refine-spec
Read a draft spec file (idea.md, mvp.md, prd.md, brief.md, plan.md, spec.md, etc.) and interactively quiz the user across multiple rounds with AskUserQuestion…
Skill by tcbuilds
Chalk Labs
Need help deploying refine-spec?
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/tcbuilds-refine-spec
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: documents
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. Checked Oct 25, 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/refine-spec
What's inside
- Skill name
refine-spec- Description
- Read a draft spec file (idea.md, mvp.md, prd.md, brief.md, plan.md, spec.md, etc.) and interactively quiz the user across multiple rounds with AskUserQuestion to fill gaps, resolve ambiguity, and revamp the document into its final form. Use when asked to "refine my idea", "tighten this spec", "turn this into an MVP doc", "interrogate my brief", "polish this PRD", "ask me questions about idea.md", or when the user points at a draft document and wants it sharpened through Q&A.
- License
- MIT
- Instructions
- 99 lines in SKILL.md · view on GitHub
Before installing
- Instructions read files in your home folder (~/.ssh, ~/.aws…)
Files (1)
- SKILL.md
SKILL.md (first lines)
# Refine Spec Transform a rough draft spec into a finalized document by reading it, identifying gaps, asking the user targeted clarifying questions in multiple rounds via AskUserQuestion, then rewriting the file in place. ## Inputs The user will name a target file (e.g., `idea.md`, `mvp.md`, `prd.md`, `spec.md`, `brief.md`). If no path is given, search the current directory for likely candidates and confirm with the user before proceeding. ## Process ### 0. Discover local context (greenfield projects only — skip for in-progress refines) **This step matters most when the user is starting fresh on a box they already operate.** Before asking the user *any* question about stack/infra/conventions, do a non-blocking scan to surface what they already run. Defaults proposed in later rounds should match the user's existing patterns, not generic best-practice. What to scan: ```bash ls ~/ # sibling project dirs — adjacent stacks cat ~/*/CLAUDE.md 2>/dev/null # absorb existing CLAUDE conventions cat ~/*/AGENTS.md 2>/dev/null # absorb existing AGENTS conventions gh repo list <user> --limit 50 # existing repos available as targets systemctl list-units --type=service \ --state=running 2>/dev/null # already-running services on the box ls ~/.config/ 2>/dev/null # existing tool configs (cloudflared, rclone, gh, etc.) which docker podman caddy nginx \ cloudflared rclone 2>/dev/null # already-installed infra binaries test -f ~/.cloudflared/config.yml \ && cat ~/.cloudflared/config.yml # cloudflared tunnel ingress (if present) ``` For each discovered tool/service/repo, note: - Is the user *already* using this for similar work? - Could the spec adopt it instead of proposing a fresh greenfield choice? Surface findings to the user in a single tight summary before Round 1, e.g.: > "Discovery scan: you already run cloudflared (tunnel for `<existing-site>`), have a Cloudflare account, run nginx for `<other-site>`, have repos `<a>`, `<b>`, `<c>` available as content targets, and use `~/golden-horizons/CLAUDE.md` patterns. I'll bias Round 1 toward reusing these. Push back if you want greenfield instead." This gates against the failure mode where the skill defaults to "best practice" choices (e.g., proposing Backblaze when the user already pays Cloudflare; proposing Caddy when the user already runs cloudflared tunnels; proposing fresh repos when umbrella patterns exist). **Skip Round 0** if the draft spec already references the user's actual stack OR if the user explicitly asks for a greenfield-best-practice pass. ### 1. Read & Diagnose Read the target file in full. Build a mental model of: - **Stated**: what the doc explicitly claims/specifies - **Implied**: what's hinted at but not nailed down - **Missing**: critical sections a finalized version of this doc-type needs but lacks - **Contradictions / ambiguity**: anything that conflicts or could be read multiple ways Match the document's apparent type (idea note, MVP scope, PRD, technical spec, marketing brief, etc.) and use that to anchor what "finalized form" means. ### 2. Plan Question Rounds Use Round 0 findings to bias every default in every later round. When an option matches the user's existing stack, mark it "(Recommended)" with the discovery reference (e.g., "Cloudflare R2 (Recommended — you already pay Cloudflare)"). Generic best-practice options come second. Group questions into **3–5 thematic rounds**, ordered from foundational to detailed. Typical progression: 1. **Vision / problem** — who is this for, what problem, why now 2. **Scope** — in vs out, MVP vs later, success criteria 3. **Mechanics** — how it works, key flows, constraints 4. **Edge cases & open decisions** — tradeoffs, risks, unresolved choices 5. **Polish** — naming, tone, format, distribution Skip rounds that the draft already answers cleanly. Add rounds if the doc-type demands it. ### 3. Ask in Batches via AskUserQuestion For each round, batch **2–4 related questions** into a single AskUserQuestion call (the tool supports up to ~5 questions per call). For each question: - Provide 3–4 concrete multiple-choice options when the answer space is enumerable - Always include an open-ended escape hatch by letting the user supply a custom answer - Keep question text tight; put nuance in the option labels Between rounds, briefly acknowledge what you learned (one sentence) before launching the next batch — so the user feels heard, not interrogated. ### 4. Synthesize & Rewrite After the final round: - Rewrite the target file in place with Edit/Write - Preserve the user's voice and any strong existing prose - Add/restructure sections so the doc reads as a finalized version of its type - Mark any remaining unknowns explicitly as `**Open question:**` rather than guessing - Do **not** invent facts the user didn't supply — the rule is fill from answers, flag the rest ### 5. Confirm End with a short summary: what changed, what's still open, suggested next step (e.g., "ready to share with stakeholders" or "needs one more pass on pricing"). ## Guardrails - **Discover before defaulting.** Run §0 (Discovery) first on greenfield projects. Bias every Round 1+ default toward the user's existing stack. Generic best-practice answers are the fallback, not the lead. - **Don't ask everything at once.** Multi-round is the point — it gives the user time to think and you time to adapt later questions to earlier answers. - **Don't ask what the doc already answers.** Read first, ask second. - **Don't ask what discovery already answers.** If `~/.cloudflared/config.yml` exists, don't ask "what reverse proxy?" — propose the existing tunnel. - **Don't fabricate.** If the user skips a question, mark the section as open rather than filling with plausible-sounding content. - **Stay in the file.** Edit the target in place. Don't spawn `idea-v2.md`, `idea-final.md`, etc. unless the user asks. - **Stop when diminishing returns hit.** If the user starts giving terse "whatever you think" answers, wrap up — don't drag them through more rounds.
From tcbuilds/nautilus, MIT.
More from tcbuilds/nautilus
- 0.570score
prompt-injection-audit
tcbuilds
Run a read-only, evidence-based audit of an LLM or agent application for prompt injection and the failures that make injection matter.
SkillAFreshness: AClaude Code +9 - 0.566score
codex-implement
tcbuilds
Execute one implementation slice or an implementation_plan.md task, phase, or full plan through parallel file-owned Codex CLI workers in isolated Git worktrees.
SkillAFreshness: AClaude Code +9 - 0.565score
codex-review
tcbuilds
Run a second-pass Codex review over changed files, produce severity-ordered findings, and escalate repeated fix-loop invariants to a Sol-max repair architect.
SkillAFreshness: AClaude Code +9 - 0.529score
pi-implement
tcbuilds
Execute one implementation slice or an implementation_plan.md task, phase, or full plan through Pi-native subagent workflows.
SkillBFreshness: BClaude Code +9 - 0.524score
roadmap
tcbuilds
Create or update implementation_plan.md with the smallest evidence-backed set of Markdown task checkboxes that fully satisfies the request.
SkillBFreshness: BClaude Code +9 - 0.524score
warmup
tcbuilds
Warm up a fresh Claude/Codex session by loading repo-local context.
SkillBFreshness: BClaude Code +9 - 0.524score
pi-review
tcbuilds
Run an independent Pi-native review over changed files using fresh-context reviewer subagents, evidence-backed severity findings, parent synthesis…
SkillBFreshness: BClaude Code +9 - 0.524score
repo-orientation
tcbuilds
Build an onboarding-quality breakdown of an existing repository.
SkillBFreshness: BClaude Code +9 - 0.520score
claude-build
tcbuilds
Claude Code-only implementation workflow for implementation_plan.md using Claude specialist agents, phase-boundary commits, and the codex-review gate.
SkillBFreshness: BClaude Code +9 - 0.494score
bro
tcbuilds
Restate the last message in plain human language, with no jargon.
SkillBFreshness: BClaude Code +9 - 0.479score
nautilus-sync
tcbuilds
Sync a personal skill or agent from ~/.claude/ into a public-facing nautilus-shaped playbook repo as a sanitized loadable artifact.
SkillCFreshness: CClaude Code +9 - 0.474score
handoff
tcbuilds
Capture the state of an in-flight session into HANDOFF.md so a future session, teammate, or agent can resume without re-deriving context.
SkillCFreshness: CClaude Code +9
Signals
- Stars
- 0
- Forks
- 0
- Open issues
- 0
- Last commit (this folder)
- May 17, 2026 (5 months ago)
From GitHub (fetched Oct 24, 2026).
Details
- License
- MIT
- Repository
- github.com/tcbuilds/nautilus
- First seen
- Oct 24, 2026
- Last checked
- Oct 24, 2026
- Source
- GitHubgithub-graphql
More from tcbuilds
- 0.570score
prompt-injection-audit
tcbuilds
Run a read-only, evidence-based audit of an LLM or agent application for prompt injection and the failures that make injection matter.
SkillAFreshness: AClaude Code +9 - 0.566score
codex-implement
tcbuilds
Execute one implementation slice or an implementation_plan.md task, phase, or full plan through parallel file-owned Codex CLI workers in isolated Git worktrees.
SkillAFreshness: AClaude Code +9 - 0.565score
codex-review
tcbuilds
Run a second-pass Codex review over changed files, produce severity-ordered findings, and escalate repeated fix-loop invariants to a Sol-max repair architect.
SkillAFreshness: AClaude Code +9 - 0.529score
pi-implement
tcbuilds
Execute one implementation slice or an implementation_plan.md task, phase, or full plan through Pi-native subagent workflows.
SkillBFreshness: BClaude Code +9 - 0.524score
pi-review
tcbuilds
Run an independent Pi-native review over changed files using fresh-context reviewer subagents, evidence-backed severity findings, parent synthesis…
SkillBFreshness: BClaude Code +9 - 0.524score
repo-orientation
tcbuilds
Build an onboarding-quality breakdown of an existing repository.
SkillBFreshness: BClaude 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 refine-spec 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
refine-spec skill. Agent Mart. https://agentmart.dev/skills/tcbuilds-refine-spec (last updated 2026-10-24)