
Linux, WSL, and Windows. The output is plain Markdown with GitHub-flavored tables
— readable in any viewer, git diff, or text editor.
When it runs & prerequisites
Run it once the graph is ready — i.e. after
.anti-legacy/requirements/requirements_graph.json exists (post
graph-translate). Coverage and config are optional enrichments:
- Required:
requirements_graph.json with ≥ 1 requirement. Without it the
skill exits non-zero — it will not write a hollow PRD.
- Optional:
coverage-report.json (executive-summary coverage % +
mean_confidence; absent → "coverage not yet computed"), config.json (project
name, migration_mode, coverage.resolve_threshold for the low-confidence
flag — defaults to 0.75).
Confirm the graph is registered:
python3 .anti-legacy/run.py manifest status
You want to see requirements-graph listed. (Coverage is produced earlier by
anti-legacy:extraction; if it is not present yet, the PRD still renders.)
Parameters
The script reads the standard .anti-legacy/... locations by default; override
only when pointing at a non-standard path.
--requirements — path to requirements_graph.json
(default .anti-legacy/requirements/requirements_graph.json).--coverage — path to coverage-report.json
(default .anti-legacy/coverage-report.json).--config — path to config.json (default .anti-legacy/config.json).--no-register — write the PRD but do not touch the manifest (dry run / preview).
Step 1: Render the PRD
Run through the dispatcher. With no flags it reads the standard inputs, writes
the PRD, and registers it:
python3 .anti-legacy/run.py prd
Dry run (no manifest write) while you eyeball the output first:
python3 .anti-legacy/run.py prd --no-register
Step 2: Read what was produced
Open .anti-legacy/deliverables/product-requirements.md and confirm:
- Header names the project, the
migration_mode, the graph it was generated
from, and the timestamp.
- Executive summary reports the domain count and the requirement counts
(active / dropped / unresolvable) and, when
coverage-report.json is present,
the coverage % + mean rule confidence. If it reads "coverage not yet computed",
run anti-legacy:extraction first if you want that line populated.
- Requirements by domain — one section per domain; per active requirement a
title, description, Legacy components (the mandatory trace), a Business
rules table (RULE id · statement · confidence, with ⚠ on any confidence below
the resolve threshold), and Validations / Error paths tables where present;
plus an Entities table per domain.
- Out of scope (dropped) lists every dropped requirement with its
disposition_reason.
- Coverage & gaps names the soft spots: active requirements with no business
rules, anything
review / unresolvable, and every low-confidence rule.
- Appendix: traceability tabulates req_id → legacy_components → rule ids.
python3 .anti-legacy/run.py manifest status | grep deliverable-prd
Done-gate
Assert BEFORE trusting the deliverable:
requirements_graph.json exists and has ≥ 1 requirement. If not, the script
prints a stderr error and exits non-zero — it does not write a PRD. (Do not
paper over this: produce the graph with anti-legacy:graph-translator first.)- The written
product-requirements.md is non-empty before registration; the
script re-checks file size and refuses to register an empty file.
- With registration on (no
--no-register), the manifest carries a
deliverable-prd artifact (fmt markdown, depends_on: ["requirements-graph"])
and audit.jsonl has gained one anti-legacy:artifact-registered row. The
phase is unchanged — a deliverable never advances the pipeline.
If any assertion fails, surface the gap and stop — do not register a partial or
hollow document.
Output
- File:
.anti-legacy/deliverables/product-requirements.md
- Artifact:
deliverable-prd (fmt markdown, status: final, produced_by: anti-legacy:prd, depends_on: ["requirements-graph"]).
Still not done (callers should not assume)
- The PRD is only as rich as the graph. Thin
business_rules or missing
legacy_components upstream yield a thin PRD — the Coverage & gaps section
names exactly what is missing. Fix the graph (re-run extraction /
graph-translator) and re-render rather than editing the PRD.
- The PRD does not gate anything and does not replace GATE_1's review packet
(
anti-legacy:review-packet). It is a stakeholder artifact, not a sign-off.
- Coverage numbers come from
coverage-report.json; if that file is absent the
PRD says so rather than inventing a figure.
Failure cases
- "no requirements graph … (or it has no domains)" — the graph is missing or
empty. Run
anti-legacy:graph-translator (after extraction) and re-run.
- "requirements graph … has 0 requirements" — the graph parsed but contains
no requirement nodes. The upstream translate produced nothing to document —
investigate graph-translator before re-running.
- "manifest absent — PRD written but not registered" — you ran outside a
workspace (no
.anti-legacy/manifest.json). The PRD was still written; run from
an initialized workspace to register it.
- Executive summary says "coverage not yet computed" —
coverage-report.json
is absent. This is graceful, not a failure; run anti-legacy:extraction to
populate the coverage line.
