truestack
A lightweight, honesty-first coding skill set for Claude Code — packaged as a plugin
that bundles 23 skills, 8 slash commands, and an MCP integration layer, all behind a single
truestack-orchestrate router that reads the request, right-sizes it, and dispatches to the right
skill in the right order. Language-aware backend skills (Express · .NET · Python), a pro React
frontend skill, multi-agent / multi-session coordination, an always-on honesty contract,
committed repo-memory, verified deep research, scheduling, and real tool integrations —
with process that sizes itself to the work.
Highlights:
truestack-orchestraterouter — classifies any request, right-sizes it, runs the canonical chain, and gates everything throughtruestack-quality-control; routes to the best skill even when it lives in another installed set.- Enforced governance — a PreToolUse hook denies catastrophic and asks on money/destructive/schema/outbound tool calls (84-case tested). Enforced once wired: automatic with plugin install, a one-time settings merge with drop-in install (step 2 below). The append-only MCP approval log is skill-directed (the model writes it under the
truestack-mcp-integrationskill), not hook-enforced. - Honesty, machine-checked — an always-on grounding contract, a code↔memory reconciliation tally, and auto-research of current-fact decisions from authoritative sources.
- Full lifecycle — prompt optimization · plan · backend/frontend · API contract · migrations · deploy · CI/CD · observability · security · dependencies · data-privacy · debugging · reverse-engineering · research · scheduling · multi-agent coordination · self-evaluation · plain-English explanation.
- Self-measuring — ships its own deterministic skill lint + a behavioral routing eval, both run in CI.
The skills (23)
| Skill | Fires on | Job |
|---|---|---|
truestack-orchestrate |
first, on any non-trivial request | front-door router: ground → right-size → route to the right skill(s) → run the canonical chain → gate through truestack-quality-control |
truestack-prompt-optimizer |
every truestack-orchestrate handoff / "optimize my prompt" / "sharpen this prompt" / "adopt an expert persona" |
turn a raw prompt into an optimized brief — task type, required components (ask when a required one can't be inferred), zero-shot / few-shot / chain-of-thought technique, matching persona — before any skill runs on it |
truestack-project-memory |
first task / "remember" / stale facts | study the repo, maintain committed memory + the always-on contracts (honesty, communication, clarify, boundaries, tool-use) |
truestack-architecture-planning |
designing/scoping new work | right-size → clarify (loops) → architecture → testable criteria → tasks (parallel vs sequential) → gate risk |
truestack-backend-development |
implementing backend | match/recommend stack (Express · .NET · Python) → accurate, load-safe code; build loop to acceptance, or an honest score |
truestack-react-frontend |
building/reshaping React UI | distinctive design and sound React engineering (arch, state, perf, a11y, browser + visual QA) |
truestack-root-cause-debugging |
a bug/error/regression | reproduce → confirm root cause → permanent fix → verify-or-loop until proven fixed |
truestack-quality-control |
after any change | tests (+property) · types/lint · 6-axis review · load/perf · safety (OWASP) · intent check |
truestack-agent-coordination |
parallel agents / research fan-out / continuing a session | decompose → worktree isolation → shared task ledger → merge protocol |
truestack-mcp-integration |
a task needs live data or a real external action | use connected MCP tools with accuracy/security/honesty discipline; configure .mcp.json |
truestack-database-migrations |
schema/data change · migration · backfill (high-risk) | expand/contract · zero-downtime DDL · reversible up/down · idempotent batched backfills |
truestack-deploy-and-runtime |
deploy/run on a VPS · "my deploy causes downtime" | Docker · nginx · TLS · systemd · SIGTERM drain · blue-green + instant rollback |
truestack-ci-and-delivery |
CI/CD · GitHub Actions · release · "ship it" | pipeline gates · SHA-pinned actions · required checks · semver release · health-gated deploy |
truestack-observability |
logging/metrics/tracing · SLOs · "no telemetry yet" | structured logs · OTel metrics/traces · RED/USE · burn-rate alerts · bounded on one box |
truestack-application-security |
"is this secure" · auth · access control · OWASP · injection · secrets | authn/authz (deny-by-default, IDOR/BOLA) · OWASP Top 10 · structural injection prevention · Argon2id · CSRF/SSRF · STRIDE |
truestack-api-design |
API contract · REST/GraphQL/gRPC · versioning · pagination · idempotency | contract-first · RFC 9457 errors · cursor pagination · idempotency keys · OpenAPI/contract tests |
truestack-dependency-management |
add/pin a dependency · CVE/SBOM/license · "is this package safe" | lockfile/pinning · Renovate cooldown · CVE/GHSA triage · SBOM · license · typosquat/SLSA |
truestack-data-privacy |
PII · GDPR/CCPA · right-to-erasure · retention · consent | PII inventory/classification · retention + erasure · consent · audit logging · breach readiness |
truestack-deep-research |
"research / compare / what's best/latest" | parallel search → primary sources → adversarial verification → cited answer + confidence |
truestack-reverse-engineering |
shared code/reference · "how does this work" · "adopt this" · upgrade from a reference | verified model (verified vs inferred) → transferable idea → gated, license-aware upgrade path |
truestack-task-scheduling |
"every day / weekly / remind me / run at" | self-contained job spec (trigger · run prompt · delivery · failure policy) wired to the host scheduler |
truestack-skill-evaluation |
"score / rate / audit / improve" a skill | static lint + judge + behavioral trigger test → scorecard with fixes; its own quality gate |
truestack-explain-plain |
after a build/fix/design lands / "explain this" / "what does this do" / "I don't get it" | explain finished work or existing behavior in short, beginner-friendly plain English — the lesson is the deliverable |
Slash commands (8)
| Command | Does | Routes to |
|---|---|---|
/truestack-verify [scope] |
full QC sweep + verdict | truestack-quality-control |
/truestack-loop <target> |
iterate to proven-done, or honest score | truestack-backend-development / truestack-root-cause-debugging |
/truestack-simplify [target] |
maintainability simplify pass, behavior unchanged | truestack-quality-control |
/truestack-run [tests|build|lint|all] |
run the stack's checks, report in a table | truestack-quality-control tooling |
/truestack-deep-research <question> |
verified, cited multi-source answer | truestack-deep-research |
/truestack-schedule <what> <when> |
set up a recurring/deferred run | truestack-task-scheduling |
/truestack-onboard [focus] |
study repo + build committed memory | truestack-project-memory |
/truestack-eval [target] |
score/audit skills — lint + judge + trigger test | truestack-skill-evaluation |
/truestack-onboard is intentionally not /init (that name is a built-in Claude Code command).
Integrations (MCP) — guidance that can finally act
truestack-mcp-integration turns the set from advice into action through connected MCP servers, while
keeping the discipline: discover before you assume, treat tool output as untrusted data,
read freely / write carefully, and keep money, destructive ops, schema changes, and
outbound sends Ask-first. Idempotency on retried writes; verify the real effect afterward. Every Ask-first call also leaves an approval receipt + append-only audit row (.ai/agents/mcp-audit.md, written by the model under the skill's direction — not machine-enforced) so external effects are reviewable like code.
Servers are declared in a committed, secret-free .mcp.json (env-var expansion ${VAR});
this ships with an empty .mcp.json plus a .mcp.example.json template — connect only what a
project needs. Full schema + per-category checklist (DB · vector · payments · webhooks · search)
in skills/truestack-mcp-integration/references/mcp-config.md.
Honesty & anti-hallucination (always-on)
Seeded into CLAUDE.md so it applies to every reply, with the full contract in
skills/truestack-project-memory/references/honesty.md. In impact order: ground, don't recall;
abstain when evidence is thin; verify consequential claims before finalizing; truth
over agreement; separate verified / inferred / unknown; never fabricate paths, APIs, or
numbers. Hallucination is structural — the contract makes it rare and always flagged, not "solved."
Multi-agent & multi-session (no overwrite, no collapse)
truestack-agent-coordination separates isolation (each writing agent in its own git worktree + branch)
from coordination (shared contracts up front, a task ledger where each agent writes only its
own row, non-overlapping scope, defined merge order). Memory writes are append-or-section; a
later session resumes from committed memory + the ledger + status files. Caps at 2–3 agents to start.
Language-aware backend & pro React frontend
truestack-backend-development/truestack-quality-control adapt to the project's language — stack idioms live in
truestack-backend-development's references/stacks.md, and the OWASP-aligned security checklist for
Express.js · .NET · Python lives in truestack-quality-control's references/security.md.
truestack-react-frontend holds two bars at once — anti-AI-slop visual design and the engineering a
design-only skill skips (architecture, state/data, React 19, Core Web Vitals, a11y, testing).
How they chain
- Entry point:
truestack-orchestratereads the request, right-sizes it, and dispatches into one of the chains below (skip it for a trivial one-off). - Backend:
truestack-architecture-planning→truestack-backend-development→truestack-quality-control - Frontend:
truestack-architecture-planning→truestack-react-frontend→truestack-quality-control - External effects:
truestack-backend-development→truestack-mcp-integration→truestack-quality-control - Parallel:
truestack-architecture-planning→truestack-agent-coordination→ (workers runtruestack-backend-development/truestack-react-frontend) →truestack-quality-control - Fixes:
truestack-root-cause-debugging→truestack-quality-control - Research / recurring:
truestack-deep-research(←truestack-agent-coordinationfan-out); wrap withtruestack-task-schedulingto recur - All read project memory first;
truestack-project-memoryowns it.
Install
This repo is itself a valid Claude Code plugin (.claude-plugin/plugin.json +
marketplace.json, skills/, commands/, .mcp.json). Pick one method.
Names are already namespaced. Every skill and command carries a
truestack-prefix (truestack-orchestrate,/truestack-verify, …) so they never collide with other installed sets. This means drop-in is the recommended install — it keeps those flat names as-is. Plugin-marketplace install also namespaces by plugin id, so it would double up totruestack:truestack-orchestrate; use it only if you prefer plugin-managed updates.
Drop-in, kept current with git pull — recommended. Two steps: link the skills/commands,
then wire the hooks (skills and commands auto-load from ~/.claude/; hooks never do — an
unwired gate is silently dead).
Step 1 — link. macOS / Linux (symlinks):
git clone <your-fork-url> ~/.claude/truestack
mkdir -p ~/.claude/skills ~/.claude/commands
for d in ~/.claude/truestack/skills/truestack-*; do ln -s "$d" ~/.claude/skills/; done
for f in ~/.claude/truestack/commands/truestack-*.md; do ln -s "$f" ~/.claude/commands/; done
# update anytime: cd ~/.claude/truestack && git pull
Windows (PowerShell — junctions for skill folders need no admin or Developer Mode; command
files can't be junctioned, so copy them and re-copy after each git pull):
git clone <your-fork-url> "$HOME\.claude\truestack"
New-Item -ItemType Directory -Force "$HOME\.claude\skills","$HOME\.claude\commands" | Out-Null
Get-ChildItem "$HOME\.claude\truestack\skills\truestack-*" -Directory | ForEach-Object {
New-Item -ItemType Junction -Path "$HOME\.claude\skills\$($_.Name)" -Target $_.FullName }
Get-ChildItem "$HOME\.claude\truestack\commands\truestack-*.md" | ForEach-Object {
Copy-Item $_.FullName "$HOME\.claude\commands\$($_.Name)" -Force }
# update anytime: cd "$HOME\.claude\truestack"; git pull # then re-run the Copy-Item loop
(True symlinks on Windows need an elevated shell — or PowerShell 7+ with Developer Mode;
Windows PowerShell 5.1 can't create them unelevated even with Developer Mode. Junctions avoid
all of that.) Or run the bundled one-shot: powershell -ExecutionPolicy Bypass -File .\install.ps1 -WireHooks
— it links/copies everything, migrates a stale pre-namespacing install to a backup folder, and
wires the hooks (Step 2) with a settings backup.
Step 2 — wire the hooks. Merge this into ~/.claude/settings.json (adjust the clone path;
full options + per-project variant in hooks/README.md):
{ "hooks": {
"PreToolUse": [ { "matcher": "*", "hooks": [
{ "type": "command", "command": "node \"$HOME/.claude/truestack/hooks/pretooluse-gate.mjs\"", "timeout": 30 } ] } ],
"UserPromptSubmit": [ { "hooks": [
{ "type": "command", "command": "node \"$HOME/.claude/truestack/hooks/truestack-orchestrate-reminder.js\"", "timeout": 10 } ] } ]
} }
Verify: node ~/.claude/truestack/hooks/test-gate.mjs (84/84) and /hooks inside Claude Code.
Finally, merge the truestack/.mcp.example.json entries you need into your project .mcp.json.
Drop-in, one-time copy (simplest; re-copy after each git pull):
git clone <your-fork-url> truestack
cp -r truestack/skills/* ~/.claude/skills/ # the 23 truestack-* skills
cp -r truestack/commands/* ~/.claude/commands/ # the 8 /truestack-* slash commands
Then do Step 2 — wire the hooks above (copying files never wires hooks).
As a plugin via marketplace (double-prefixes the names — see note above; hooks auto-load, no Step 2 needed):
git clone <your-fork-url> truestack
Then, in Claude Code:
/plugin marketplace add ./truestack
/plugin install truestack@truestack
Validate the bundle anytime with claude plugin validate .. Claude Code auto-detects skills and
commands — no restart. Configure .mcp.json with your own servers (secrets via env vars) before
using truestack-mcp-integration.
No comments yet
Be the first to share your take.