track-roadmap
Use this skill to maintain ROADMAP.md whenever the user wants to "add an item to the roadmap", "put these in the roadmap", "update the roadmap", "mark a feature done", "log the work I shipped", "create a roadmap", "what should we build next", "brainstorm features", "audit the roadmap", "work through the roadmap", or file hands-on checks like "add the release sign-off to the plan" — even if they don't mention the roadmap file by name. Covers the generate, update, audit, brainstorm, and resume modes. Do NOT use this skill for session-level task progress (see track-session) or free-form idea backlogs that aren't landing in a roadmap file.
How do I install this agent skill?
npx skills add https://github.com/antjanus/skillbox --skill track-roadmapIs this agent skill safe to install?
- Gen Agent Trust Hubpass
The skill is generally safe and focuses on maintaining a project roadmap. It presents a low risk of indirect prompt injection as it scans project files (including TODO/FIXME comments and READMEs) to generate or audit the roadmap, which could allow malicious instructions in those files to influence the agent's behavior.
- Socketpass
No alerts
- Snykpass
Risk: LOW · No issues
- Runlayerpass
1 file scanned · No issues
What does this agent skill do?
Track Roadmap
Overview
Maintain ROADMAP.md in the project root as a living, high-level feature plan, so decisions about what to build next are intentional, not reactive. Keep it to features (user-visible capabilities), not tasks — task-level work belongs in track-session. A healthy roadmap is 5-15 committed features readable in under 2 minutes.
Workflow
Every mode reads the current state and writes the file. Whether it proposes and waits in between depends on where the change came from. A change the user stated ("add X", "dark mode shipped yesterday", "drop the export item") is written directly and reported in one line. A change that is your own judgment (a generated feature list, an audit ruling, a reword or removal nobody asked for) is proposed first and written after the user confirms. Per mode:
| Mode | Command | Propose / confirm |
|---|---|---|
| Generate | /track-roadmap or generate | Offer a codebase scan, ask 4 discovery questions, propose a feature list, write after the user confirms |
| Update | /track-roadmap update | Apply what the user stated; ask what changed only when the request does not say |
| Audit | /track-roadmap audit | Scan code to mark each feature Done/In-Progress/Not-Started/Unclear, review relevance with the user, update |
| Brainstorm | /track-roadmap brainstorm | Divergent ideation — explore directions before committing; viable ideas land in "Future Ideas" as status:idea |
| Resume | /track-roadmap resume | Check session state, present remaining features, user picks one, hand off to /track-session |
Update is the common case: adding an item, marking one done, or logging shipped work all land there. There is no save mode — treat /track-roadmap save as Update. When the user asks to work through the whole roadmap, Resume takes the open features in file order, one session per feature, and moves to the next without asking once each session closes. In Update and Audit, edit the affected items in place rather than regenerating the file — ids, ordering, and untouched text survive, and a whole-file rewrite costs more tokens for the same result. Generate is the only mode that writes the file from scratch.
Full per-mode procedures (discovery questions, brainstorm question banks, audit steps): reference/MODES.md.
ROADMAP.md format
---
schema: cc-dash/roadmap@1
project: project-name-here
description: One-line project purpose
last_updated: YYYY-MM-DDTHH:MM:SS-TZ
---
# Roadmap
> Project purpose in one sentence.
## Core Features
<!-- category:core -->
- <!-- id:r_XXXXX status:planned --> **Feature Name** - What it does and why it matters.
> Research finding, with its source link.
> Open: the question still to settle.
- <!-- id:r_XXXXX status:in-progress started:YYYY-MM-DD --> **Feature Name** - Short description.
## Future Ideas
<!-- category:future -->
- <!-- id:r_XXXXX status:idea --> **Feature Name** - Short description.
## Completed
<!-- category:completed -->
- <!-- id:r_XXXXX status:done completed:YYYY-MM-DD --> ~~**Feature Name**~~ - Short description. *(Completed: YYYY-MM-DD)*
Format rules. The cc-dash dashboard parses this file, so a dropped marker silently removes the item from the board:
- Frontmatter requires
schema,project,description,last_updated. - Every item: an
idand astatus. Ids arer_+ 2–10[a-z0-9_]chars and permanent — never change; new ones default to 5 random chars. Grep the file for a new id before writing it, because a duplicate makes two items indistinguishable on the board. - Statuses:
idea/planned/in-progress/done, plusdeferred(still wanted, not now) anddropped/skipped/rejected/cancelledwhen the user wants a record kept of why an item stopped instead of deleting it. The last four land in the board's Closed column. - Optional metadata:
started:andcompleted:asYYYY-MM-DD, anddepends:r_abc12,r_def34. cc-dash ignores any other date key (at:,done_date:) and deletes it on its next write, so convert those tocompleted:when you touch the item. - Every category heading gets
<!-- category:slug -->on the next line. Categories themselves are flexible — invent whatever groups fit the project. - One line per feature: bold title + 1-2 sentence description.
- Notes hold what the description can't: research findings with their source, open questions, and rejected alternatives with the reason. Write them as
>lines directly under the item; cc-dash shows them on the card and keeps them through its own writes. Keep an item to about five note lines — when research outgrows that, move it to a file underdocs/and leave one note linking to it, because a roadmap that carries its research inline stops being readable in two minutes. Don't put notes in the description or in a second paragraph under the item; the dashboard reads neither. - When an item moves to Completed, cut its notes to what a later audit needs (the decision and why), since the research has done its job.
Rules
- The user drives the roadmap — write what the user asked for without a second confirmation, and propose and wait only for additions, rewords or removals that are your own judgment.
- High-level only — "User authentication", not "Add bcrypt to hash passwords".
ROADMAP.mdis the source of truth — if it's not in the file, it's not on the roadmap.- Completed features move, they don't disappear — relocate them to the Completed section with a
completed:date instead of deleting, because audits use that history as evidence.
Examples
❌ Bad — - tasks / - lists / - make it look good / - fix bugs: no descriptions, no groupings, mixes features with tasks, no IDs or purpose. The format block above is the ✅ counterpart.
✅ Good — update mode, user says "dark mode shipped yesterday": keep the id, flip the status, add the date, move the line to Completed:
- <!-- id:r_j4t8v status:done completed:2026-03-14 --> ~~**Dark mode**~~ - System-level dark/light theme preference. *(Completed: 2026-03-14)*
Full ✅/❌ comparisons for generate, audit, update, brainstorm, and resume modes: reference/EXAMPLES.md.
Integration
- track-session — after picking a feature, use
track-sessionto plan and track the implementation; SESSION_PROGRESS.md references the ROADMAP item ID. - Hands-on verification (a playthrough, a parity gate, a release sign-off) is an ordinary roadmap item here — there is no separate checklist file, so nothing files back from one.
Gotchas
- Renaming a feature keeps its id. Edit the title text but leave
<!-- id:r_XXXXX -->alone — SESSION_PROGRESS.mdroadmap_refvalues and cc-dash history both point at it. On a merge, keep the surviving id; on a split, keep the original on one half and generate a new id for the other. roadmap_refholds exactly one id. cc-dash drops a comma list with a warning, which unlinks the session from the board. When one session advances several features, link the main one and name the rest in the session's## Current Status.- Resume overwrites an active session plan. If SESSION_PROGRESS.md has uncompleted tasks, ask whether to continue it before presenting the roadmap; switching features replaces that file, so offer
/track-session savefirst. - Ambiguous audit evidence is
Unclear, not a guess. Partial or experimental code (asrc/auth/with no tests) gets surfaced to the user for a ruling rather than being scored Done. - A 30+ item roadmap is a signal, not a state to preserve. Audit it, move speculative entries to "Future Ideas" as
status:idea, and delete what no longer fits — git holds the history, so in-file archiving just adds noise. - Roadmap drift is expected. When the file stops matching what's being built, run audit to reconcile rather than patching individual lines from memory.
Extended edge cases (codebase-scan noise, brainstorm dead ends, priority paralysis, tightening vague feature descriptions): reference/TROUBLESHOOTING.md.
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/antjanus/skillbox/track-roadmap">View track-roadmap on skillZs</a>