slice-authoring

name	slice-authoring
description	Slice Authoring Skill

Slice Authoring Skill

Tags

slice, delegation, context, prompt, handoff, acceptance-criteria, scope, sub-agent, implementer

When to Use

Writing a delegation prompt for Gemini/Codex/Local; structuring a sub-agent task; defining acceptance criteria; determining what context to include vs exclude; assigning a role to a delegatee.

---

1. Minimal Viable Slice Context

A well-formed slice prompt contains exactly this — no more, no less:

slice_objective: <1-2 sentences — what to build, not how>
role: implementer  # or reviewer, architect
phase: Phase 86

relevant_files:
  - scripts/ai/lib/attention_queue.py   # paths only — agent reads them
  - dashboard/backend/api/routes/aistack.py

acceptance_criteria:
  - GET /api/aistack/alerts/status returns 200 with {"queue_depth": N, "pending": [...]}
  - aq-qa 0 passes (77/77) after change
  - No new AppArmor denials in journalctl -u apparmor.service

constraints:
  - Port from options.nix only — never hardcode 8889
  - NoNewPrivileges=true on this service — use ix not Ux for AppArmor subprocess rules
  - Async handlers only — no blocking I/O in FastAPI routes

reference_skills:
  - testing-patterns   # for http_get tuple, QA check authoring
  - coordinator-api    # if touching coordinator routes

resume_checkpoint: .agent/collaboration/RESUME.json

---

2. What to Include vs Exclude

Include	Exclude
File paths (not content)	Full HANDOFF.md (link to it)
Acceptance criteria as explicit conditions	Previous agent outputs
1-2 constraints most likely to bite	Full AGENTS.md
Skill names (not skill content)	Unrelated file contents
RESUME.json path for recovery	Full session history
Known gotchas specific to this slice	General system orientation

Why: A sub-agent receiving 6000 tokens of context will spend its budget parsing noise. A sub-agent receiving 500 tokens of signal will deliver better output.

---

3. Acceptance Criteria Format

Good acceptance criteria are binary — pass or fail, no ambiguity:

# GOOD:
- GET /api/route returns 200 with field "status" == "ok"
- aq-qa 0 passes (77/77 checks)
- File X contains pattern Y (verifiable with grep_search)
- No Python syntax errors (py_compile passes)

# BAD:
- "The code should work correctly"
- "The dashboard looks good"
- "Performance is acceptable"
- "Handle edge cases"

For Gemini (no shell): use file-verifiable criteria only:

- acceptance_criteria:
  - The file scripts/ai/aq-alerts contains a function named format_alert_line
  - The file imports attention_queue from scripts/ai/lib/attention_queue.py
  - No obvious syntax errors visible in the Python code

---

4. Role Assignment in Delegation Prompt

# In delegate-to-gemini:
scripts/ai/delegate-to-gemini --role implementer --prompt-file /tmp/slice.txt

# In delegate-to-local:
scripts/ai/delegate-to-local --mode agent --role implementer --prompt-file /tmp/slice.txt

# In Codex:
scripts/ai/delegate-to-codex --prompt-file /tmp/slice.txt
# (include "Your role: implementer" in the prompt text)

Role assignment in the prompt also sets the context tone:

• implementer: tell it exactly what to build, where, and what done looks like
• reviewer: give it the implementation + acceptance criteria; ask for explicit verdict
• architect: give it the problem space; ask for design + risk analysis

---

5. Slice Boundaries (Scope Control)

Before writing the slice prompt, declare the scope boundary explicitly:

scope_boundary:
  in_scope:
    - dashboard/backend/api/routes/aistack.py  (add 1 route)
    - scripts/testing/harness_qa/phases/phase86.py  (add check)
  out_of_scope:
    - Any changes to ai-stack/ services
    - NixOS module changes
    - Other dashboard routes

This prevents the implementer from "helpful" scope creep that breaks other things.

---

6. Token Budget by Agent

Agent	Prompt budget	File reading budget	Output budget
Claude	~4000	Unlimited	Unlimited
Gemini	<2000 (routing 429 above this)	auto_edit reads on demand	~1400
Codex	Any (use --prompt-file)	reads on demand	Any
Local (direct)	<512 for best results	None — direct mode has no tool access	180 (coordinator ceiling)
Local (agent)	<3500	reads on demand	512

For Gemini: if the slice prompt + context exceeds ~1500 tokens, split into 2 slices.

---

7. VERDICT Request (for Reviewer Slices)

When assigning reviewer role, always request an explicit verdict at the end:

After reviewing, output your final verdict as the LAST LINE in this exact format:
VERDICT: PASS|FAIL|REQUEST_REVISION — <one-line reason>

If REQUEST_REVISION: list specific changes needed with file:line references.

The orchestrator reads the last line to determine whether to accept or re-delegate.