Imported from dimagi/open-chat-studio (.claude/skills/extract-adrs/SKILL.md). Install upstream with npx skills add dimagi/open-chat-studio --skill extract-adrs. Copyright stays with the author.

name: extract-adrs description: Extract Architecture Decision Records from a stable design or spec document. Use when a design doc is finished and its decisions should be crystallised into citable ADRs at docs/adr/. Refuses to run on docs with status: active frontmatter. Pass the source doc path as the argument.

Extract ADRs from a design doc

Drive a guided conversation that converts a stable design or spec document into multiple ADRs at docs/adr/, then either shrinks or deletes the source doc.

Argument: path to the source doc (e.g. docs/superpowers/specs/2026-03-23-elevenlabs-voice-provider-design.md).

Format reference: the canonical ADR template at .claude/skills/extract-adrs/templates/adr-template.md (and its twin at docs/adr/_template.md) defines the file structure. Status is an inline HTML <span> pill — there is no YAML frontmatter on ADR files. See docs/adr/0000-record-architecture-decisions.md for a worked example.

Phase 1 — Gate check

1. Read the source doc's YAML frontmatter (the leading --- block, if any).

2. Branch on status:

   • active → STOP. Tell the user the doc is marked active and refuse to proceed. Suggest they flip the status to stable if the design is settled.
   • extracted → Ask: "This doc has already been extracted. Re-extract?" — only proceed on explicit yes.
   • stable → continue to Phase 2.
   • No frontmatter at all → ask the user via AskUserQuestion to classify the doc. Options:
     • Active — still evolving (stop here; instruct user to add status: active frontmatter manually)
     • Stable — ready for extraction (continue to Phase 2)
     • Pure plan — just delete it (no decisions to extract; work has shipped; file is scaffolding)
     • Mixed — let me think (escape hatch — stop and let the user decide manually)

3. If the user picks "Pure plan — just delete it" → confirm once more, then git rm <path>, commit with message docs: remove shipped implementation plan <basename>, and exit.

Phase 2 — Candidate decisions

1. Read the source doc in full.
2. Identify candidate decisions. A decision is something that:
   • Was a real choice (multiple plausible options existed).
   • Has consequences that constrain future work.
   • Would be useful to cite from code or PRs later.
   • Would plausibly be superseded or revised independently of the others (the split-vs-fold test). A choice that exists only as a forced consequence of a bigger decision — a stub library dictated by the type-checker you picked, a serializer forced by your framework — is NOT its own ADR. Fold it into the parent decision's Decision/Consequences/Alternatives. Relatedness alone is not a reason to split; the extends: link captures relationships.
3. Present a numbered list to the user. Each entry must include:
   • Short title (5–10 words, kebab-case ready).
   • One-line summary of the decision.
   • Source section(s) the decision came from (heading or line number).
4. Ask the user via AskUserQuestion (multi-select where possible) to:
   • Drop candidates that aren't real decisions.
   • Merge near-duplicates.
   • Split overly-broad candidates.
   • Rename if the title is poor.
5. Iterate until the user confirms the final list.

Phase 3 — Per-ADR drafting

For each approved candidate one at a time (not batched):

1. (At the start of Phase 3 only) Allocate the next number:

   ls docs/adr/[0-9]*.md 2>/dev/null | grep -oE '[0-9]{4}' | sort -n | tail -n1

   Take the highest existing number, then reserve N+1, N+2, ... for this run. Show the user the planned numbering.

2. Read .claude/skills/extract-adrs/templates/adr-template.md once at the start of Phase 3 — use it as the model for every ADR you draft.

3. Draft the ADR. The structure follows the template exactly:

   • # ADR-NNNN: Short title heading.
   • <span class="adr-status adr-status-{lowercase-status}">{UPPERCASE-STATUS}</span> pill.
   • <p class="adr-meta">Author: {user name} · Created: {today's date as YYYY-MM-DD}</p> meta line.
   • Optional Extends: [ADR-NNNN](NNNN-prior-title.md) line (filled in during Phase 4).
   • ## Context — paraphrase the source doc's framing of why this decision was needed. Keep tight: one or two paragraphs.
   • ## Decision — a short "We will…" lead-in; one paragraph for a simple decision, or bullets when it has several parts.
   • ## Consequences — bullets, good and bad.
   • ## Alternatives considered — bullets, one line each, name + rejection reason.

   Content discipline. ADRs are immutable once accepted, so anything that decays must stay out. Inside the body: cut file paths, file.py:lineno references, private helper names (_foo, internal underscore-prefixed methods), code-walk paraphrases of the implementation, exact log strings, and dated migration filenames. Keep identifiers the decision creates as public contracts: model + field names, DB constraint / index names, enum values, waffle flag IDs, URL routes and query-parameter surfaces, and ORM lookup paths when the join is the decision (e.g. "filter through run.config, not evaluator"). Deciding heuristic: "If I rename this tomorrow, do I need a migration or just a refactor?" Migration → keep. Refactor → cut.

   Concision discipline. An ADR records the decision, not the implementation — keep it dense. Verbosity is the most common failure mode, so draft tight from the start rather than trimming later:

   • One idea per sentence. If a sentence chains three clauses with em-dashes or semicolons, split it or cut the weakest clause.
   • State each fact once. Don't restate the Context reasoning in Consequences or Alternatives — cross-reference it ("the main fork, covered in Context") instead of repeating it.
   • No editorializing or self-praise ("hits the right point on the rigidity/flexibility curve", "elegantly", "honestly", "the conservative default"). State the trade-off plainly and let it stand.
   • No mannered prose. Metaphor and flourish in place of direct statement — "a dial worth turning" for "a parameter worth varying", "earns its keep" for "still matters". Metaphors drag in connotations you did not choose. Say what you mean; when a literal phrase is available, use it.
   • Bullets are one sentence: each Consequence and Alternative is name → reason. If you're adding a clause to explain the explanation, cut it.
   • Soft budgets: Context ≤ 2 short paragraphs; ≤ 1 sentence of rationale per Decision bullet. A single-decision ADR rarely needs more than ~400 words; a complex multi-model one more than ~700.
   • Final pass: re-read the draft and delete every sentence that wouldn't change what a future engineer does. If cutting it loses no decision, constraint, or durable identifier, it was prose.

4. Use AskUserQuestion to confirm the ADR status:

   • accepted (default for extraction from shipped work).
   • proposed (decision is recorded but not yet implemented).
   • rejected (rare — captures a decision considered and turned down).
   • draft (still under discussion; unusual for extraction).

5. Verify behavioral claims against the code (for accepted ADRs). Every statement that asserts how the system actually behaves — data model, scoping (e.g. team-scoped vs global), defaults, control flow, named functions/modules/settings — must be checked against the implementation. Use Grep/Read to find the code and confirm each claim. The code is ground truth: where the source doc contradicts the implementation, correct the draft to match the code and flag the correction when you show the user. Show the verifying symbol (file + function) to the user when you walk through the draft so they can spot-check where you looked, and record it in the corrections summary if you flag a code-vs-doc fix. Do NOT pin file paths or :lineno refs into the ADR body — the ADR is immutable and those references rot on the next refactor (see Content discipline above). Skip this step for proposed/draft ADRs — there is nothing implemented to check yet.

6. Show the user the drafted ADR, noting any code-vs-doc corrections from step 5. Ask if they want edits before moving to the next candidate.

7. Do not write the file to disk yet — keep all drafts in conversation until Phase 5.

Phase 4 — Cross-references

After all candidates are drafted:

1. Scan the drafts. When ADR B's Context paraphrases or relies on something ADR A decided, B "extends" A.
2. Build the proposed extension graph as a bulleted list:
   • ADR-0042 extends ADR-0040
   • ADR-0043 extends ADR-0040, ADR-0042
3. Show the user. Ask via AskUserQuestion if the graph is correct (Yes / Edit it).
4. Apply confirmed Extends: lines into each ADR draft (immediately below the meta line):

   <p class="adr-meta">Extends: <a href="0040-...">ADR-0040</a></p>

   If extending multiple: separate with , inside the same <p class="adr-meta">.

Phase 5 — Write & wire up

1. Write each ADR file to docs/adr/NNNN-kebab-title.md. Write atomically per file using the Write tool.

2. Update the source doc per disposition rule:

   • Design doc with surrounding narrative (user stories, requirements, backlog mapping, etc.): replace the body (everything after the H1 title) with a brief overview paragraph plus a ## Decisions section listing the extracted ADRs as links. Set status: extracted in frontmatter (add the frontmatter block if missing).
   • Pure spec with no narrative beyond the decisions: delete the file with git rm.
   • Mixed / unclear: ask the user via AskUserQuestion which disposition to apply.

3. Update mkdocs.yml:

   • Locate the nav: block, then the Architecture: → Decisions: subsection.
   • Append entries for each new ADR in numeric order, using 6-space indentation to match existing entries:

           - NNNN Title: adr/NNNN-kebab-title.md

4. Update docs/adr/index.md:

   • Append a row per new ADR to the markdown table (after the existing rows, before EOF).
   • Use the format documented in the HTML comment at the top of the table:

     | [NNNN](NNNN-kebab-title.md) | <span class="adr-status adr-status-{lowercase-status}">{UPPERCASE-STATUS}</span> | Short title |

5. Run the verification build:

   • uv run --no-project zensical build --clean
   • Expected: completes with no new errors. (Pre-existing warnings on unrelated docs are OK.) If errors, surface them to the user; do NOT auto-commit.

6. Print a summary:

   • "Wrote N ADRs: 0042 …, 0043 …, 0044 …"
   • "Source doc: <shrunk to pointer block | deleted | left unchanged>"
   • "Updated: mkdocs.yml (nav), docs/adr/index.md (table)"
   • "Next ADR number for future runs: {N+1}"
   • "Review the diff and commit when ready."

Safety properties

• Never commit from inside the skill. The user reviews the diff and commits.
• No file writes until Phase 5. If interrupted between Phase 1 and Phase 5, nothing is on disk yet — safe to restart.
• Atomic per file in Phase 5. Each ADR file is written via a single Write call; the source doc is touched only after all ADR files are written.
• If the build fails in Phase 5 Step 5, surface the error and DO NOT proceed to print the summary as if successful. The user may need to roll back the ADR files (git clean -f docs/adr/) before retrying.

When NOT to use this skill

• The source doc has status: active. Mark it stable first.
• The source doc is purely an implementation plan with no decisions. Use plain git rm + commit; don't run this skill.
• You want to write a single new ADR by hand. Just copy docs/adr/_template.md to docs/adr/NNNN-kebab-title.md, fill it in, and add nav + index entries manually.

Worked example trigger

User: "Can you turn the elevenlabs voice provider spec into ADRs?"

Skill kicks off with the source path. Phase 1 detects no frontmatter, asks the user to classify, user picks "Stable — ready for extraction". Phase 2 proposes two candidate decisions ("Add ElevenLabs as a voice provider", "Map OCS voice config to ElevenLabs voice IDs"); user drops the second as too implementation-specific. Phase 3 drafts ADR-0001. Phase 4 has nothing to cross-reference. Phase 5 writes docs/adr/0001-add-elevenlabs-voice-provider.md, deletes the source spec, appends to mkdocs nav and the index. Build passes. User reviews and commits.