Ask Dev Loop
The human does not remember every skill. Ask.
This is an advisory router: name the skill to type next and why, then stop.
Execution routers can invoke reusable engines within an authorized task, but
this skill only recommends. grilling, domain-modeling, tdd, and debug
may be named as what the chosen workflow will run.
Contract
Inputs:
- A situation in plain language (an idea, a bug, a PR, a foggy large effort, a
setup question, or "I don't know where to start")
Outputs:
- The recommended skill to type, one sentence of why, and the next step after that
- Neighbours when two skills are easy to confuse
Creates/Modifies:
External Side Effects:
Confirmation Required:
Delegates To:
- None. Name the skill; the human types it.
Precondition
If docs/agents/issue-tracker.md is missing, recommend /setup-agent-routing
first. The other engineering skills read that routing block.
The main flow: idea → ship
The route most work travels.
/interview — sharpen the idea. Repo-grounded; runs grilling and
domain-modeling; leaves an interview brief. Start here whenever the working
directory is a real repo.- Branch — does a design question need a runnable answer? Detour through
/prototype (throwaway code that answers one question), then return to the
brief.
- Branch — is this a multi-session build?
- Yes →
/prd write (prd-writer) then /prd intake (feature-intake)
or writing-plans on the issue, then /loop / executing-plans per ticket.
- No →
writing-plans in this session, then executing-plans (or just
implement with /tdd).
Keep grilling, spec, and tickets in one context window. Each /loop /
executing-plans run starts fresh from the ticket.
On-ramps
- Bugs and incoming requests piling up →
/prd intake (feature-intake) or
github-inbox. Tickets that prd-task-creator already wrote are agent-ready; do not
re-intake them.
- Something's broken →
/debug (or systematic-debugging when previous fixes
failed). Tight red loop first; no theory without a loop.
- A huge, foggy effort →
roadmap-analyzer / roadmap-to-milestones to chart
the destination, then merge onto the main flow at /interview or /prd write.
Do not skip the collapse into a buildable PRD.
Codebase health
Not feature work — upkeep.
/codebase-advisor — survey, produce plans for another agent. Read-only on
source./tech-debt — ranked debt register (interest over principal).codebase-design — deep-module vocabulary when the question is the shape
of a module, not an inventory.
Review
/review (review-dispatch) — pick the review depth and target.code-review — correctness and security gate, plus spec fidelity against
the originating issue.
Standalone
/wait-what — the last message did not land; re-pitch it.grilling — the interview primitive with no wrapper. Reach for it only when
the interview itself is the whole ask.domain-modeling — the words are the problem (fuzzy term, overloaded
"account", missing ADR)./wizard — steps only a human can perform (dashboards, secrets, cutovers).fix-merge-conflicts — already mid-merge or rebase.
How to answer
Match the user's situation to one row above. Reply with:
- The skill to type (slash name when it has a command).
- Why this one, in one sentence.
- The neighbour they might have meant, if any.
- What happens after that skill finishes.
If two flows both fit, ask one question that splits them, then recommend.
1---
2name: ask-dev-loop
3description: Ask which Dev Loop skill or flow fits the current situation. A router over the flagship idea-to-ship path.
4disable-model-invocation: true
5license: MIT
6metadata:
7 version: "1.0.2"
8 tags: "dev-loop, router, planning, dispatch"
9 author: Ship Shit Dev
10when_to_use: "which skill, what should I run, ask-dev-loop, how do I start, idea to ship, which flow"
11---
12
13# Ask Dev Loop
14
15The human does not remember every skill. Ask.
16
17This is an **advisory router**: name the skill to type next and why, then stop.
18Execution routers can invoke reusable engines within an authorized task, but
19this skill only recommends. `grilling`, `domain-modeling`, `tdd`, and `debug`
20may be named as what the chosen workflow will run.
21
22## Contract
23
24Inputs:
25
26- A situation in plain language (an idea, a bug, a PR, a foggy large effort, a
27 setup question, or "I don't know where to start")
28
29Outputs:
30
31- The recommended skill to type, one sentence of why, and the next step after that
32- Neighbours when two skills are easy to confuse
33
34Creates/Modifies:
35
36- None
37
38External Side Effects:
39
40- None
41
42Confirmation Required:
43
44- None. Advisory only.
45
46Delegates To:
47
48- None. Name the skill; the human types it.
49
50## Precondition
51
52If `docs/agents/issue-tracker.md` is missing, recommend `/setup-agent-routing`
53first. The other engineering skills read that routing block.
54
55## The main flow: idea → ship
56
57The route most work travels.
58
591. **`/interview`** — sharpen the idea. Repo-grounded; runs `grilling` and
60 `domain-modeling`; leaves an interview brief. Start here whenever the working
61 directory is a real repo.
622. **Branch — does a design question need a runnable answer?** Detour through
63 `/prototype` (throwaway code that answers one question), then return to the
64 brief.
653. **Branch — is this a multi-session build?**
66 - **Yes** → `/prd write` (`prd-writer`) then `/prd intake` (`feature-intake`)
67 or `writing-plans` on the issue, then `/loop` / `executing-plans` per ticket.
68 - **No** → `writing-plans` in this session, then `executing-plans` (or just
69 implement with `/tdd`).
70
71Keep grilling, spec, and tickets in **one context window**. Each `/loop` /
72`executing-plans` run starts fresh from the ticket.
73
74## On-ramps
75
76- **Bugs and incoming requests piling up** → `/prd intake` (`feature-intake`) or
77 `github-inbox`. Tickets that `prd-task-creator` already wrote are agent-ready; do not
78 re-intake them.
79- **Something's broken** → `/debug` (or `systematic-debugging` when previous fixes
80 failed). Tight red loop first; no theory without a loop.
81- **A huge, foggy effort** → `roadmap-analyzer` / `roadmap-to-milestones` to chart
82 the destination, then merge onto the main flow at `/interview` or `/prd write`.
83 Do not skip the collapse into a buildable PRD.
84
85## Codebase health
86
87Not feature work — upkeep.
88
89- **`/codebase-advisor`** — survey, produce plans for another agent. Read-only on
90 source.
91- **`/tech-debt`** — ranked debt register (interest over principal).
92- **`codebase-design`** — deep-module vocabulary when the question is the *shape*
93 of a module, not an inventory.
94
95## Review
96
97- **`/review`** (`review-dispatch`) — pick the review depth and target.
98- **`code-review`** — correctness and security gate, plus spec fidelity against
99 the originating issue.
100
101## Standalone
102
103- **`/wait-what`** — the last message did not land; re-pitch it.
104- **`grilling`** — the interview primitive with no wrapper. Reach for it only when
105 the interview itself is the whole ask.
106- **`domain-modeling`** — the words are the problem (fuzzy term, overloaded
107 "account", missing ADR).
108- **`/wizard`** — steps only a human can perform (dashboards, secrets, cutovers).
109- **`fix-merge-conflicts`** — already mid-merge or rebase.
110
111## How to answer
112
113Match the user's situation to one row above. Reply with:
114
1151. The skill to type (slash name when it has a command).
1162. Why this one, in one sentence.
1173. The neighbour they might have meant, if any.
1184. What happens after that skill finishes.
119
120If two flows both fit, ask one question that splits them, then recommend.
shipshitdev