From desk to pack: assembling a reusable local AI writing kit

Affiliate disclosure: Some links are affiliate links. As an Amazon Associate I earn from qualifying purchases. See Disclosure.

Most “local AI writing setups” die as a pile of one-off Markdown files. A kit is different: the same folder map, the same prompts, and the same quality gates every time you start a post. On a 16 GB Apple silicon Mac mini that is the difference between a quiet desk and a second-model fight.

This post is the pack map I actually use under ~/AI/ and the published Hugo tree. It pairs with the quiet desk overview and the 16 GB checklist.

What the kit must do

A reusable kit answers five questions before any model runs:

Where does research live, and what is allowed to become a claim?
Which local model writes, and which only extracts or audits?
How do we reject AI filler before Hugo sees the draft?
Which /go/<slug>/ paths are real rows in links.csv?
Who signs off — human gate, or auto-publish only above a score threshold?

If any answer is fuzzy, stop. Shipping volume without those rails is how you get 50 thin posts and zero trust.

Folder layout that survives overnight jobs

Keep prose work off the LTX and Freqtrade lanes. A layout that has held up:

~/AI/
  prompts/          # deslop-audit, deslop-rewrite, niche-research, …
  bin/askai         # stdin → Ollama (tiered models)
  notes/            # rubrics, swap maps, ops logs
  tmp/deslop-YYYY-MM-DD/
~/Downloads/AI-outputs/
  local-ai-blog-affiliate/   # briefs, drafts, claims worksheets
  qa/                        # extract-claims.sh output
~/AI/tmp/publish/abhishekrhl.github.io/
  content/posts/             # Hugo markdown that ships
  content/go/                # /go/ stubs
  links.csv                  # allow-list only
  static/_redirects          # generated from links.csv

Rule: drafts never overwrite outlines. Copy forward (brief → outline → draft → claims → publish), and keep a dated backup under ~/AI/notes/backups-YYYY-MM-DD/ before a destructive rewrite.

Prompt pack (minimum viable)

Ship these four beside ~/AI/bin/askai:

Prompt	Profile	Job
`niche-research.md`	extract	Demand + gaps; mark unknowns as “no data”
`deslop-audit.md`	extract	Score 0–10 slop; quote phrases to cut
`deslop-rewrite.md`	write	Delete-first rewrite ≤ 90% length
`delete-first-shorten.md`	write	Hit ≤ 0.9×N when rewrite lands hot

Add hook / flow / specificity / CTA honesty prompts once the quality rubric gates settle — the local copy lives at ~/AI/notes/deslop-quality-rubric-2026-10-25.md on the desk.

Balanced tier on this machine: chat and extract on qwen3.5:9b; reserve heavier write models for nights when unified memory is free and Freqtrade dry_run is left alone.

Quality gates before publish

Do not “hope” the draft is fine. Score it:

Hook — first 2 sentences name the reader’s constraint (RAM, noise, privacy) and a concrete outcome.
Flow — each H2 advances one decision; no restating the 16 GB lecture three times.
Specificity — real paths, model tags, IST windows, or measured limits; no “studies show”.
CTA honesty — only /go/ slugs that exist in links.csv; placeholders OK until Associates UK lands; never invent ASINs.
Claims — run extract-claims.sh; empty [CITE:] blocks human review.

Auto-publish only if slop score ≤ 3 and all five gates pass. Otherwise the draft stays in AI-outputs/ until a human gate.

Affiliate inject without breaking trust

Affiliate links are a closed set. The pack includes:

links.csv rows → scripts/generate-go-redirects.py → static/_redirects + content/go/*.md
In-body links only as /go/<slug>/ (trailing slash)
One disclosure block at the top; a short gear box at the end is enough
Amazon.in tag abhishekrhl-21 where approved; UK URLs stay untagged placeholders

See affiliate allow-list hygiene for why invented slugs fail QA.

What this kit deliberately excludes

Fake reviews or sockpuppet promotion
Amazon Special Links inside ebook PDFs
Overnight video (LTX) overlapping prose hours — keep that to 01:00–06:30 IST
Stopping Freqtrade dry_run to free RAM for a write job

Next steps in the series

Once the pack is assembled, schedule it: LaunchAgents and the quiet desk schedule, overnight windows, and anti-patterns. The goal is a sustainable cadence of posts readers finish — not a raw 50/day firehose.

Desk gear that actually helps a 16 GB Mac mini (affiliate)

If model files and drafts are fighting the internal SSD, external NVMe via a Mac mini dock is the practical upgrade — not a bigger cloud bill.

UGREEN Mac mini dock with NVMe slot (affiliate) — ports + model-library drive in one stand.
Satechi Mac mini hub & stand (affiliate) — design-matched alternative.
Samsung T7 Shield portable SSD (affiliate) — fast offload for GGUF libraries when a full dock is overkill.
UK readers: UGREEN on Amazon.co.uk (placeholder — Associates tag lands after UK approval).

Disclosure: as an Amazon Associate I earn from qualifying purchases. See /disclosure/.