spec-token-budget

Files (1)

• SKILL.md

SKILL.md (first lines)

# Spec token budget

## Profiles (pick one per turn)

| Profile | Max output | Input discipline |
|---------|------------|------------------|
| **advisory** | 800 words | Max 3 files read; explore subagent for search |
| **feasibility** | 400 words + table | Platform matrix only; no install unless asked |
| **implement** | HANDOFF + diff summary | Read approved specs only; no chat history |
| **handoff** | 500 words | Paths + blockers only; no narrative replay |
| **docs-touch** | 600 words | Whitelist paths only (README, docs/, ROADMAP) |
| **release** | 15 lines + YAML path | Run collect/estimate scripts; paste summaries not logs |

## Mandatory rules

1. **Verdict first** — 3 bullets before detail.
2. **No tool dumps** — >30 lines of output → summarize or write file and cite path.
3. **Delegation** — orchestrator ≤500 words, mostly paths (Principle 8). Use spawn allowlist; **never** attach conversation summaries.
4. **HANDOFF** — max 5 key decisions; next agent gets paths — **do not paste this HANDOFF block**.
5. **External fetch** — prefer `raw.githubusercontent.com`; skip HTML repo pages.
6. **Uploads** — if user attached a doc, do not re-fetch the same URL.

## Compression (Gate / subagent end)

If response or HANDOFF feels long:

- Drop repeated playbook explanations — cite `SPECFORGE_HOME/ENGINEERING-PLAYBOOK.md`.
- Replace paragraphs with a table.
- Set **For next agent — paths only** to a path list (never a chat summary).

Hook reminder: subagentStop may ask for compression before next delegation. Do not attach a chat summary to the next spawn.

## Orchestrator

State in every delegation: `Token profile: [name]`

From sushilti80/specforge-engineering-team, MIT.

More from sushilti80/specforge-engineering-team