Menü
Use on an existing project to scan code once and produce both architecture docs and a MemPalace seed pass.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Basierend auf der SOC-Berufsklassifikation
Use when adding a feature to a production project. Full cycle: brainstorm, plan, implement, docs.
Use when fixing a bug in a production project. Full cycle: diagnose, fix, regression test, review, docs.
Use to batch-validate all skills in this plugin. Runs deterministic repo-wide static checks (frontmatter, command delegation, cross-references, Architect model= rule, orphan references, README<->MemPalace sync) and reports pass/fail per skill.
Use when creating a new project. Bare AI shell by default; interactive menu adds docs / backend-infra / agents on demand.
Use after init-project to fill docs/product/start-project.md with AI research on competitors, monetization, valuation, and marketing.
Use when unsure which skill to use. Lists all vladyslav skills, workflow, and integration.
| name | ingest |
| description | Use on an existing project to scan code once and produce both architecture docs and a MemPalace seed pass. |
Type: Architect
Single-pass project intake. Two bash scripts produce the discovery JSON; Opus main does the narrative synthesis (architecture docs) and the decision extraction (MemPalace records). Both outputs derive from the same source-of-truth scan, so they cannot disagree.
init-project — that already scaffolds docs/architecture/ and there is no git history to seed from.write-project-docs instead.Apply the verify-working-directory contract from <plugin>/skills/_shared/references/verify-pwd.md: confirms CLAUDE.md exists, derives the canonical MemPalace wing name, warns on stale-wing duplicates, and establishes the mandatory path-validation rule for the rest of this skill's MemPalace reads.
If CLAUDE.md is missing → STOP and suggest /vladyslav:attach-project first to bootstrap the AI workflow shell.
Glob ~/.claude/plugins/cache/vladyslav-marketplace/vladyslav/*/scripts/scan-architecture.sh and take the directory two levels up. Fall back to /Volumes/DevSSD/Development/vladyslav-skills (development clone).
Run the two scripts in parallel — they are independent and write to separate JSON outputs (one shell invocation, backgrounded):
<plugin-root>/scripts/scan-architecture.sh --pwd . > /tmp/ingest-arch.json &
<plugin-root>/scripts/gather-seed-signals.sh --pwd . > /tmp/ingest-signals.json &
wait
ARCH=$(cat /tmp/ingest-arch.json); SIGNALS=$(cat /tmp/ingest-signals.json)
ARCH schema:
{
"stacks": {<from detect-stack.sh: ios/python/go/...>},
"entry_points": [<paths>],
"routes": {"framework": "fastapi|flask|express|go-stdlib|none", "handlers": [{method, path, file}, ...]},
"schema_files": [<SQL migrations, Prisma, Drizzle, Alembic>],
"deps": {<manifest path>: <summary string>},
"doc_files": [<paths under docs/>]
}
SIGNALS schema:
{
"git": {
"available": true|false,
"head_commit": "<short sha>",
"branch": "<current branch>",
"first_commit_date": "<YYYY-MM-DD>",
"recent_themes": [<last 30 non-merge commit subjects>],
"decision_commits": [<feat/refactor/fix/decision-prefixed subjects, up to 40>],
"most_edited": [{"path": "...", "edits": N}, ...]
},
"manifests": {<path>: <summary string>},
"existing_docs": [<paths under docs/>],
"adr_files": [<paths under docs/architecture/adr/ or docs/decisions/>],
"claude_md": {"exists": true|false, "size": <bytes>}
}
Together, ARCH answers "what does the code look like now", SIGNALS answers "what changed historically and what decisions exist already". The LLM combines them.
Before writing seed records, search the current wing to avoid duplicates:
mempalace_search wing=<wing> "architecture"
mempalace_search wing=<wing> "decision"
mempalace_search wing=<wing> "stack"
Categorise existing records:
Path-validation rule applies to every search result — drawers referencing non-existent paths are marked [STALE] and excluded.
Orchestration (Steps 4 + 5): once the Step 3 user decision is settled (a serial gate), Steps 4 and 5 are independent — Step 4 writes only
docs/, Step 5 writes only MemPalace. Dispatch them concurrently: Step 4 as asonnetsubagent (narrative generation from JSON), Step 5 onopus(decision extraction is judgment — keep it on the strongest model). The Opus main session merges results and runs Step 6. See_shared/references/orchestration-conventions.md. (MemPalace writes inside Step 5 still runcheck_duplicatesequentially — never parallelize writes.)
sonnet subagent)Using ARCH, write or update:
docs/architecture/system.md — high-level narrative derived from stacks, entry_points, and the codebase layout. 2-4 paragraphs. Preserve any user-edited sections if the file already exists; merge rather than overwrite.docs/architecture/api.md — only if routes.framework is not none. Table of method | path | handler-file from routes.handlers. Group by feature area when possible.docs/architecture/db-schema.sql — only if schema_files is non-empty. Stitch the relevant SQL migrations / Prisma / Drizzle schemas into one consolidated reference. Preserve user-edited sections.Do NOT touch:
docs/product/* (that's discover / init-project's job)docs/plans/*, docs/testing/*, docs/release/*, docs/operations/*, docs/marketing/* (these are not architecture)If claude_md.exists is false → also write a minimal CLAUDE.md with the Source-of-Truth table pointing at the docs you just wrote. If it already exists, leave it alone (it's a high-touch user document — don't blast it).
Using ARCH + SIGNALS together, identify 10–20 records worth seeding. Apply <plugin>/skills/_shared/references/mempalace-record.md for shape and room-type rules. Quality over quantity — re-seedability matters.
What to capture:
decision) — derived from manifests, framework choice (routes.framework), schema decisions (schema_files), notable patterns from most_edited files.decision) — when manifests reveal a deliberate choice (FastAPI over Flask, Drizzle over Prisma, etc.). If the rationale isn't documented, write the choice as [WHAT] and leave [WHY] blank with a <unknown — confirm with team> placeholder.problem) — signals.decision_commits filtered to fix: prefixes. One drawer per recurring class, not per individual fix.milestone) — release-like commit subjects (e.g. release: v...) and the first-commit date.decision) — for each path in signals.adr_files, write a short pointer drawer so future searches surface them.What to skip:
first_commit_date makes no sense to writeBefore each mempalace_add_drawer, run mempalace_check_duplicate to avoid pollution. For relationship facts (module X depends on Y, decision D supersedes E), also call mempalace_kg_add with subject/predicate/object.
Run 3–5 mempalace_search queries within the wing to confirm the new records surface for likely future queries. Example:
If a record does not surface for an obvious query, rewrite its content with better keywords (the [WHAT] line is what the search indexes most heavily) and re-add.
If a ## Memory section is absent, append:
## Memory
This project has been ingested into MemPalace. Wing: `<wing-name>`.
Before scanning the codebase, search the wing: `mempalace_search wing=<wing-name>`.
Last ingested: <YYYY-MM-DD>.
If a ## Memory section already exists, update the Last ingested: date in place.
✓ Architect report — Ingest
- Project: <name>
- Wing: <wing-name>
- Detected stacks: <list from ARCH.stacks>
- Routes detected: <count> across <framework>
- Schema files: <count>
Architecture docs:
- docs/architecture/system.md <created | merged | unchanged>
- docs/architecture/api.md <if backend present>
- docs/architecture/db-schema.sql <if schemas present>
MemPalace records:
- Decisions added: <count>
- Problems added: <count>
- Milestones added: <count>
- Duplicates skipped: <count>
- Verification searches passed: <count>/<count>
CLAUDE.md memory pointer: <added | updated | unchanged>
Next steps:
- /vladyslav:add-feature — build new features with both architecture docs and MemPalace context now ready
- /vladyslav:design-sync — if UI work is upcoming
- /vladyslav:discover — only if docs/product/start-project.md is incomplete (rare on an existing project)
ARCH and SIGNALS into a narrative docs/architecture/system.md requires reading the structured data, identifying the load-bearing patterns, and writing in human prose. No bash script can substitute.signals.decision_commits); the LLM curates.scan-architecture.sh and gather-seed-signals.sh) do all the I/O. The LLM doesn't read files directly — it reads the JSON. That's the leverage. The scripts run in parallel (Step 2); doc synthesis (sonnet) and decision extraction (opus) run as a concurrent fan-out (Steps 4+5).docs/architecture/system.mddocs/architecture/api.md (backend only)docs/architecture/db-schema.sql (schemas only)CLAUDE.md — only if missing; never overwrittenswift-calories wing — only if a new Apple-pattern decision is discovered (rare; mostly discover-apple-check's job)If the project already has architecture docs (e.g. from a previous ingest run), Step 4 merges rather than overwrites — user-edited sections are preserved and only new findings are added.
If the MemPalace wing is already populated, Step 3 detects this and asks whether to add only newly-discovered records or re-seed from scratch. Re-seeding is safe: mempalace_check_duplicate prevents exact duplicates, and Step 6 verifies searchability after the write.