Design Doc Creator Skill

Design Doc Creator

Create well-structured design documents for AILANG features following the project's conventions.

Current State

Current version: !'cat std/VERSION'
Planned docs: !'ls design_docs/planned/ 2>/dev/null | head -10'
Recent implemented docs: !'ls design_docs/implemented/ 2>/dev/null | tail -3'
Active changelog: !'ls changelogs/ | grep current 2>/dev/null'

Use the data above first. Only re-run these commands manually if the injected context is empty or you need to refresh after making changes.

Quick Start

Most common usage:

# User says: "Create a design doc for better error messages"
# This skill will:
# 1. AUTO-SEARCH for related design docs (Ollama neural embeddings)
# 2. Show matches from implemented/ and planned/ directories
# 3. Auto-populate "Related Documents" section in template
# 4. Proceed automatically (no confirmation needed)
# 5. Create design_docs/planned/better-error-messages.md
# 6. Fill template with proper structure

Automatic Related Doc Search (v0.6.3+):

When you run the create script, it automatically:

Converts doc name to search query (e.g., m-dx2-better-errors → "better errors")
Runs both SimHash (instant) and Neural (better quality) searches
Shows results from both methods so you can compare
Merges unique results (neural preferred) for the template
Auto-populates the "Related Documents" section

$ .claude/skills/design-doc-creator/scripts/create_planned_doc.sh m-semantic-caching

🔍 Searching for related design docs...

Implemented docs matching "semantic caching":
  [SimHash - instant]
  1. design_docs/implemented/v0_4_0/monomorphization.md (1.00)
  2. design_docs/implemented/v0_3_18/M-DX4-SPRINT-PLAN.md (0.95)
  [Neural - semantic matching]
  1. design_docs/implemented/v0_6_0/m-doc-sem-lazy-embeddings.md (0.45)
  2. design_docs/implemented/v0_6_0/semantic-caching-complete.md (0.42)

Planned docs matching "semantic caching":
  [SimHash - instant]
  1. design_docs/planned/v0_7_0/M-REPL1_persistent_bindings.md (1.00)
  [Neural - semantic matching]
  1. design_docs/planned/v0_7_0/semantic-caching-future.md (0.50)

ℹ Related docs found above - review them after creation if needed.

Why dual search? SimHash is instant but keyword-dependent. Neural finds semantically related docs even when keywords don't match. You see both so you can judge which results are more relevant.

When to Use This Skill

Invoke this skill when:

User asks to "create a design doc" or "write a design doc"
User says "plan a feature" or "design a feature"
User mentions "document the design" or "create a spec"
Before starting implementation of a new feature
After completing a feature (to move to implemented/)

Coordinator Integration

When invoked by the AILANG Coordinator (detected by GitHub issue reference in the prompt), you MUST output this marker at the end of your response:

DESIGN_DOC_PATH: design_docs/planned/vX_Y/design-doc-name.md

Why? The coordinator uses this marker to:

Read the design doc content for GitHub comments
Track artifacts across pipeline stages
Provide visibility to humans reviewing the issue

Example completion:

## Design Document Created

I've created the design document...

**DESIGN_DOC_PATH**: `design_docs/planned/v0_6_3/m-feature-design.md`

Available Scripts

`scripts/create_planned_doc.sh <doc-name> [version]`

Create a new design document in design_docs/planned/.

Version Auto-Detection: The script automatically detects the current AILANG version from CHANGELOG.md and suggests the next version folder. This prevents accidentally placing docs in wrong version folders.

Usage:

# See current version and suggested target
.claude/skills/design-doc-creator/scripts/create_planned_doc.sh
# Output: Current AILANG version: v0.5.6
#         Suggested next version: v0_5_7

# Create doc in planned/ root (no version)
.claude/skills/design-doc-creator/scripts/create_planned_doc.sh m-dx2-better-errors

# Create doc in next version folder (recommended)
.claude/skills/design-doc-creator/scripts/create_planned_doc.sh reflection-system v0_5_7

What it does:

Searches for related docs using ailang docs search --neural (Ollama embeddings)
Shows top 3 matches from both implemented/ and planned/
Auto-populates "Related Documents" section with clickable links
Detects current version from CHANGELOG.md
Suggests next patch version for targeting
Creates design doc from template
Places in correct directory (planned/ or planned/VERSION/)
Fills in creation date
Shows version context in output

`scripts/move_to_implemented.sh <doc-name> <version>`

Move a design document from planned/ to implemented/ after completion.

Usage:

.claude/skills/design-doc-creator/scripts/move_to_implemented.sh m-dx1-developer-experience v0_3_10

What it does:

Finds doc in planned/
Copies to implemented/VERSION/
Updates status to "Implemented"
Updates last modified date
Provides template for implementation report
Keeps original until you verify and commit

Workflow

Creating a Planned Design Doc

1. Gather Requirements

Ask user:

What feature are you designing?
What version is this targeted for? (e.g., v0.4.0)
What priority? (P0/P1/P2)
Estimated effort? (e.g., 2 days, 1 week)
Any dependencies on other features?

⚠️ CRITICAL: Audit for Systemic Issues FIRST

Before writing a design doc for a bug fix, ALWAYS ask: "Is this part of a larger pattern?"

The Anti-Pattern (incremental special-casing):

v1: Add feature for case A
v2: Bug! Add special case for B
v3: Bug! Add special case for C
v4: Bug! Add special case for D
...forever patching

The Pattern to Follow (unified solutions):

v1: Bug report for case B
    BEFORE writing design doc:
    1. Search for similar code paths
    2. Check if A, C, D have same gap
    3. Design ONE fix covering ALL cases
v2: Unified fix - no future bugs in this area

Concrete Example (M-CODEGEN-UNIFIED-SLICE-CONVERTERS, Dec 2025):

Bug reported: [SolarPlanet] return type panics

❌ Quick fix design doc: Add ConvertToSolarPlanetSlice
   (Will need ConvertToAnotherRecordSlice later...)

✅ Systemic design doc: Audit ALL slice types
   Found: []float64 ALSO broken!
   Found: []*ADTType partially broken!
   One unified fix covers all 3 gaps.

Analysis Checklist (do BEFORE writing design doc):

[ ] VERIFIED every "AILANG does/doesn't support X" claim with ailang check (HARD GATE — see below)
[ ] Is this a one-off or part of a pattern?
[ ] Search codebase for similar code paths
[ ] Check if other types/cases have the same gap
[ ] Look at git history - has this area been patched repeatedly?
[ ] If citing eval data: segmented by recent date + confirmed the cited construct is the ACTUAL cause
[ ] Design fix to cover ALL cases, not just the reported one

⚠️ CRITICAL: Verify Every Language Claim Before Asserting It (HARD GATE)

Any statement of the form "AILANG does / does not support X" MUST be verified with a live ailang check before it goes in the doc. This is not optional. A design doc that mischaracterizes the language sends the implementer to build something that already exists, reject something that's fine, or fix a non-problem.

The check takes 10 seconds:

tmp=$(mktemp -d)
cat > $tmp/claim.ail <<'EOF'
module test/claim
-- the exact construct you're claiming is (un)supported
import std/list as L
export func main() -> () ! {} = ()
EOF
ailang check $tmp/claim.ail   # COMPILES = supported; PAR_/type error = not
rm -rf $tmp

Required in the doc: every "supported"/"unsupported" claim must carry either (a) an ailang check result/transcript, or (b) a citation to a reference page or implemented design doc. Unverified assertions are treated as unproven and the doc cannot proceed to sprint.

Also verify the FREQUENCY claim if the doc cites eval data ("fails in N models", "X% of failures"): segment by recent date (not all-time aggregate — old baselines mix in pre-fix runs), and confirm the cited construct is the ACTUAL cause of the failure, not a co-occurring line. A regex that flags as <word> in an import is NOT proof the import is the cause.

Case study (2026-06-03, this exact failure): Two hand-written eval-gap docs asserted language limitations without running ailang check:

m-type-constraints claimed "AILANG has no typeclasses, use explicit comparators." FALSE — AILANG infers Ord/Num via dictionary passing; mymax[a](x:a,y:a)=if x>y then x else y runs on int AND string. The doc would have built an unnecessary feature.
m-import-alias claimed "AILANG has no import aliases, implement them." FALSE — import M as L and import M (f as g) both compile. The cited "6% of failures" was a detection-heuristic false positive: 0 of 16 flagged failures actually failed on an alias. Doc was REJECTED.

Both errors were a 10-second ailang check away from being caught. The neural related-doc search (below) is passive context; THIS gate is the active check. Do not skip it by hand-writing content over the scaffold — fill the scaffold, and verify as you fill.

The same "verify against the code, not your assumption" rule covers four more claim classes the Verification Log routinely marks "Confirmed" without actually checking:

A newly-proposed diagnostic/error code MUST be verified unallocated. Error codes (MODxxx, PARxxx, TCxxx, EFF_*, Exxx) are a shared namespace. Before writing Success: exit ≠ 0 + MODnnn: … into a doc, grep the codebase:
```
grep -rn "MOD011" internal/ cmd/   # empty = free; any hit = TAKEN, pick the next free code
```
Put the grep result in the Verification Log. The intended code is a claim like any other.
A Conflict-Surface routing/mechanism claim MUST be verified by reading the code path, not inferred from observable output. "Construct X routes to runSingle, never reaches validateModulePath" is a claim about control flow — confirm it by reading the dispatch, not by observing that X produces the right answer. A correct output can hide a wrong mechanism, and the wrong mechanism yields a wrong guard/fix that the implementer must then override mid-sprint.
Every cited regression fixture MUST exist, and every claim about an existing test's BEHAVIOR must come from reading the test body. The Conflict Surface's "Programs that MUST still work" list and any "test X asserts Y" statement are claims like the rest:
```
ls examples/lambdas_higher_order.ail        # cite only files that exist
sed -n '/func TestName/,/^}/p' path_test.go # read what it actually asserts
```
Case study (2026-07-13, m-arity-style-diagnostic): the doc cited examples/lambdas_higher_order.ail + examples/no_loops_fold.ail as regression fixtures — neither exists — and claimed TestCurriedMismatchStillFails "asserts on the old bare string" requiring "the ONE intentional test-text change"; the test asserts only err != nil, so the planned edit was fictional. Both were one ls/one read away at doc time; the sprint-planner had to correct the premises mid-loop.
Every NEGATIVE-EXISTENCE claim about an internal structure the design RELIES ON MUST carry its own Verification Log row. A design that says "the AST has no ParenExpr node," "no shared safe-write helper exists," "there is no fmt subcommand," or "field X is not retained" is making a claim about the absence of something — and a reject-by-default reviewer will (correctly) block on any such claim that lacks a grep/read proving it, because if the thing DOES exist the exhaustive-visitor / reuse / mechanism premise silently breaks. These are easy to leave implicit because "it's not there" feels like a non-claim — but it is exactly as load-bearing as a positive claim. Grep/read and record each one:
```
grep -rin "paren" internal/ast/*.go        # prove "no ParenExpr node" (empty/only-comments = confirmed)
grep -rn "os.Rename\|SafeWrite\|Atomic" internal/ cmd/   # prove "no shared write helper" (or name the one to reuse)
```
Rule of thumb: if a design sentence contains "no", "none", "doesn't exist", "not retained", "unallocated", or "must reconstruct because … isn't stored", it needs a log row. Sweep the doc for these BEFORE submitting to quorum — one row each is cheap; discovering them one-per-round via reviewer rejects ping-pongs a sound design across the bounded revision limit and parks it. Case study (2026-07-18, m-ailang-fmt, iter 49): the formatter design was architecturally sound and the controller passed it, but gemini-3-1-pro (reject-by-default) blocked it TWICE — round 1 on "atomic-write uses the repo's helper if one exists" (no helper exists; the temp+os.Rename pattern is ad-hoc at 4 sites → V19), round 2 on "the AST has no parenthesis node" (true — only // move to LPAREN comments → V20). Both were TRUE negatives missing a log row, surfaced one-per-round; the re-quorum-ONCE guardrail was then exhausted and a complete design parked for a human ratification it did not need. Both rows were a 10-second grep away at authoring time.

Case study (2026-07-12, m-module-less-run-fail-loud): the doc's Verification Log marked two things "Confirmed" that a code-check refuted. (a) It proposed error code MOD011 — already the live module-path-collision diagnostic since v0.10.9; the reality-check reassigned it to MOD014 (one grep would have caught it). (b) It marked "bare-expr eval unaffected — routes to runSingle, never reaches validateModulePath" Confirmed, on the strength of 1+1 → 2 still working; the mechanism was false (a bare-expr file does reach validateModulePath, and the parser mirrors the expression into Statements+Decls), so the doc's proposed 3-way Funcs||Statements||Decls guard would have broken ailang run 1+1. The executor had to override it to Funcs-only mid-sprint. Both were a one-command code-check away at doc time.

⚠️ CRITICAL: Conflict Surface Analysis (REQUIRED for parser/typechecker/codegen changes)

If the design touches internal/parser/, internal/lexer/, internal/ast/, internal/types/, internal/elaborate/, internal/iface/, internal/codegen/, internal/eval/, internal/vm/, internal/effects/, or cmd/ailang/exec.go:

The design doc MUST include a Conflict Surface section (see resources/design_doc_structure.md) enumerating:

What syntactic/semantic positions does this change extend?
What OTHER valid constructs already live in those positions?
How does the parser/typechecker disambiguate?
Which existing programs MUST still work post-change? (3-5 fixtures)
What deliberately changes (intentional incompatibilities)?

Why this is required: most language regressions come from "I didn't realize X also uses this position." The author is the only one who can credibly enumerate the conflict surface. Reviewers can sanity-check but can't generate the list.

Concrete case study: M-TAINT-TYPES (v0.14.3) added T{not LABEL} refinement syntax without enumerating that func ... -> bool { not f(x) } already used the same { not <ident> prefix in function bodies. The 2-token-lookahead disambiguation was insufficient. The motoko_agent fork (still on v0.13.0) hit ~14 mis-parses when migrating to v0.15.x. Caught months after release. See M-PARSER-REFINEMENT-LOOKAHEAD changelog entry for the fix.

The honest answer is almost never "no conflicts": if the section says that for a parser/typechecker change, the author hasn't looked hard enough.

Check existing work: The create script auto-searches for related docs using both SimHash and neural embeddings. Review the results before filling in the template.

Warning signs of fragmented design (expand scope if you see these):

Multiple maps tracking similar things
Switch statements with growing case lists
Bug fixes that add || specialCase conditions

Axiom Compliance: Every feature must score against all 12 axioms. Hard violations on A1/A3/A4/A7 = automatic rejection, net score must be ≥ +2. See resources/design_doc_structure.md for full scoring matrix and examples.

Quorum review (OPTIONAL, off-Anthropic — M-MISSION-FLEET-AB Phase B): Before a design doc proceeds to planning, you MAY run an independent N-reviewer quorum that scores it against these same hard gates (premise verification, Conflict Surface, axiom compliance) using non-Anthropic frontier models — a cheap (cents/doc), off-quota second opinion that catches bad premises before an Opus sprint is spent on them. This is an optional documented step, not a contract change: the design-doc-creator flow is unchanged if you skip it.

# One reviewer (reject-by-default; exits non-zero if it can't produce a verdict):
ailang design-review design_docs/planned/vX_Y/my-doc.md --reviewer gpt5-6-sol --json

# Full quorum (parallel reviewers + your own IN-SESSION verdict as the controller):
ailang design-quorum design_docs/planned/vX_Y/my-doc.md \
  --reviewers gpt5-6-sol,gemini-3-1-pro \
  --controller-verdict pass --controller-note "<your in-session judgement>" \
  --mission-log design_docs/v1-mission-log.md

Reviewers are reject-by-default (each must state a strongest_objection). Synthesis: any present reviewer or the controller rejects → blocked (exit 3); the objection goes back to you, the author, before planning. A reviewer that is unreachable / over-budget / mis-authed is recorded by NAME with its reason and the quorum degrades to N−1 — never a silent pass. Each run writes a machine artifact under .ailang/state/mission-quorum/ (seeds the Phase E assignment table) and, with --mission-log, a human markdown block. Gemini reviewers use Vertex ADC (not GEMINI_API_KEY). See ailang design-quorum --help.

2. Choose Document Name

Naming conventions:

Use lowercase with hyphens: feature-name.md
For milestone features: m-XXX-feature-name.md (e.g., m-dx2-better-errors.md)
Be specific and descriptive
Avoid generic names like improvements.md

3. Run Create Script

# If version is known (most cases)
.claude/skills/design-doc-creator/scripts/create_planned_doc.sh feature-name v0_4_0

# If version not decided yet
.claude/skills/design-doc-creator/scripts/create_planned_doc.sh feature-name

4. Duplicate / Coverage Gate (MANDATORY — do this before creating the file)

Before creating any doc, read the top matches from the search and apply this gate:

| Similarity score | Action | |---|---| | ≥ 0.75 (neural) on a planned doc | REJECT — the topic is already queued. Reply with the path and explain what's already covered. Do NOT create a new doc. | | ≥ 0.65 (neural) on an implemented doc | REJECT — already shipped. Reply with the path + version it shipped in. Do NOT create a new doc. | | 0.45–0.65 on any doc | Warn — read the doc, confirm your topic is genuinely distinct before proceeding. Note the distinction explicitly in the "Related Documents" section. | | < 0.45 | Proceed normally. |

Rejection reply format (use this when rejecting):

⛔ Duplicate / Already Covered

This topic is already addressed in:
  [doc title](path/to/doc.md) — [planned v0.X.Y / implemented in vX.Y.Z]

Key overlap: <one sentence on what the existing doc covers that would duplicate this request>

If your request is genuinely distinct, please clarify how it differs from the above.

Read related docs found by search before proceeding (non-rejected cases):

Look at their structure and patterns
Note any design decisions that apply
Check for overlap with your feature
Reference them in your "Related Documents" section

What to look for in related docs:

Architecture patterns used
Testing strategies employed
Edge cases already handled
Implementation trade-offs documented
Success/failure metrics to compare against

5. Customize the Template

The script creates a comprehensive template. Fill in:

Header section:

Feature name (replace [Feature Name])
Status: Leave as "Planned"
Target: Version number (e.g., v0.4.0)
Priority: P0 (High), P1 (Medium), or P2 (Low)
Estimated: Time estimate (e.g., "3 days", "1 week")
Dependencies: List prerequisite features or "None"

Problem Statement:

Describe current pain points
Include metrics if available (e.g., "takes 7.5 hours")
Explain who is affected and how

Goals:

Primary goal: One-sentence main objective
Success metrics: 3-5 measurable outcomes

Solution Design:

Overview: High-level approach
Architecture: Technical design
Implementation plan: Break into phases with tasks
Files to modify: List new/changed files with LOC estimates. Use one repo-relative path per top-level bullet and make it the first backticked token; continuation lines and later backticks are prose.

Examples:

Show before/after code or workflows
Make examples concrete and runnable

Success Criteria:

Checkboxes for acceptance tests
Include "All tests passing" and "Documentation updated"

Timeline:

Week-by-week breakdown
Realistic estimates (2x your initial guess!)

6. Review and Commit

git add design_docs/planned/feature-name.md
git commit -m "Add design doc for feature-name"

Moving to Implemented

When to move:

Feature is complete and shipped
Tests are passing
Documentation is updated
Version is tagged/released

1. Run Move Script

.claude/skills/design-doc-creator/scripts/move_to_implemented.sh feature-name v0_3_14

2. Add Implementation Report

The script provides a template. Add:

What Was Built:

Summary of actual implementation
Any deviations from plan

Code Locations:

New files created (with LOC)
Modified files (with +/- LOC)

Test Coverage:

Number of tests
Coverage percentage
Test file locations

Metrics:

Before/after comparison table
Show improvements achieved

Known Limitations:

What's not yet implemented
Edge cases not handled
Performance limitations

3. Update design_docs/README.md

Add entry under appropriate version:

### v0.3.14 - Feature Name (October 2024)
- Brief description of what shipped
- Key improvements
- [CHANGELOG](../CHANGELOG.md#v0314)

4. Commit Changes

git add design_docs/implemented/v0_3_14/feature-name.md design_docs/README.md
git commit -m "Move feature-name design doc to implemented (v0.3.14)"
git rm design_docs/planned/feature-name.md
git commit -m "Remove feature-name from planned (moved to implemented)"

Design Doc Structure

See resources/design_doc_structure.md for:

Complete template breakdown
Section-by-section guide
Best practices for each section
Common mistakes to avoid

Notes

All design docs should follow the template structure
Update CHANGELOG.md when features ship (separate from design doc)
Link design docs from README.md under version history
Keep design docs focused - split large features into multiple docs
Use M-XXX naming for milestone/major features

Agent Skills: Design Doc Creator

Install this agent skill to your local

Skill Files