Coding agents
Call a skill to research a change, plan the work, or edit a hook.
Skills do not start by themselves. Put the slash name in the prompt. If the slash name does not start the skill, put the skill file on line 1.
Rules live in .agents/rules/. Load each rule by exact path from catalog.md. Do not glob that folder.
Outcome
You know which skill to call, what that skill writes, and when the pipeline stops.
Assumptions
- The workspace is a Logic Bee API repo that contains
.agents/skills/and.agents/rules/. - MCP tools are available for jobs that need them.
- This site does not host the skill files.
Pipeline
logic-discovery → logic-plan → logic-plan-critic
↓ (when the change is a hook)
hook-plan → hook-plan-critic
↓
hook-coder
↓
hook-critic
Work one stage at a time. Do not mix two skills in one message.
What each stage writes
| Stage | Call | Writes |
|---|---|---|
| Discover | /logic-discovery or @.agents/skills/logic-discovery/SKILL.md | optional *-discovery.md / *-requirments.md |
| Plan the workspace | /logic-plan or @.agents/skills/logic-plan/SKILL.md | {slug}-plan.md |
| Criticize the plan | /logic-plan-critic or @.agents/skills/logic-plan-critic/SKILL.md | — |
| Plan one hook | /hook-plan or @.agents/skills/hook-plan/SKILL.md | {collection}-{slug}-hook-plan.md (not .ts) |
| Criticize the hook-plan | /hook-plan-critic or @.agents/skills/hook-plan-critic/SKILL.md | — |
| Implement | @.agents/rules/hook/coder.md | *.hook.ts |
| Criticize the hook | /hook-critic or @.agents/skills/hook-critic/SKILL.md | — |
hook-coder is a rule, not a skill. Call it with the rule path. The plan file uses the format in .agents/rules/logic/plan-format.md.
Critics and MCP
Critics return PASS, REVISE, or REJECT. Critics do not write the next artifact.
MCP is mandatory for jobs that need tools. If the tools are missing, the skill stops.
logic-plan and hook-plan start with a local intake. They do not call MCP until the job is impact, patch, spawn, or retire.
explain and every named handoff never call MCP.
Every skill loads .agents/rules/logic/thinking.md (legacy name logic-thinking.md). Obey Scope in that file. Do not recopy the five rules into a prompt.
The implementer rule .agents/rules/hook/coder.md does not load thinking.md.
Pick the altitude
| Reader question | Call |
|---|---|
| How does this work? What does this touch? | /logic-discovery |
| What do we build, and in what order? | /logic-plan |
| Is this plan honest? | /logic-plan-critic |
| Design this one hook. Do not write TypeScript. | /hook-plan |
| Write the hook. | @.agents/rules/hook/coder.md |
Review this hook .ts. | /hook-critic |
If the work spans documents and hooks, use /logic-plan. If the work is one hook, use /hook-plan.
Type sets
The two type sets are different.
discovery: bug | feature | idea | requirement | question
plan: bug | feature | fix | user-request | research
Name type in the first /logic-discovery prompt, or omit type so the skill asks.
Do not name a plan type in the first /logic-plan message. Do not name a job there. The skill asks after intake. Answer in the question box.
Gradual order
Skip a missing step. Never reorder. The order and blast radius colors are on Plan the workspace.
Agent builder
The surgery skill is not in the hook pipeline. It authors and cuts skills and rules only.
@.agents/skills/surgery/SKILL.md
type: build | cut | split | consolidate | audit
target: skill | rule
<path>
Lexicon: .agents/skills/surgery/lexicon.md.