agentleFS
The-Artificer-of-Ciphers-LLCThe-Artificer-of-Ciphers-LLC/skills-from-the-artificer/adr-phase-coverage/SKILL.md

adr-phase-coverage

Skill6 starsChanged 33 days agoMITView on GitHub

Use when a user needs to turn a design doc, ADR, RFC, or spec into a phased implementation plan — or verify that an existing phase breakdown actually covers everything the design promised. The failure mode this catches: Phase N defers a deliverable to Phase N+1, Phase N+1 never claims it, and it ships as promised-but-not-built. Trigger when users say: 'break this into phases', 'turn this ADR/RFC into a roadmap', 'is my phase plan complete', 'what slips between phases', 'sanity check my phase breakdown', 'audit this epic for coverage gaps', or ask to sequence any architectural work into milestones. Also trigger on `/adr-phase-coverage`. Do NOT trigger for: authoring a new ADR, updating ADR rationale, writing a PRD, debugging a stuck deploy, or dependency analysis.

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 2 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    When a design document is sliced into implementation phases, the failure mode is rarely a phase doing its job badly — it is a deliverable …

    line 18
  2. Distinctive:

    Do this when decomposing a design into phases, or to audit an existing decomposition.

    line 47
  3. Very specific:

    Write a JSON trace mirroring assets/trace.example.json (a passing example you can copy).

    line 82
  4. Very specific:

    When unsure whether something is user_facing, ask: "is there a sentence anywhere — in the design, the PRD, or the docs — that a *user* …

    line 121

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

ADR phase coverage: nothing falls between phases

Why this exists

When a design document is sliced into implementation phases, the failure mode is rarely a phase doing its job badly — it is a deliverable that is nobody's job. Each phase has locally-coherent acceptance criteria, so each can honestly call itself "done," while a cross-cutting deliverable falls into the seam between two phase boundaries. The classic shape: Phase 4 builds the library and says "the user-facing command lands in Phase 5"; Phase 5's actual scope is something else; the command is now owned by no one — and the docs cheerfully describe it as shipping. (A real case study is in references/anti-patterns.md.)

This skill makes that gap mechanically visible before the plan is locked. You extract a small trace of decisions↔phases; a deterministic checker does the exhaustive cross-referencing humans get wrong. It is language-, framework-, and process-agnostic — it reads your decisions and your phases, whatever form they take.

The deep "why" — requirements traceability, walking skeletons, vertical slicing, the horizontal- layering trap — is in references/methodology.md (with citations). You don't need it to run the check.

The four gaps it catches

  1. Orphaned decision — an ADR decision owned by no phase.
  2. Dangling deferral — Phase N says "handled in Phase M," but Phase M neither owns nor accepts it. The hand-off points into the void.
  3. Multi-owner ambiguity — two phases both claim a decision; confirm it's a deliberate split, not two phases each assuming the other does it.
  4. Keystone unwired — a user-facing capability (a command/route/UI the user invokes) is owned, but no owning phase's acceptance criteria wire a user-reachable surface. A library can be fully built and still be unreachable; this forces an explicit "user can invoke X" acceptance.

Workflow

Do this when decomposing a design into phases, or to audit an existing decomposition.

1. Gather the two inputs

  • The design: the ADR / RFC / spec / design doc.
  • The decomposition: the phase breakdown / roadmap / epic — proposed or existing.

These live wherever your project keeps them (a docs/adr/*.md, an RFC in the wiki, a roadmap file, a tracker epic with sub-issues). See references/adapting-to-your-repo.md for how to locate decisions and phases across common conventions.

2. Extract decisions — and separate internal from user-facing

List every decision the design makes (numbered decisions, "Decision"/"Consequences" bullets, or requirements). Then do the step that prevents the keystone miss: add a row for every user-facing capability the design promises — every "user can …" story, every command/route/screen the design (or the docs describing it) says will exist. These are often not their own decision heading; they are implied by the decisions, which is exactly why they get orphaned. Mark them user_facing: true.

A reliable source is any documentation that describes the feature as shipping: if a doc says a command/endpoint exists, that is a user-facing capability some phase must own and wire.

3. Extract each phase's owns / accepts / defers

For each phase capture:

  • owns: the decision/capability ids this phase is responsible for delivering.
  • accepts: its acceptance criteria — what is observably TRUE when the phase is done. For a user-facing capability, at least one accept must name the wired surface ("user can run the foo command", "route /x returns …"), not just "the library exists."
  • defers: explicit hand-offs {to: <phaseId>, what: <id or text>} for anything this phase consciously pushes to a later phase. Capture every "see Phase N", "handled later", "deferred to" — those are the seams.

4. Write the trace file and run the checker

Write a JSON trace mirroring assets/trace.example.json (a passing example you can copy). Then:

python3 scripts/check_coverage.py <your-trace>.json

It prints the traceability matrix + findings and exits 1 if any ERROR exists (--json for a machine-readable result you can gate in CI). The matrix alone is valuable — a decision row showing owning phase(s): — NONE is the gap, at a glance. (Pure Python 3 stdlib; no dependencies.)

5. Fix every ERROR before locking the plan

  • orphan-decision → assign the decision to a phase's owns (with a real acceptance criterion).
  • dangling-deferral → go to the target phase and add the deferred item to its owns/accepts, by the same id/name. "Deferred to Phase M" is not coverage until Phase M says "I own this."
  • keystone-unwired → add an acceptance criterion to an owning phase that names the user-reachable surface and its wiring ("user can run X, dispatched to the Phase-K library"). If no phase wires it, you have found the next stranded keystone — schedule the owning work now.
  • multi-owner (warning) → confirm the split is intentional; otherwise consolidate ownership.

Re-run until the result is ✅. Then the decomposition is safe to lock.

6. After the owning phase ships (optional, code-layer)

The checker works at the planning layer (before code exists). Its real-code counterpart is a reachability / dead-code pass: confirm each new exported symbol is reachable from a registered entry point (CLI parser, HTTP router, event handler). If an export has no such path, the phase built a layer without wiring it. See references/methodology.md §4 for tools per ecosystem (Knip, ts-prune, Go deadcode, Vulture). Use both: the plan check prevents the gap; the code check proves it stayed closed.

Authoring the trace from messy docs

The script is deliberately dumb — it does exhaustive set-math, not reading comprehension. You do the reading: pull decisions and user-facing capabilities out of prose, infer the implied commands, and be honest about defers. The leverage is entirely in step 2 (naming the user-facing capabilities the design never gave their own heading) and step 3 (recording every "handled later" as a real defers edge). Skip those and the checker has nothing to catch.

When unsure whether something is user_facing, ask: "is there a sentence anywhere — in the design, the PRD, or the docs — that a user would read as 'I can do X'?" If yes, it's user-facing and needs a phase that wires it.

Files

  • scripts/check_coverage.py — the deterministic checker (Python 3 stdlib; no dependencies).
  • assets/trace.example.json — a passing trace to copy.
  • assets/trace.adr1244-broken.json — a real-world gap; run the checker on it to see a live catch.
  • references/methodology.md — RTM, walking skeleton / tracer bullet, vertical slicing, with sources.
  • references/adapting-to-your-repo.md — locating decisions & phases across ADR/RFC/roadmap/tracker conventions (with one worked example).
  • references/anti-patterns.md — the catalog, with a full case study end-to-end.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.