convert-prompt-to-saill
Convert a natural-language prompt into a SAILL agent-team flow — roles with model groups, control-flow flags (parallel/wait/Loop/if needed/if asked/ask user), boxes for sub-teams and nesting, and -context- for run-time values. Use when the user types /convert-prompt-to-saill, says "turn this into SAILL", "convert my prompt to SAILL", or "express this as an agent team / SAILL flow".
- License 许可
- MIT
- 最近更新
- 2026-07-06
- Stars
- ★ 1
- Forks
- 0
01 前言 Front Matter
- name
- convert-prompt-to-saill
- tools
- Read, Bash
- description
- Convert a natural-language prompt into a SAILL agent-team flow — roles with model groups, control-flow flags (parallel/wait/Loop/if needed/if asked/ask user), boxes for sub-teams and nesting, and -context- for run-time values. Use when the user types /convert-prompt-to-saill, says "turn this into SAILL", "convert my prompt to SAILL", or "express this as an agent team / SAILL flow".
03 SKILL.md 正文
Skill: /convert-prompt-to-saill
Model preference: #midcost (per horizon_aios_model_prefs.md; overridable by a prompt directive).
Turn "here's my prompt" into its SAILL form — the standard, shareable agent-team notation
(roles + model groups + flags + boxes + loops + -context-). The deliverable is the compact,
portable representation behind a plain-English request, which the user can then run, save as a
team, or share. Authoritative grammar: documentation/system/agent_teams.md (SAILL, §8) and the
flag catalog $HORIZON_ETC/agent_team_flags.md.
When to invoke
/convert-prompt-to-saill <prompt>, or the user asks to turn a prompt into SAILL / express
work as an agent team.
Step 1 — Load the vocabulary (do not invent flags)
1.1 Recognized flags: python $HORIZON_BIN/resolve_agent_teams.py --flags.
1.2 Existing teams (reuse one if it fits): python $HORIZON_BIN/resolve_agent_teams.py --json.
Use only flags in the registry and #groups defined in horizon_aios_model_prefs.md.
Step 2 — Analyze the prompt
Decompose the prompt into a flow:
- Roles — the distinct units of work; for each, a short charter (what it does).
- Order — sequential vs concurrent; mark concurrency with
parallel+ a followingwait. - Iteration — "until it passes / retry" → a
**Loop:**back to a named role, with a cap. - Gates — "ask me / approve first" →
ask user; "only if needed" →if needed; "only if I ask" →if asked. - Sub-agents / nesting — group related roles in a box
[ … ]; name it for an ephemeral sub-teamName[ … ]; nest as deep as the prompt implies. - Context values — anything left to runtime (a scope, a pass condition, a cap) →
-context-/-context-<name>.
Step 3 — Map to SAILL
3.1 If the prompt matches a shipped team (e.g. "investigate and fix X"), reuse it by name and add only the prompt's specifics (scope, flags) — do not reinvent.
3.2 Otherwise compose roles. Assign each a #group by the kind of work (mirror
horizon_aios_model_prefs.md Task-Class Routing): security / architecture / review → #highcap;
research / exploration → #investigate; mechanical / edits / summarizing → #lowcost;
debugging / coding → #debug; otherwise → #midcost.
3.3 Apply the flags, boxes, loops, and -context- from Step 2. Prefer named loop targets and
the most one-line-readable form that stays faithful.
Step 4 — Output
4.1 The SAILL block — either the agent_teams.md role format (### Name + numbered roles)
or a one-line box expression, whichever is clearer for this flow.
4.2 A short gloss: map each role / flag back to the phrase in the prompt it came from.
4.3 Call out any -context- placeholders and what fills them at run time.
Step 5 — Offer next steps
Offer to save it as a reusable team (/agent-teams), run it ("send this team …"), or
test it (/test-agent-teams <name>). Do NOT execute the work here — this skill only converts.
Notes for the executing agent
- Use ONLY flags from the registry and
#groupsfrom model-prefs — never invent vocabulary. If the prompt needs a new one, offer/agent-teams(add a flag) or/model-prefs(add a group). - SAILL governs control flow; a role's work is charter prose (§1.2) — keep "run skill X" /
"scope to Y" in the charter or as
-context-, never as a flag. - Keep the output terse and human-readable — the value of SAILL is a compact, shareable form.
- This converts only; it never runs the flow.
04 同仓库更多 Skills
try-my-prompt
Take a natural-language prompt and return two outputs — an agent_teams.md SAILL entry and a terse one-line invocation — by invoking /convert-prompt-to-saill and formatting the result.
test_saill_in_skill
Proof-of-concept SAILL-in-skill. Replaces a wordy dirtree-check skill with a tight SAILL block. Audits tested_implementations nonce coverage via the Nonce Trace Audit team.
test-agent-teams
End-to-end self-test of the Agent Teams system — walk every defined team (or one named team) and spawn each role so it echoes a self-generated nonce, its role, what it was told to do (its charter), and the model it actually ran as. Use when the user types /test-agent-teams --all-teams (all) or /test-agent-teams <team-name> (one team), says "test the agent teams", "run the agent-team self-test", or wants to verify team resolution + model routing. The skill generates its own nonce.
saill-to-english
Translate a SAILL agent-team definition into plain English — render every primitive, loop, box, flag, and -context- placeholder as natural language a non-technical reader can follow. Use when the user types /saill-to-english, says "translate this SAILL", "explain this team definition", "what does this loop do", or "describe this in plain English".
model-prefs
Configure or inspect the in-context model-preference layer — model groups (incl. local/BYO models), per-session slots, and task-class routing — by editing the gitignored extend file. Use when the user types /model-prefs, asks to "set up model groups", "define
model-prefs-test
Test which model the model-preference config actually resolves to per group, and (in live mode) spawn small test agents by group name to confirm the spawn honors it. Use when the user types /model-prefs-test, or says "test my model prefs", "check which model