The Instruction Diet Kit (free)
The complete starter kit for this week’s play: build a local scanner that finds the instruction debt inside AI development agencies' repositories, and sell it at $249 a quarter per workspace. Free for one issue cycle, then it moves into the maintained Vault.
This kit assumes nothing but a laptop, a coding agent (Claude Code, Codex, or Cursor), and the willingness to send 20 emails. No prior product, no portfolio, no code you have to write by hand.
What you build: a single-file scanner that inventories a workspace’s agent instruction files (AGENTS.md, CLAUDE.md, Cursor rules, Markdown skill files), runs 18 deterministic checks, and outputs a review queue plus a proposed cleanup patch that a human approves. It never edits anything on its own.
What you sell: the scan, the patch proposal, and a quarterly re-scan, at $249 per quarter per workspace.
1. THE WORKED EXAMPLE (read this first)
Northstar Dev is a hypothetical 14-repository AI development agency. Its developers have written agent instructions for two years, across three model generations and four staff members. Nobody has ever deleted anything.
You scan one client workspace. The inventory counts 11,600 always-loaded tokens across AGENTS.md, CLAUDE.md, and three skill files. The checks find:
# | Finding | Check | Proposed verdict |
|---|---|---|---|
1 | “Explain the plan before editing” appears 5 times across 3 files | Near-duplicate rule | MERGE into one scoped rule |
2 |
| Contradicting imperatives | INVESTIGATE, assign an owner |
3 | A 1,900-token migration walkthrough loads on every task | Oversized global example | MOVE TO ON-DEMAND |
4 | 7 rules name Opus 4.8; the workspace moved to Opus 5 | Stale model name | UPDATE |
5 | 2 rules point at | Dead file reference | UPDATE or KEEP with fixed path |
The proposed patch cuts always-loaded context from 11,600 tokens to 4,200, about a 64% reduction, every change traceable to a file and line.
The claim discipline (this is what makes you credible): the scanner proves structure, not performance. You may say “your always-loaded context shrinks 64% and the contradiction in your migration rules is resolved.” You may NOT say “your agents will be 64% better.” The buyer’s own tests decide whether the patch merges. Sell the visibility and the traceability. The honesty is the moat.
Conservative math on the business: 50 researched contacts take about 5 hours. A modest result is 2 paying workspaces. After roughly $9 payment processing and $10 delivery infrastructure per workspace, you net $230 per sale, about $70 per hour including outreach and onboarding. The renewal quarter runs about half an hour of support per workspace. Round down, then let renewals surprise you.
2. THE FIXTURE REPOSITORY (your demo, built in 10 minutes)
Never ask a prospect for production code before they pay. Demo on a fixture: a fake, deliberately messy workspace that mirrors Northstar.
Create a folder fixture-workspace/ with three files:
AGENTS.md (the global file). Include, on purpose:
- the rule “Explain the plan before making any edits.” (this will be duplicated elsewhere)
- “Never create a database migration.”
- a 1,900-token worked example of a database migration, pasted inline
- two rules that name “Opus 4.8” explicitly
- a rule referencing docs/style-guide.md (do not create that file)
CLAUDE.md. Include:
- “Before editing, explain your plan.” (near-duplicate of the AGENTS.md rule)
- “Always explain the plan first, then edit.” (a third variant)
- three rules naming “Opus 4.8”
- a note with no imperative (“The team prefers clean code.”)
skills/migrations.md. Include:
- “Always create a migration when the schema changes.” (the contradiction)
- another copy of “Explain the plan before editing.”
- a reference to docs/style-guide.md
That single folder now contains every one of the five Northstar findings. The scan of this fixture IS your sales demo: record 90 seconds of the scanner finding all five.
3. THE BUILD PROMPT (a coding agent writes v1 in one session)
Paste this into Claude Code, Codex, or Cursor, from an empty project folder. Budget 12-16 focused hours across build, fixture, and demo before your first outreach; the first scanner session is usually an afternoon.
Build a single-file Python 3 script called instruction_diet.py. No external
dependencies beyond the standard library. It is a READ-ONLY scanner: it must
never modify, move, or delete any file it scans.
INPUT: a directory path. Recursively find agent instruction files:
AGENTS.md, CLAUDE.md, any *.md under directories named skills/ or rules/,
and files matching .cursor/rules/* or *.mdc.
PASS 1, INVENTORY: for each file, split into sections by markdown heading
and into individual rules (a rule = a paragraph or list item containing an
imperative). For each rule record: file path, section heading, start line,
end line, load scope (global if the file is AGENTS.md/CLAUDE.md, otherwise
on-demand), rule type (imperative / example / note), any model name matched
by the regex (opus|sonnet|haiku|fable|gpt-[0-9.]+|o[0-9]|gemini|claude)
[-\s]?[0-9.]*, any file path referenced in backticks or links, and an
approximate token count (characters / 4).
PASS 2, CHECKS: run the 18 checks provided in the checklist file
(checks.md) and emit one finding per hit with: check id, severity
(high/medium/low), file, line range, the offending text (first 120 chars),
and a one-line reason.
OUTPUT 1: inventory.csv with the 9 inventory columns.
OUTPUT 2: findings.md, a review queue grouped by check, each finding with
a suggested verdict from exactly this set: KEEP / MERGE / MOVE TO
ON-DEMAND / UPDATE / INVESTIGATE, and an empty "approved:" field.
OUTPUT 3: patch-preview.md, a unified-diff-style proposal covering ONLY
findings whose approved field is set to yes on a re-run. On first run it
contains only the header "No approved findings yet."
Also print a summary: files scanned, rules counted, total always-loaded
tokens, findings by severity.
Include a --fixture flag that explains what the fixture workspace is.
Write clean, commented code. Then run it against ./fixture-workspace and
show me the summary.Then iterate twice: once to fix whatever the fixture scan gets wrong, once to make the findings.md readable enough to screenshot. Stop there. Do not add platforms, a web UI, or auto-fixing before someone pays.
4. THE 18 DETERMINISTIC CHECKS (checks.md, ships with the scanner)
Duplication:
1. Exact duplicate rule text appearing in more than one file. (high)
2. Near-duplicate rules: same imperative verb + object, different wording. (medium)
3. The same rule repeated inside one file. (medium)
Contradiction:
4. Contradicting imperatives: a “never X” and an “always X” on the same object anywhere in the workspace. (high)
5. Two rules claiming the same trigger with different required outcomes. (high)
6. Conflicting output-format demands (e.g. one rule requires JSON, another prose, same scope). (medium)
Staleness:
7. A rule naming a model that no other current config in the workspace names. (high)
8. A rule referencing a tool, command, or flag that appears nowhere else in the repo. (medium)
9. A referenced file path that does not exist. (high)
10. A rule with an expired date qualifier (“until”, “before”, a past month). (medium)
Scope and load:
11. An example block over 1,000 tokens in an always-loaded file. (high)
12. A rule with no trigger or scope in an always-loaded file. (low)
13. Task-specific keywords (migration, deploy, release, billing) in a global file rather than a skill. (medium)
14. An always-loaded file whose total exceeds 4,000 tokens. (high)
Hygiene:
15. A “rule” with no imperative verb: a preference note doing nothing. (low)
16. Commented-out or struck-through blocks still present in a loaded file. (low)
17. Formatting artifacts that break parsing: unclosed code fences, curly quotes inside command examples. (medium)
18. An orphaned skill file referenced by no config and no other file. (medium)
These thresholds (1,000 and 4,000 tokens) are defensible defaults, not laws. Let a buyer change them; record the change in the report header.
5. THE INSTRUCTION INVENTORY (the 9 columns)
Every scan produces inventory.csv with exactly these columns. This sheet is half the deliverable; agencies have never seen their instruction surface as a table.
file(path)section(nearest heading)lines(start-end)load_scope(global / on-demand)rule_type(imperative / example / note)named_model(or blank)referenced_path(or blank)approx_tokensfinding_ids(comma-joined, or blank)
6. THE CONFLICT MATRIX + REVIEW QUEUE
Conflict matrix (one row per check-4/5/6 finding): rule A file and lines, rule B file and lines, conflict type, the two quoted lines, owner (a human at the agency), status.
Review queue (findings.md): every finding gets exactly one verdict from the five labels:
- KEEP: the rule earns its place; record why.
- MERGE: duplicates collapse into one scoped rule; the patch shows the survivor.
- MOVE TO ON-DEMAND: examples and task rules leave the global files for a skill loaded when relevant.
- UPDATE: stale model names, fixable paths, expired dates.
- INVESTIGATE: contradictions and anything needing a human decision; never patch these silently.
The patch rules (non-negotiable, print them in every report): the scanner proposes, the human approves, the agency applies through its normal code review. No auto-apply. No claim about performance. Every change cites its source lines. INVESTIGATE items never appear in a patch.
7. THE BUYER QUALIFICATION SHEET
Score a prospect before you pitch. A YES needs all four:
Question | Pass |
|---|---|
Do they build with coding agents for clients (site, cases, or job posts say so)? | Yes required |
Plausibly 10+ active repositories (team of 2-20, multiple case studies)? | Yes required |
Multiple instruction surfaces (they mention Claude Code, Cursor, or agents in how they work)? | Yes required |
A named human you can email from public info? | Yes required |
Agencies with one tidy CLAUDE.md are a NO. Solo freelancers are a NO (they are your competitors, not your buyers). Enterprises are a NO for now (procurement will out-wait you).
Where to find 50: Clutch’s AI development and generative AI directories list agencies with team size, location, and website. Five hours of qualification gets you 50 rows: agency, site, contact name, contact path, qualification score, date sent, reply.
8. THE OUTREACH NOTE (20 sends, then stop and read replies)
Subject: two rule files that contradict each other
Body:
Quick question for whoever owns your agent setup: who removes agent instructions after a model upgrade, and how do you catch two rule files that contradict each other? I ask because Anthropic deleted more than 80% of Claude Code’s own system prompt at the Opus 5 launch with no measurable eval loss. Most agencies' AGENTS.md and skill files have been accumulating since 2024 and nobody owns the deletion. I run a local scanner that inventories your instruction files and flags duplicates, contradictions, and stale rules with exact line references, then proposes a cleanup patch your team reviews. It runs on your machine; your code never leaves the building. 90-second demo on a fixture workspace: [link]. Founding price is $249 a quarter per workspace, first scan included. If your instruction files are already clean, you don’t need me, and the demo will show you that too.
Rules: send from a real business email, name yourself honestly, include a working unsubscribe path or an explicit “tell me if you’d rather not hear from me,” and honor every opt-out immediately (US commercial-email law, CAN-SPAM, applies to cold B2B email). No follow-up spam: one send, one bump a week later, done.
9. THE PRICING SCRIPT
Founding: $249 per quarter per workspace. First scan, review-queue walkthrough, one patch proposal, and the quarterly re-scan with updated checks.
Hold the price. If they ask for a discount, offer to scan ONE workspace instead of three, never a lower rate.
When to say no: fewer than 10 repositories, or they want auto-apply (“the scanner will never edit your files; that is a feature”), or they want a performance guarantee (the claim discipline above is the answer).
After two renewals or five workspaces, the standard rate for NEW buyers moves to $349 a quarter. Founding buyers keep $249. Tell them that at the founding pitch; it is true and it closes.
10. THE INSTALL SOP (day by day)
Day 1 (build, ~4-6 hrs): run the build prompt, create the fixture workspace, iterate until the scan finds all five Northstar-class findings.
Day 2 (demo + list, ~4 hrs): record the 90-second fixture demo. Put the $249 founding offer on a one-page site or a simple doc. Start the Clutch qualification sheet.
Day 3 (finish the 50, ~3 hrs): complete 50 qualified rows.
Day 4 (send 20): first 20 outreach notes, logged.
Day 5-7: answer replies with the demo, book scans. Run each first scan live on a call if the buyer prefers; the inventory lands either way.
Quarter+1: re-run every paying workspace with the updated checks file; deliver the delta report. This half hour is what the renewal buys.
Works-every-time checklist (before any paid scan):
- [ ] Scanner runs clean on the fixture and finds all five seeded findings
- [ ] Report prints the patch rules and the claim discipline verbatim
- [ ] inventory.csv opens correctly in a spreadsheet
- [ ] You have said, in writing, that the scanner is read-only and runs locally
- [ ] The buyer has named the human who owns verdicts
Legal and safety notes: cold email must comply with CAN-SPAM (identify yourself, no deceptive subject, honor opt-outs). The scanner must stay read-only; you are never taking custody of client code, and say so in writing. No performance claims without the buyer’s own test results. This kit is educational, not legal advice.
THINKING THIS IS A LOT OF WORK?
It is. Read back what you just agreed to do.
Build the scanner. Write the eighteen checks. Assemble a fixture workspace so you can prove the checks actually fire. Build the inventory sheet, the conflict matrix, the review queue. Write the qualification sheet, the outreach note, the pricing script. Then find fifty agencies and start sending.
That is a weekend if it goes well. Two if it does not.
Or you open The Cleanroom and it is already built.
The scanner, running, with all eighteen checks implemented and tested. The Northstar fixture workspace, filled in and ready to demo. The conflict-matrix and review-queue sheets. The outreach-to-renewal pricing ladder. The quarterly update SOP that turns one sale into a renewal instead of a one-off.
You skip the build entirely and start at the outreach, which is the only part that actually pays.
And it is not one play. Pro opens the whole Vault: 45 plays, each a complete playbook with the buyer, the offer, the price you can charge, and the honest part telling you why it might fail. Six finished deeper builds sit on the shelf beside The Cleanroom. A new play banks most Mondays, and your price never moves.
$129 for the year, locked forever, for the first 25 members. After the 25th seat the price starts climbing toward $399 and never comes back down.
Out-yield the average.
Overyield is educational, not financial, legal, or business advice.