| name | forge-bootstrap-project |
| description | Meta-skill that bootstraps a new project end-to-end from a one-liner idea by chaining the deterministic Forge skills: forge-build-prd → (pause for review) → forge-build-agent-team → (pause for review) → optionally forge-assign-models. Use this skill when a user wants to go from idea to a reviewable PRD and a generated agent team in a single guided flow, without losing the human review gates between steps.
|
Skill: Bootstrap a Project from an Idea (Meta-Skill)
You are orchestrating the standard Agent Forge bootstrap flow on behalf of the
user. Your job is to remove copy-paste friction between the deterministic Forge
skills without removing the human review gates between them.
You do not re-implement what the underlying skills do. You invoke them in
order, pause at each handoff, present a verification checklist, and only
continue when the user explicitly approves.
This skill chains:
forge-build-prd — interview the user and produce docs/PRD.md.- Pause 1 — PRD review gate (mandatory).
forge-build-agent-team — generate the specialist agent team from the
approved PRD.- Pause 2 — Agent team review gate (mandatory).
forge-assign-models — discover models and recommend a per-agent
assignment (optional, opt-in at Pause 2).
Operating Principles
- Never skip a pause. Even if the user says "just do everything," you still
stop at each gate, present the checklist, and require an explicit "approved"
/ "continue" / "looks good" before moving on. The pauses exist because PRDs
and agent teams are high-leverage artifacts that are expensive to fix later.
- Do not duplicate the underlying skills' work. When it is time to run a
step, invoke the corresponding skill (
forge-build-prd,
forge-build-agent-team, forge-assign-models) and let it own its full
process — including any clarifying questions it normally asks.
- Preserve all existing outputs. The artifacts produced (
docs/PRD.md,
.github/agents/*.md, .github/skills/*/SKILL.md, docs/MODEL-PLAN.md,
etc.) must be identical to what you would get by running the underlying
skills directly. This skill is glue, not a rewrite.
- Stay inside the existing model. You are one skill calling the procedures
of other skills. Do not introduce new files, formats, or tools beyond what
the underlying skills already produce.
- Be explicit about which step you are in. At every message, name the
current step (e.g., "Step 1 of 3: forge-build-prd") so the user knows where
they are in the flow.
- Resumability. If the user starts this skill in a repo that already has a
docs/PRD.md or existing agents in .github/agents/, detect it and offer to
resume from the appropriate step instead of overwriting.
Process
Step 0: Confirm the Idea and the Flow
The user invokes this skill with (typically) a one-liner idea, e.g.:
/forge-bootstrap-project I want to build a CLI that summarizes my git history into a weekly changelog.
Do the following before invoking any other skill:
-
Echo the idea back in one or two sentences so the user can confirm you
understood it.
-
State the flow explicitly:
- Step 1:
forge-build-prd → produces docs/PRD.md
- Pause 1: PRD review
- Step 2:
forge-build-agent-team → produces agent + skill files
- Pause 2: Agent team review
- Step 3 (optional):
forge-assign-models → recommends per-agent models
(These are Steps 1–3 of the bootstrap flow itself; this confirmation is
Step 0 — the pre-flight before any underlying skill is invoked.)
-
Check repo state and flag anything that affects the flow:
- Does
docs/PRD.md already exist? If yes, ask whether to keep, replace,
or extend it before running Step 1.
- Do agent files already exist in
.github/agents/ (beyond the forge
templates project-orchestrator and forge-team-builder)? If yes,
warn that Step 2 will run in Feature Increment Mode rather than
Full Build, and confirm whether the user wants that.
- Does
docs/product-vision.md with docs/features/*.md exist? If yes,
Step 2 will run in Vision + Features Mode; confirm with the user.
-
Wait for confirmation. Do not proceed to Step 1 until the user says to
start.
Step 1: Run forge-build-prd
Invoke the forge-build-prd skill, passing the user's idea as the input.
Let that skill drive its own clarifying-questions process; do not answer on
the user's behalf and do not collapse its interview into a single batch.
When forge-build-prd finishes and docs/PRD.md is saved, transition to
Pause 1. Do not silently continue into Step 2.
Pause 1: PRD Review Gate (mandatory)
Post a message that clearly says you are paused. Include:
- The path to the artifact:
docs/PRD.md
- A one-paragraph summary of what was produced (scope, key tech stack
choices, number of phases, number of functional requirements).
- A list of Open Questions the PRD itself flagged (copy them out of the
PRD's "Open Questions" section so the user does not have to hunt).
- The verification checklist below.
- An explicit prompt: "Reply
approved to continue to Step 2
(forge-build-agent-team), revise: <notes> to iterate on the PRD, or
stop to end here."
PRD review checklist (emit verbatim):
PRD review checklist — verify before continuing:
Scope & intent
- [ ] The Overview matches the idea you actually want to build
- [ ] Goals and Non-Goals are correct (nothing important is missing, nothing
out-of-scope has crept in)
- [ ] Target users / personas are right
Requirements
- [ ] Every must-have capability you care about appears as a functional
requirement with a priority
- [ ] Non-functional requirements (performance, security, privacy,
accessibility) reflect your real constraints
- [ ] Security & Privacy section is correct (data handling, auth, compliance)
Technical choices
- [ ] Technology stack is current, available to you, and acceptable
- [ ] Project Structure is something you are willing to live with
- [ ] No deprecated or end-of-life dependencies were selected
Plan
- [ ] Implementation Phases are ordered correctly and each phase is shippable
- [ ] Testing Strategy matches how you actually plan to validate the product
- [ ] Acceptance Criteria are concrete enough that "done" is unambiguous
Open items
- [ ] Every entry under "Open Questions" has either an answer or an explicit
"accept the default assumption" decision
- [ ] Any flagged risks have a mitigation you can live with
Do not invoke forge-build-agent-team until the user responds with an
approval keyword (e.g., approved, continue, looks good, proceed).
If the user replies with revisions, hand control back to forge-build-prd
to iterate, then re-present Pause 1 with the updated checklist.
Step 2: Run forge-build-agent-team
After Pause 1 is approved, invoke the forge-build-agent-team skill against
the approved PRD (or against docs/product-vision.md + docs/features/*.md
if that layout exists). Let that skill detect its own mode (Full Build,
Vision + Features, or Feature Increment) — do not override its mode
detection.
When it finishes and agent files have been written under .github/agents/
(and any skills under .github/skills/), transition to Pause 2.
Pause 2: Agent Team Review Gate (mandatory)
Post a paused message. Include:
- The list of agent files written or modified, with a one-line role
description for each.
- The list of skill files written or modified, if any.
- A pointer to the responsibility matrix if one was produced (e.g.,
docs/agent-responsibility-matrix.md).
- The verification checklist below.
- An explicit prompt: "Reply
approved and assign models to continue to
Step 3 (forge-assign-models), approved (without models) to stop here
with the team in place, revise: <notes> to iterate on the team, or
stop to end here."
Agent team review checklist (emit verbatim):
Agent team review checklist — verify before continuing:
Coverage
- [ ] Every PRD requirement (or every feature in docs/features/) maps to
exactly one primary owner agent
- [ ] No requirement is unowned
- [ ] No two agents claim primary ownership of the same area
Boundaries
- [ ] Each agent's responsibilities are scoped to a clear domain (no
"do everything" agents)
- [ ] Cross-cutting concerns (testing, security, docs, ops) have explicit
owners rather than being implied
Hygiene
- [ ] Each agent file has valid YAML frontmatter and the `name:` matches
the filename
- [ ] The forge templates (`project-orchestrator`, `forge-team-builder`)
are still present and untouched
- [ ] Any newly generated skills have a `SKILL.md` and follow the existing
skill format
Fit
- [ ] The team size feels right for the project (not so large that
coordination dominates, not so small that agents are overloaded)
- [ ] You are comfortable handing each agent its stated scope
Do not invoke forge-assign-models unless the user explicitly opts in at
this gate. If the user replies with only approved (no model-assignment
request), end the flow here and tell them they can run
forge-assign-models later at any time.
If the user replies with revisions, hand control back to
forge-build-agent-team to iterate, then re-present Pause 2 with the
updated checklist.
Step 3 (Optional): Run forge-assign-models
Only run this step if the user explicitly opted in at Pause 2.
Invoke the forge-assign-models skill. Default to its Recommend mode —
which produces docs/MODEL-PLAN.md without modifying agent files — unless
the user explicitly asked for Apply. This preserves the underlying
skill's own opt-in safety: nothing gets written into agent YAML without a
second confirmation.
When the skill finishes:
- If it ran in Recommend mode: point the user at
docs/MODEL-PLAN.md
and tell them how to invoke forge-assign-models again in Apply mode
when ready.
- If it ran in Apply mode: list the agent files that were modified
and summarize which model each agent received.
End the flow with a short summary of what was produced and the recommended
next step (typically: commit the changes, then invoke
@project-orchestrator to start executing Phase 1).
Guidelines
- Be a conductor, not a soloist. The substantive work — interviewing the
user, drafting the PRD, designing the team, classifying workloads — belongs
to the underlying skills. Your value is sequencing, pausing, and surfacing
the right checklist at the right moment.
- Never collapse a pause into a "looks fine, moving on" message. The
whole point of this meta-skill is that the review gates remain real.
- Repeat checklists verbatim. Do not paraphrase the checklists; users
rely on them being stable across runs.
- Quote, do not invent. When summarizing Open Questions or agent roles
at a pause, pull the text from the artifact rather than restating from
memory.
- Idempotent re-entry. If the user re-invokes this skill mid-flow
(e.g., after closing their session), inspect the repo state and resume
from the earliest step whose artifact is missing or unapproved, rather
than starting over.
- No new file formats. Do not introduce a new state file, manifest, or
config to track progress between pauses. The artifacts on disk
(
docs/PRD.md, .github/agents/*.md, docs/MODEL-PLAN.md) are the
state.
