一键导入
roadmap-bootprompt
Author the boot prompt — capture session-tacit cognitive stance before context dies
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
Author the boot prompt — capture session-tacit cognitive stance before context dies
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Assess convergence, review session, write successor, close cleanly
Generate convergence-oriented roadmap specs
Autonomous roadmap execution with rich reporting
Run ad-hoc cause-routing diffusion whenever a roadmap node's REALIZED value falls short of its PROJECTED value — not only on RED/AMBER/BLOCKED, but on a GREEN that under-delivered (a shallow green). Identifies actual cause vs assumed cause, enumerates recovery axes, dispatches a non-symbolic recovery worker, and accepts the shortfall only when every axis yields a cited empirical zero. The general roadmap diffusion trigger; project-agnostic. Sibling to /core-loop · /cross-page-sweep · /diffuse-the-pipeline. Supersedes the narrower diffuse-on-not-green, which fired only on non-green and missed the shallow green entirely.
Self-orient at session start — roadmap position, fleet state, what to do next
| name | roadmap-bootprompt |
| description | Author the boot prompt — capture session-tacit cognitive stance before context dies |
| user-invocable | true |
The spec encodes what to prove. This skill encodes how the drafting session wanted to think. Drift-prevention, dead ends, register, user concerns, the micro-doctrine that emerged in conversation but isn't formal enough for CLAUDE.md — all of it dies when the session ends unless captured now.
A fresh agent in a future session loads r<N>.boot.md via /roadmap-orient and inherits the cognitive stance. Without this skill, that residue is lost and the next session re-discovers (or worse, re-falls-into) the same patterns.
🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪
PRIMARY immediately after /roadmap-spec runs `roadmap make`
session-tacit knowledge is still alive · capture before it dies
SECONDARY mid-execution when fresh drift-prevention emerges
user pushes back on a pattern · agent notices a recurring miss ·
cognitive context shifted enough to warrant a refresh
NEVER by a fresh agent that wasn't there for the drafting session
a renderer can fill scaffold sections from spec fields,
but Stance + Watch require having been in the conversation
🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪
Write to .roadmap/heads/r<N>.boot.md. One file per ROUND (not per DAG). N is the round number from dag-id prefix (r7-... → N=7), or from sidecar.round, or from dag_desc / Round. If none present, the spec lacks round encoding and the boot.md falls back to .roadmap/heads/<dag-id>.boot.md (legacy).
Round-scoped because cognitive context is round-scoped — multiple DAGs within a round share stance, drift-prevention, and substrate inheritance. Re-writing on each intra-round successor is by design: the latest boot.md folds prior round-stable residue with new shifts.
.roadmap/heads/r6.boot.md round 6 cartridge
.roadmap/heads/r7.boot.md round 7 cartridge (reads r6 for cross-round residue)
glob: .roadmap/heads/r*.boot.md all rounds, sorted lexicographically = chronologically
# <Plain-English DAG title from dag_desc line 1>
## Context
<1-2 sentences · what repo, what lane, what round>
## Intent
<pulled from dag_desc · what the human is actually asking for>
## Scenario
given <starting state>
when <the human acts>
then <the human can ___>
## Round
Round <N> · falsifier: <one line>
Carriers inherited from r<N-1>: <list, or "none">
## Substrate state at round boundary
<MANDATORY · see "Required boot.md section" below · jq-runnable facts, not prose>
## Start
**First action: invoke `/roadmap-auto`.** Do not orient separately — auto handles orientation internally. Do not ask the user "should I begin" — the user pasted this boot prompt because they want execution to start. (§No-spontaneous-checkpoints applies.)
Branch: <current branch>
First frontier (from compile-time orient): <ready node ids>
Read CLAUDE.md for project-specific execution protocol.
## Skills
/roadmap-orient · /roadmap-spec · /roadmap-bootprompt · /roadmap-auto
## Stance
<AUTHORED · session-tacit · the load-bearing part>
- <drift this session revealed · what to NOT re-explore>
- <dead ends · "we tried X, abandoned because Y">
- <patterns the user wants enforced · things to flag>
- <register/voice notes for this lane>
- <micro-doctrine that emerged in conversation but isn't in CLAUDE.md yet>
If the project has no tighter stance, the floor still travels:
- subtract before adding · removing a surface > handling a case
- extend, don't bolt · refactor when extension isn't natural
- thin and long > short and fat
- ~400 LOC goldilocks · functions 10-40 lines
## Watch
<AUTHORED · what to flag back to the user on encounter>
- <things the user wants surfaced if seen>
- <known AMBER candidates from prior rounds>
- <boundaries where surprise is likely>
## Trajectory so far
<optional · only if a prior session ran>
Last N nodes: <WIN WIN PARTIAL LOSS ...>
Notable pattern: <e.g. "LOSS streak at nodes 5-7 routed to upstream">
🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪
Every boot.md MUST include a top-level section titled ## Substrate state at round boundary between intent/thesis and stance. The section is mechanically
populated, not narrated · it inventories the substrate that exists at round
boundary so the next session opens with state, not prose.
Mandatory sub-sections (each one paragraph or short list):
· Canonical artifacts touched by this round · jq model/canon/index.json · cite path + sha256 + relevant entity counts. · Cross-machine substrate · jq model/canon/forensic-sources.json for any .sources[] whose role intersects the round's problem domain · cite .access.remote_host + .access.capture_emit_dir + .known_emitter_state flags. · Ephemeral substrate state · ls newest 1-3 epochs under model/raw// · cite actual VALUES sampled (not just field names) · note coverage percentages. · Predecessor round empirical findings · cite specific receipts (path + verdict + key numeric finding) · NOT prose summaries.
Anti-pattern · "prose narrating cognitive stance about substrate" instead of "jq-runnable substrate facts." Substrate facts survive context compression · prose summaries rot within one round..
§Substrate-inventory-precedes-DAG-authoring (sibling /roadmap-spec discipline)
1. SCAFFOLD pull from spec fields
· title from dag_desc line 1
· Intent/Scenario/Round from dag_desc structured blocks
· Branch from `git rev-parse --abbrev-ref HEAD`
· First frontier from `roadmap orient` JSON output
· Skills list is static
2. STANCE AUTHORED from session memory · the load-bearing work
walk back through the drafting conversation
surface:
· pushback the user gave on patterns
· approaches we tried and abandoned
· register/voice cues
· concerns the user voiced about quality/drift
· micro-doctrine not yet in CLAUDE.md
if the session didn't surface any of these, the floor
stance ships unchanged. that's honest; better than fake.
3. WATCH AUTHORED · what should the next agent flag back?
AMBER candidates · likely boundaries of surprise ·
things the user wants eyes on
4. WRITE one file, .roadmap/heads/r<N>.boot.md (round-scoped)
if r<N>.boot.md exists from a prior DAG in the round, FOLD
still-relevant residue into the new write · do not blindly
overwrite, do not blindly append
≤ 80 lines total · cartridge, not doctrine repo
MANDATORY the ## Start section MUST contain the directive
"First action: invoke `/roadmap-auto`." verbatim and prominent.
The boot prompt is a self-executing cartridge — pasting it
into a fresh session must trigger execution without the user
typing /roadmap-auto themselves. A boot.md missing this
directive failed authoring.
🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪
this skill requires session context · never dispatch to a background agent
the Stance + Watch sections are not derivable from spec fields
the scaffold is not the point — the cognitive residue is
brevity is mercy · a fresh agent reads this cold, not as a doctrine treatise
the floor stance always ships · the ceiling raises from there
✗ a renderer a renderer can fill scaffold sections but cannot
author Stance + Watch · those require having been there
✗ a session log not a transcript · not a "what we did" recap ·
capture FORWARD-LOOKING cognitive cues, not history
✗ doctrine permanent doctrine goes to CLAUDE.md ·
boot.md is per-ROUND and rotates on round boundary
✗ a retro not "what went well/poorly" · that's roadmap-auto's
terminal review · this is about the next session's mind
called immediately after /roadmap-spec runs `roadmap make`
also invokable standalone mid-session when drift emerges
output: .roadmap/heads/r<N>.boot.md at canonical path (round-scoped) ·
glob `.roadmap/heads/r*.boot.md` lists all rounds chronologically
read by: /roadmap-orient on session start
chain: spec → bootprompt → (user) → orient (next session) → auto
🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪🟥🟧🟨🟩🟦🟪
💀 The session that drafted the spec is the only one that can write its boot prompt.