Stop Burning Tokens
You've been paying Opus prices for Haiku work. Stop Burning Tokens fixes that.
Category: skill Author: npub1carj2jw…h8c9 Date: 5 May 2026 Votes: 1
---
name: stop-burning-tokens
description: >
Cost-optimised execution for multi-step planning pipelines. Produces an explicit,
dependency-sequenced plan before any work begins, segments steps into intelligence
tiers, prompts model switches at tier boundaries, and maintains a concise living
plan document throughout. Trigger on any request involving a multi-phase build,
migration, refactor, or investigation where planning and execution are distinct stages.
version: "1.0.0"
license: MIT
allowed-tools: Read, Write, Edit
metadata:
author: agentlist
tags:
- planning
- cost-optimisation
- orchestration
- pipeline
- agents
---
# SKILL: tiered-planning
## What this skill does
Structures multi-step pipelines to minimise cost without sacrificing quality. Before any
work begins, produces an explicit plan segmented by dependency order and intelligence
tier. Prompts the user to switch to the right model before each tier boundary. Maintains
a concise living `PLAN.md` throughout execution. Offers session compaction and new-session
handoffs at appropriate points.
---
## Phase 0 — Plan first
Before touching anything, produce a plan in this format and ask the user to confirm it.
Do not begin execution until confirmed.
```markdown
## Plan: <title>
| # | Step | Depends on (why) | Tier |
|---|------|-------------------|------|
| 1 | Define data schema | — | Architect |
| 2 | Scaffold directory structure | 1 (schema determines layout) | Execute |
| 3 | Implement parser | 2 | Execute |
| 4 | Review error handling strategy | 3 (parser output may reveal edge cases) | Architect |
```
Rules for writing the plan:
- Steps must be concrete enough that "done" is unambiguous
- Dependency column: list step numbers *and* a brief reason why that order matters
- Tier is one of **Architect** or **Execute** (see tier reference below)
- Group consecutive same-tier steps where possible — minimises model switches
- If the full scope is unclear upfront, plan as far as is reliable; flag remaining steps as `TBD` and revisit after step N
---
## Intelligence tier reference
| Tier | When to use | Anthropic | OpenAI |
|------|------------|-----------|--------|
| **Architect** | Design decisions, ambiguous requirements, debugging with unclear root cause, cross-cutting changes, anything where the wrong call is expensive to undo | Claude Opus 4 | o3 (high reasoning effort) |
| **Execute** | Well-defined implementation, boilerplate, mechanical refactors, formatting, steps where the approach is already decided | Claude Sonnet 4.5 / Haiku 4.5 | GPT-4.1 mini / nano (low–medium reasoning effort) |
When in doubt, escalate to Architect. The cost difference is small; the cost of a wrong Execute-tier decision is not.
> **Note:** Model switching mid-conversation is only possible in interfaces that expose model selection (e.g. Claude.ai, Claude Code with separate task invocations). If you are using a fixed-model API setup, apply Architect-level reasoning to all ambiguous steps regardless of tier labelling.
---
## Phase 1 — Execution
### At each tier boundary
Before beginning the first step of a new tier block, pause and output:
> **Model switch.** Steps N–M are [Tier] tier. Switch to [recommended model] before continuing. Reply when ready.
Issue this prompt **once per consecutive block**, not once per step. Do not proceed until the user confirms.
### During each step
1. State the step: **Step N of T — [title]**
2. Complete the work
3. Update `PLAN.md` immediately (see below)
4. If the step reveals something that changes a future step: update the plan table now, state the change explicitly, and determine whether the revision requires an Architect-tier re-evaluation before proceeding
---
## Plan document (`PLAN.md`)
Create this file before step 1 begins. Update it after every step — never batch updates.
```markdown
## Plan: <title>
_Progress: N/T steps complete_
### Steps
| # | Step | Status | Notes |
|---|------|--------|-------|
| 1 | Define data schema | ✅ done | |
| 2 | Scaffold directory structure | 🔄 in progress | |
| 3 | Implement parser | ⏳ pending | revised: must handle null ids (from step 1) |
| 4 | Review error handling strategy | ⏳ pending | |
### Learnings
- [Step 1] Schema has no unique constraint on `slug` — step 3 parser must deduplicate
```
Status values: `✅ done` · `🔄 in progress` · `⏳ pending` · `❌ blocked`
**Learnings rules:**
- Only log if the finding changes the approach of a future step
- One line per learning. No prose, no "as expected" entries
- If a learning makes a future step's plan incorrect, update that step's Notes immediately
---
## Session management
### Compacting
When either of these is true, offer to compact:
- `PLAN.md` exceeds ~80 lines, OR
- More than half the steps are `✅ done`
Offer:
> "PLAN.md is getting long. Want me to compact it? Completed steps collapse to a summary row; only learnings relevant to remaining steps are kept."
If the user agrees:
1. Collapse all `✅ done` rows into: `| 1–N | Steps 1–N | ✅ done | [summary of any persistent constraints] |`
2. Remove learnings that applied only to completed steps
3. Retain all learnings that affect remaining steps — do not summarise these further
4. Separately, if the model/interface supports `/compact` or equivalent context compression, offer to run that too — it is a distinct operation from compacting the document
### New session handoff
Once **at least 50% of steps are complete**, if the remaining steps are self-contained (i.e. `PLAN.md` plus any artefacts produced so far contain all necessary context), offer once:
> "Remaining steps look self-contained. A fresh session with PLAN.md as context may reduce cost and latency. Want to hand off?"
If the user agrees, output a handoff prompt they can paste into a new session:
```
Read PLAN.md. Continue from step N. All prior context is in the file and [list any other relevant artefacts].
```
---
## Rules
- **No execution without a confirmed plan.** If the user tries to skip confirmation, pause and re-ask.
- **PLAN.md is updated after every step, not at the end.** If the session dies mid-plan, the document must be recoverable.
- **Model switch prompts are mandatory at tier boundaries.** Never silently continue on the wrong tier.
- **Learnings section is for surprises and course-corrections only.** If it is not going to change a future step, do not write it.
- **Compact and new-session suggestions require explicit user confirmation.** Never compact or split unilaterally.
- **One new-session suggestion maximum.** Issue it once, after 50% completion. Drop it if the user declines.
Discussion