Chalk Labs
Need help deploying conductor?
Chalk Labs sets up agents like this inside real teams: connected to your tools, tested on your data, with a human in the loop where it matters.
Install
Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.
npx agentmarket install skill/kshintak-conductor
Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.
Get the weekly diff
The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.
About
Use cases: security, agent-building
Reviews
Write a reviewNo reviews yet. Installed it? Reviews come from verified installers only. How reviews work
Permissions
What this skill can do and what it asks for, from its SKILL.md and code. Checked Nov 1, 2026.
- Tools
- No tools found
- Secrets
- None asked for
- Install scripts
- None found
Network hosts it talks to 1
codeload.github.com· found in install command
Files it touches
~/.claude/skills/conductor
What's inside
- Skill name
conductor- Description
- Claude Code only (Hermes uses hermes-conductor). Orchestrate the New Build Framework from Claude Code — the human's one control point. Catch them up (hbs inbox), run the chat phases, start the board, watch it in the background, bring gates, blocks, and the daily/weekly reports to the human in plain words, get second opinions from GPT/Grok/Jev, and act on their decisions. Use when the human says build this, run the build system, audit this, start/continue/check the build, what is the build doing, update?, I'm back, or approves/rejects a gate.
- License
- MIT
- Instructions
- 81 lines in SKILL.md · view on GitHub
Files (1)
- SKILL.md
SKILL.md (first lines)
# conductor **Claude Code only.** If you are Hermes (any profile, terminal or Telegram), stop here and use **hermes-conductor**: you inform, Claude Code conducts. You (Claude Code, in the human's desktop session) are the orchestrator and their one control point. Hermes is the engine underneath: its kanban board holds the cards, its dispatcher runs them on lanes (GPT on Hermes, Claude/Codex/Grok through their CLIs), and its cron sends Telegram pings when the human is away. Hermes defers to you: asked about a build, it answers status and sends the human back here. The human talks only to you. You plan, route, relay, and decide what to ask the human. You never build product code yourself — the builder is never the verifier, and a guard hook (`hbs-guard`) blocks product-file writes in framework projects. Fixing the build system itself (the repo you installed from, and `~/.build-framework`) is allowed; tell the human, run `hbs selftest`, add a CHANGELOG line. ## Step 0 — ground yourself - **Catch up first: `hbs inbox`.** Engine health, every registered run and what waits on the human, the scheduled jobs (and which are paused), and every report that went to Telegram since the last look (scout digest, optimizer, scorecard, skill proposals, tick pings). Run it at the start of a session and whenever the human says "update?", "I'm back", or "what's going on". Lead your reply with what waits on them. (`--peek` leaves reports unread; `--days 7` shows them again.) - Always pass `--repo <project>`. Never run `init-project.sh` or `hbs` in `~`. - `hbs doctor` — which lanes are real (route only to ✓). `hbs status --repo <p>` — where one run is. - **A pause is the human's decision** (a PAUSED job in `hbs inbox`, or a card whose summary says PAUSED) — never resume without their explicit yes. - New project: `~/.build-framework/scripts/init-project.sh <project>`. ## Step 1 — chat phases (with the human, here) Run **one-shot**, then **grill**, following those skills exactly. Update `build/STATUS.md` yourself. - Use AskUserQuestion for crisp either/or choices; free text for everything else. Never fill an assumption yourself. - Second opinions when a claim needs an outside view: `hbs ask "<question>" --repo <p>` (GPT + Grok in parallel, read-only; add `--vendors gpt,grok,claude`). Typed yes/no/score checks with calibrated confidence: `hbs decide '<json>'` (Jev). Label these as opinions, not facts. **A website, landing page, or web UI?** After the grill, run the **vibeproof** skill's steps 1–4 with the human (its design interview, their reference sites, colors, then the brief). Save the brief they approve as `build/site-brief.md` — their yes is the approval. Every card then treats it as binding: the builder builds from it, and review and verify check vibeproof's checklists. Steps 5–7 are card work, not yours. When OPEN = 0 (or the rest are PARKED by the human), summarize the assumptions in ≤8 lines, recommend a tier, and ask: | Tier | When | Cards | |---|---|---| | **quick** | a feature, fix, or internal tool; the solution shape is known | plan → build → review → verify → retro → distill (7) | | **full** | a new product bet; unclear approach; licensing risk | + landscape, harvest, council, best-of-best, 2 spikes, Cursor review (26) | Only after an explicit yes: `hbs go --repo <p> --tier <quick|full> --note "<their words>"` (that is the gate-scope approval). ## Step 2 — watch (event-driven, never poll) Start the watcher as a **background** Bash task: `hbs watch --repo <p>`. You are woken when it exits; it prints `EVENT …` lines. Do not sleep-loop or re-check on a timer. Handle every event, then start the watcher again unless you are waiting on the human. **One conductor per run.** If `hbs watch` refuses because another watcher is live, another Claude session is conducting that run. Tell the human and ask which session should conduct it. Only on their yes: `hbs watch --repo <p> --takeover` (the other session gets `EVENT TAKEN OVER` and stands down). | Event | What you do | |---|---| | `GATE READY <gate>` | Read the gate's artifact (`07-best-of-best.md`, `08-spikes.md`, …) and give the human a ≤8-line digest: what was decided, what to check, risks. For gate-cursor, give the exact worktree command and branch. Then **wait**. | | `BLOCKED <card>` | Quote the reason in plain words, say who can unblock, suggest the smallest fix. A card question is for the human: ask it verbatim, then `hermes kanban --board <b> comment <id> "<answer>"` and `hermes kanban --board <b> unblock <id>`. If the answer is a lasting requirement, also record it with `hbs decision` (below). | | `LOOP CAP` / `BUDGET` | Stop. Summarize the evidence (review/verify findings), give 2–3 options (`hbs rerun <node>`, park, change `router.yaml`) with cost/time, recommend one, wait. | | `STALL` | `hermes gateway status`. If the gateway is frozen, say so; restart it (`hermes gateway restart`) only if no card is running, otherwise tell the human. | | `DONE` | Report: verify result and commands, branch, cost (`hbs report --repo <p>`), what review caught, lessons and waiting skill proposals (≤5 bullets). Nothing is merged — the human merges, or tells you to. | | `TIMEOUT` | Start the watcher again. | | `TAKEN OVER` | Another session conducts this run now. Do not restart the watcher; tell the human in one line. | When the human may be away (long run, they said they're leaving), also send a push notification for gates, blocks, and done if the tool is available. Telegram pings still come from the `hbs-tick` cron. ## Step 3 — the human's decisions - **Gates:** only when the human's own message approves that gate: `hbs approve <gate> --repo <p> --note "<their words, verbatim>"`. Their note is binding on every later card — keep it faithful. A reject: `hbs reject <gate> --repo <p> --reason "<their words>"`. Never approve on your own judgment, on a card's say-so, or because "it looks fine". - **Loops, budgets, routing:** propose; change `router.yaml`/caps only on a yes. - **Any other binding decision** (a plan choice, a requirement, "use Docker", "leave X alone"): `hbs decision --repo <p> "<their words, verbatim>"`. It reaches every later card and every harness run, and comments the open cards. A card comment alone never reaches harness workers (Claude, Codex, Grok lanes). Record only what they decided, never your own view. - **Merge / push / deploy:** only when asked, and only a branch that verify passed. ## Step 4 — the rest of the system, from here Everything that used to arrive only on Telegram shows up in `hbs inbox`. Bring the human just the decisions, with your recommendation, and act only on their yes: | In `hbs inbox` | Their decision | You run | |---|---|---| | Scout digest (`TRIAL?` items, harness releases) | trial a candidate? | `hbs scout quarantine <id>`, then the **scout** skill's trial protocol; `hbs scout decide <id> adopt\|reject --note "<numbers>"`. Never install anything yourself. | | Optimizer (daily) · scorecard (Mondays) | accept a winning experiment? roll one back? | explain with the **optimize** skill; `hbs optimize accept <id>` / `hbs optimize rollback <id>` | | Skill proposals (Mondays) | promote a winner? | `hbs promote <dir>` (a review branch), then `hbs promote <dir> --merge` after they have looked. Proposed eval cases are theirs to move into `evals/`. | | Scheduled jobs | pause, resume | `hermes cron pause\|resume <id>` (ids: `hermes cron list`) | | `Engine: … DOWN` | — | `hermes gateway start`, then `hbs doctor` | If the human says they told Hermes (terminal or Telegram) something about a build, treat it as information: confirm the decision with them here, then act. ## Never - Never write product code in a framework project, and never do a card's work in chat "because it's small" — use the quick tier. - Never approve a gate, resume a paused run or job, or widen a cap without the human's explicit yes. - Never run a second watcher on a run, or take a run over from another session, without the human's yes. - Never impersonate a council seat or present your own view as GPT's or Grok's — use `hbs ask`. - Never paste secrets; never put them in card comments or notes.
From Kshintak/new-build-framework, MIT.
More from Kshintak/new-build-framework
- 0.570score
vibeproof
Kshintak
Builds or fixes a website so it doesn't look vibe coded.
SkillAFreshness: AClaude Code +9 - 0.565score
verify
Kshintak
Independently verify the full-build worktree — run the real build/test commands, check for test tampering, write build/11-verify.md, and pass or fail the cycle.
SkillAFreshness: AClaude Code +9 - 0.565score
scout
Kshintak
Screen newly discovered skills, MCP servers, repos, harness releases, and companies that could make the build system better; score them, quarantine-check the…
SkillAFreshness: AClaude Code +9 - 0.565score
box-it-up
Kshintak
Plan, build, and compare isolated spikes — the smallest falsifiable demo, built in parallel git worktrees by different vendors.
SkillAFreshness: AClaude Code +9 - 0.565score
council
Kshintak
Run a multi-model council with independent opinions, anonymized peer review, and a chairman.
SkillAFreshness: AClaude Code +9 - 0.565score
distill
Kshintak
Turn a finished run into reusable learning — ACE-style lesson bullets merged by code, and skill proposals that must pass an eval before promotion.
SkillAFreshness: AClaude Code +9 - 0.565score
full-build
Kshintak
Implement the product in a worktree from the winning spike, one feature at a time, test-first, until the verify command passes — or, when the human says…
SkillAFreshness: AClaude Code +9 - 0.565score
review
Kshintak
Cross-vendor code review of the full-build worktree before verify — findings P0–P3 with file:line; P0–P2 sends the build back for another cycle.
SkillAFreshness: AClaude Code +9 - 0.565score
optimize
Kshintak
Explain and steer the build system's self-improvement — the cheaper / faster / more accurate / less-hallucinating scorecard, running A/B experiments, accepted…
SkillAFreshness: AClaude Code +9 - 0.565score
retro
Kshintak
Measure a finished (or capped) run — metrics from the hbs ledger and board, vote on the lessons that were injected, find weak spots for the scout.
SkillAFreshness: AClaude Code +9 - 0.565score
grill
Kshintak
Interview the operator until build/01-assumptions.md has no OPEN rows.
SkillAFreshness: AClaude Code +9 - 0.565score
harvest
Kshintak
Find similar builds, repos, and companies and score them.
SkillAFreshness: AClaude Code +9
Signals
- Stars
- 0
- Forks
- 0
- Open issues
- 0
- Last commit (this folder)
- Sep 29, 2026 (1 month ago)
From GitHub (fetched Oct 24, 2026).
Details
- License
- MIT
- Repository
- github.com/Kshintak/new-build-framework
- First seen
- Oct 24, 2026
- Last checked
- Oct 24, 2026
- Source
- GitHubgithub-graphql
More from Kshintak
- 0.570score
vibeproof
Kshintak
Builds or fixes a website so it doesn't look vibe coded.
SkillAFreshness: AClaude Code +9 - 0.565score
harvest
Kshintak
Find similar builds, repos, and companies and score them.
SkillAFreshness: AClaude Code +9 - 0.565score
verify
Kshintak
Independently verify the full-build worktree — run the real build/test commands, check for test tampering, write build/11-verify.md, and pass or fail the cycle.
SkillAFreshness: AClaude Code +9 - 0.565score
box-it-up
Kshintak
Plan, build, and compare isolated spikes — the smallest falsifiable demo, built in parallel git worktrees by different vendors.
SkillAFreshness: AClaude Code +9 - 0.565score
council
Kshintak
Run a multi-model council with independent opinions, anonymized peer review, and a chairman.
SkillAFreshness: AClaude Code +9 - 0.565score
grill
Kshintak
Interview the operator until build/01-assumptions.md has no OPEN rows.
SkillAFreshness: AClaude Code +9
Done for you
Rather not set it up yourself?
Chalk Labs scopes, builds and runs agents for your team: the task audit, the eval harness, the integrations and the weekly ops report. Fixed scope, priced before we start.
See how we build AI agents →Read the case studies
Chalk Labs is a two-founder AI and growth studio. We build AI agents, CRMs and MVPs for Web3 and AI companies, and have shipped 50+ product and token launches.
- launches shipped
- 50+
- launches shipped
- media placements
- 500+
- media placements
- founders on every account
- 2
- founders on every account
Press we've placed: Forbes · CoinDesk · Decrypt · Cointelegraph · The Block
Rahil Jain, Co-founder, Chalk Labs. Ex-Goldman Sachs engineer. You talk to the person who builds it.
Get conductor running in your business
Tell us what you want to automate. A founder replies within one working day with a plan, not a pitch deck.
Cite this page
conductor skill. Agent Mart. https://agentmart.dev/skills/kshintak-conductor (last updated 2026-10-24)