anatomy
An agent skill for creating detailed, interactive isometric illustrations of real-world machines and complex systems. Generates polished SVGs with optional WebGL shaders and 3D effects.
- 29k
- JavaScript
- MIT
- 5 wk ago
Turn a product idea into a grounded, buildable SRD suite — a skills.sh agent skill
Repo maxgfr/construct, read from GitHub 12 h ago. maxgfr has not joined Take My Repo.
maxgfr/construct is a public repo on GitHub by maxgfr. Turn a product idea into a grounded, buildable SRD suite — a skills.sh agent skill. It declares the MIT license on GitHub. It is written mainly in JavaScript, has 1 stars and last changed on 2026-11-20. Nobody has reviewed it on Take My Repo yet, which says nothing about its quality. Read from GitHub on 2026-11-22.
In the maintainer's words. Shown from GitHub; Take My Repo does not keep a copy.
On request by default. construct runs when you ask for it: the agent may call
it when you request a product SRD, and you can invoke it directly with $construct
in Codex, /construct in Claude Code or OpenCode. The agent does not pick it up on
its own, and CLI commands are unchanged. One setting makes it
explicit-only — see Manual or automatic.
Turn a product idea into a grounded, buildable SRD suite — a Software Requirements Document whose requirements and decisions rest on real research (competitors, open-source prior art, technology docs, known pitfalls), not the model's memory. A skills.sh agent skill.
construct is the companion to reconstruct
(rebuild a repo into PRDs) and ultradoc
(answer questions grounded in a repo). Same engineering: a single committed,
zero-dependency Node bundle + a thick agent playbook, fully tested, released
by Conventional Commits.
brief.json.[E#] ids:SRD.json + a machine-readable BUILD-PLAN.json task DAG. At
--level complex it also renders a design system (design/: principles,
design tokens, components, screens/flows, an accessibility contract); pass
--no-design to skip it.--min-grounding threshold), and — via
review + check --semantic — an opt-in claim-support gate that fails
on any cited evidence that doesn't actually back its claim.BUILD-PLAN.json (status --json lists the buildable task frontier so
independent same-milestone tasks can be built in parallel); construct verify
referees it against the SRD — declared files exist, every requirement is
referenced by a test, and (with --run-tests) the declared test commands pass.No API keys. No npm install at skill-use time.
npx skills add maxgfr/construct
skills add detects the agents you use and installs for them; pass
-a <agent> (e.g. -a claude-code, -a codex, -a opencode) to pick one
explicitly. The skill is agent-neutral: any host that loads Agent Skills can run it.
node scripts/construct.mjs init --idea "a self-hosted read-it-later app" --out ./readpile
# optional: diverge first, then fold kept ideas into the brief
node scripts/construct.mjs brainstorm --out ./readpile # scaffold BRAINSTORM.md
node scripts/construct.mjs brainstorm --out ./readpile --merge # kept → brief.json
# …fill ./readpile/brief.json via the interview…
node scripts/construct.mjs research --out ./readpile --angles market,oss,tech
node scripts/construct.mjs analyze --out ./readpile # what's thin? drill it
node scripts/construct.mjs render --out ./readpile --level complex
node scripts/construct.mjs check --out ./readpile # add --min-grounding 70 to enforce
node scripts/construct.mjs review --out ./readpile # adjudicate each cited [E#] → verdicts.json
node scripts/construct.mjs check --out ./readpile --semantic # gate refuted/unsupported claims
node scripts/construct.mjs verify --out ./readpile --app ./readpile-app --run-tests --strict
Add --merge to also emit a single-file SRD.md. Run --help for the full
surface, or see SKILL.md for the agent playbook and
DOCUMENTATION.md for internals.
00-overview/ VISION.md · SCOPE.md
requirements/ FUNCTIONAL.md · NON-FUNCTIONAL.md
architecture/ SYSTEM-CONTEXT.md · DATA-MODEL.md · INTERFACES.md · decisions/NNNN-*.md
design/ PRINCIPLES.md · DESIGN-TOKENS.md (+ design-tokens.json) · COMPONENTS.md · SCREENS.md · ACCESSIBILITY.md (complex; --no-design to skip)
competitive/ LANDSCAPE.md
BUILD-PLAN.md · BUILD-PLAN.json (task DAG for the build phase) · TRACEABILITY.md
evidence/ EVIDENCE.md · evidence.json · meta.json · brief.json · SRD.json
VERIFY.md · VERIFY.todo.json · VERIFY.json (claim-support review, from `review`)
construct check separates the two axes. The structural gate fails the build
on an incomplete SRD (unresolved 🧠 decisions, an FR with no acceptance
criteria, a dangling reference, a missing required NFR category, a malformed
ADR). The grounding coverage is a report — it tells you how well-cited the
SRD is so you can invest research where it matters, but it never fails the build
by default. When you do want it enforced, --min-grounding <0-100> opts into
a second gate that fails below the threshold.
Coverage counts citations; it does not check they hold. construct review
builds a claim↔evidence worklist (one pair per cited [E#]); an agent (or a
fan-out of skeptic subagents — see references/orchestration.md) adjudicates
each as supported | partial | refuted | unsupported, and check --semantic
turns that into a third opt-in gate that fails on a refuted or unsupported
claim.
The skill shells out to the CLI and parses its output. An MCP server skips both: your agent calls construct as typed tools, with JSON schemas in and structured results out. Same engine, same page cache, no wrapper.
Any MCP client works; register it with whichever host you use.
# stdio — the default; the server is just this command:
# node /abs/path/to/scripts/construct.mjs mcp
claude mcp add construct -- node /abs/path/to/scripts/construct.mjs mcp # Claude Code
codex mcp add construct -- node /abs/path/to/scripts/construct.mjs mcp # Codex
opencode mcp add construct -- node /abs/path/to/scripts/construct.mjs mcp # OpenCode
# or over HTTP, on loopback — any client takes the URL
node scripts/construct.mjs mcp --transport http --port 7342
claude mcp add --transport http construct http://127.0.0.1:7342/mcp # Claude Code
codex mcp add construct --url http://127.0.0.1:7342/mcp # Codex
opencode mcp add construct --url http://127.0.0.1:7342/mcp # OpenCode
// Claude Desktop takes stdio servers only — a remote URL here will not work.
{ "mcpServers": { "construct": { "command": "node", "args": ["/abs/path/to/scripts/construct.mjs", "mcp"] } } }
// Cursor, HTTP:
{ "mcpServers": { "construct": { "url": "http://127.0.0.1:7342/mcp" } } }
It serves all three MCP primitives, because a skill is three things: the engine
(tools), the method (prompts), and the documentation the method refers
to (resources). Here that matters more than usual: check gates structure
and grounding is advisory, so a client given only the tools produces a
well-shaped SRD full of decisions nobody researched — and every gate goes green.
| Tool | What it does |
|---|---|
construct_status |
What exists in the run, and the exact next command |
construct_research |
The only command that grounds anything — market, OSS, tech → a dossier |
construct_research_angle |
Probe ONE angle (web/oss/tech/so), persist nothing |
construct_analyze |
What is thin, and the command that fills each gap |
construct_render |
The SRD suite: requirements, ACs, NFRs, ADRs, build plan, traceability |
construct_check |
The structural gate (grounding coverage is advisory) |
construct_review |
Claim↔evidence worklist — where advisory grounding becomes real |
construct_verify |
Referee a built app against its SRD |
construct_cache |
What the page cache holds |
construct_read |
A file, or a line range, from the run |
--allow-write additionally exposes construct_init, which scaffolds a run
folder on disk. Pass --out <run> at startup to dedicate the server to one SRD
— run then becomes optional on every tool.
| Prompt | Arguments | What it drives |
|---|---|---|
interview_idea |
idea |
The questions whose answers change what gets built — one at a time, following the surprising answer |
enrich_srd |
run |
evidence → testable requirements → check → review |
judge_adr |
run, decision? |
Ground a technology choice in its docs, its open issues, and what people hit in production |
Each states the thing the gates cannot: the rigor is yours. A green
construct_check means the SRD is well-formed, not well-researched.
SKILL.md and all 19 references/*.md are served under skill://, read off
disk at request time — so a documentation fix reaches every client without a
rebuild.
Two things worth knowing:
construct_research is the slow one — a network fan-out across angles,
minutes on a first run. Re-runs are nearly free thanks to the page cache.127.0.0.1 and refuses anything else unless you
pass --allow-remote. This server fetches arbitrary URLs and reads local
files; an exposed port is a fetch-anything primitive for whoever finds it.node scripts/construct.mjs semantic up # Qdrant + Ollama + SearXNG, fully local, no key
node scripts/construct.mjs firecrawl up # Firecrawl, fully local, no key (~3 GB, own profile)
The first adds embedding-based re-ranking (--semantic) and keyless web
discovery. The second swaps the built-in regex HTML stripper for browser-based
main-content extraction on every page fetch — the only way a JS-rendered page
yields evidence at all. Both are optional and degrade quietly: when a stack is
down, research runs exactly as it did before and says so in the dossier notes.
See references/semantic-setup.md.
verify without --acceptance checks static consistency; FR tags are not proof
that a criterion ran. For the final build gate, inspect
verify --out <run> --acceptance --json (nonzero until executed), bind each
current {frId,index,fingerprint} to a dedicated test command under its done
task's verify.criteria, then run:
node scripts/construct.mjs verify --out <run> --strict --acceptance --run-tests --json
Every criterion reports passed, failed or not-tested with its full-FR
fingerprint and bounded command output. Missing, stale, duplicate or unexecuted
criteria fail. The command always executes afresh; it does not import old green
reports. Review the actual assertions before claiming implementation: exit 0
does not prove a test exercises its mapped requirement. Commands run locally
with your privileges, through the existing shell runner, not a sandbox. Schema,
limits and rebinding protocol: verify reference.
MIT © maxgfr
A .pdf URL or an application/pdf response goes through an extractor
ladder (src/research/pdf/): npx @firecrawl/pdf-inspector@1 → npx @firecrawl/anydoc@0.1 (the PDF on
stdin, in a child process) → the self-hosted Firecrawl → pdftotext → a
built-in dependency-free reader — stopping at the first rung whose output passes
a quality gate, and REFUSING rather than quoting a PDF none of them could read.
Office documents — .docx/.doc/.odt/.rtf, .pptx/.ppt/.odp,
.xlsx/.xls/.ods, .epub, .csv — go through their own two-rung ladder
(src/research/doc/): npx @firecrawl/anydoc@0.1 (the bytes on stdin, converted to
GitHub-Flavored Markdown) → the self-hosted Firecrawl. Same gate, same refusal.
The refusal is the point: these are ZIP and OLE containers, so the fall-through
this replaced did not degrade the evidence, it fabricated it — a .docx was
quoted into requirements as if it were prose, as kilobytes of replacement characters, silently. anydoc needs
Node 20+, so an unavailable converter is a normal outcome rather than a
misconfiguration; CONSTRUCT_DOC_ENGINE=none disables the ladder.
Scanned PDFs — no text layer at all, so every rung above fails — are
rescued by a final OCR rung: copyable-pdf
tesseract, when both are installed. It is last because it is the only
expensive one (~2.7s per page at 300 DPI) and is budgeted per process
(CONSTRUCT_OCR_MAX, default 3, 0 disables). Both binaries are checked before
the tool is spawned: asked for a missing tesseract, copyable-pdf offers to run
brew install / sudo apt-get install -y and waits on stdin, and a research
run must never install a system package as a side effect.Without it a PDF body was returned verbatim: its bytes decoded as UTF-8, cached, and quoted into requirements as if it were prose. The gate rejects text laced with C0/C1 control bytes or U+FFFD at ANY length — the built-in reader can emit 16 MB of image-stream garbage for a 12 MB paper, which every length-limited check waves through.
CONSTRUCT_NO_NPX=1 drops the npx rung; CONSTRUCT_PDF_ENGINE=<rung> pins one.
The stack is shared with the sibling skills (ultrasearch, construct, ultradoc): one compose project, one set of containers, one set of volumes. They used to define three separate projects on the same host ports, so only one could be up at a time — starting a second failed on the port after leaving its sidecars running. Bringing it up from any of them now targets the same containers, so the second is a no-op and the RAM is paid once.
Compose files and bind-mounted settings use the shared directory
~/.cache/skills/compose. Set ULTRA_STACK_CACHE_DIR to the same directory
for all three tools to override its parent. Per-tool HTTP and clone cache
overrides still apply only to those caches. The first startup after upgrading
from per-tool Compose directories may recreate SearXNG once to move its mount;
subsequent startups from another tool reuse it.
Upgrading from a version with per-skill container names? Remove the old ones once — this file can no longer stop them, and they still hold the ports:
docker rm -f $(docker ps -aq --filter name='^(ultrasearch|construct|ultradoc)-')
See shared engine maintenance for pins, source adoption checks and the daily repin workflow.
construct ships model-invocable, on request, and skills add installs it
that way: the agent may call it through its skill tool, but the description
restricts it to explicit requests, so it runs when you ask for it, not on its
own. You can still invoke it directly with $construct in Codex, /construct in
Claude Code or OpenCode.
Making it explicit-only again is one setting, read by Claude Code and OpenCode, applied to the installed copy of the skill:
| Host | Shipped, on request | Explicit-only |
|---|---|---|
| Claude Code, OpenCode | no disable-model-invocation in SKILL.md |
add disable-model-invocation: true |
Claude Code can do it without touching the file:
"skillOverrides": { "construct": "user-invocable-only" } in settings.json
leaves /construct working while hiding the skill from the model. Updating or
reinstalling the skill restores the shipped default, so reapply the change afterwards.
maxgfr/construct is a public repo on GitHub by maxgfr. Turn a product idea into a grounded, buildable SRD suite — a skills.sh agent skill.
It declares the MIT license on GitHub.
Its maker has not tagged a system, so it is unknown here. Its README says what it needs.
Open "Take it" on its page: Take My Repo writes an install plan for your machine and hands it to your AI agent (Claude Code, Codex or Cursor), which asks you before each step. Take My Repo has not tested it on any machine.
Its last change on GitHub was on 2026-11-20. Take My Repo does not judge from that whether it is maintained.
Nobody has reviewed it on Take My Repo yet, which says nothing about its quality.

An agent skill for creating detailed, interactive isometric illustrations of real-world machines and complex systems. Generates polished SVGs with optional WebGL shaders and 3D effects.

Reverse engineer anything with agents, from app behavior down to native binaries.
A privacy-first app that strips AI watermarks from content you own.

A self-hosted, browser-based AI CSV analyzer
My GitHub stars, sorted by AI and searchable, rebuilt daily by starsmith

Self-hosted fundamental stock screener — 150k+ equities, zero API keys forever. Transparent scores: Piotroski, Altman, Ohlson, Zmijewski, Beneish, Montier, Greenblatt magic formula, Graham — full provenance, DuckDB-fast (<1s full-universe).
Client-side database schema visualizer with AI-powered analysis. Upload SQL, visualize, share, export. No backend needed.
Github action which lets you to change a value from a json file (e.g. package.json)
This page is also available as structured data, with the same sources and dates.
/api/v1/resources/gh/maxgfr/constructTake My Repo has not tested this on your machine. Requirements come from the project's own documentation.