Claude Code Accelerator — digest (seed content for the Studio slice)
For: Vince, Sead, Agon; Mathew for the questions in §6. Source: the three archives Mathew shared after the workshop (anthropic-skills.zip, claude-code-accelerator-workspace.zip, all-skills.zip) and his Loom. Use: design input and seed content for cards V14, V16, V17, V18, V19, V20, S14, S15, S9 and A15; questions folded into D32. It changes no sequence, gate or owner.
1. What the material is
A methodology and a local skill library, not product code. The core is a single-operator "workspace OS" for Claude Code + Obsidian:
- The Script (
CLAUDE.md) — a thin top file: who I am, how I work, where things live; detail lives in folders, never in the Script. - The data dictionary — the rule for what compounds into memory: clear wins and losses are promoted as they happen; the second repeat of the same pattern becomes a reusable rule or play; by the third repeat it should already be a workflow.
- Workbench vs memory — production output (
copy/,creatives/) is separate from reusable lessons (winners/,losers/,brand/,workflows/). "Save the work, promote the lesson — two moves, not one." - The folder model —
brand/voice.md(rules, especially corrections after misses, banned words),clients/<name>.md(facts, proof we are allowed to use, current winners, dead angles),swipes/(organised by mechanism, not by brand, each with "the one thing to steal"),winners/andlosers/(hook, why it worked or died, where, the rule it taught, reuse conditions),inbox/(unsorted capture),daily/(running log),archive/(never delete, move). - The rituals (skills):
reflect(end of session — what is worth promoting, where),remember-this(mid-session capture of a standing rule),save-as-skill(third-time-becomes-a-skill with approval),handoff(compact restart context),adversarial-review(break it before shipping),convince-me(restate before acting). - Anthropic's
skill-creator— a real eval/benchmark loop for writing and changing skills.
World-class: the promotion rule, the workbench/memory split, mechanism-first swipes, proof-we-may-use, losers with the rule they taught, the eval loop. Generic: plan-as-page, see-your-setup, upskill, create-workflow-diagram, dashboard-setup (good craft, not workflow scope).
2. Already in the plan (no change)
| Accelerator idea | Skynet |
|---|---|
| Thin Script + detail in layers | V14 prompt layers: doctrine → workflow → brand/product → personal overlay, never forked |
| Taste / rules / pairs per operator; personal first, team after approval | V17 / V19 instinct loop (from Mathew's Instinct Transfer) |
| Workbench vs memory; only what is proven compounds | S15 validated memory: human-confirmed winners or human-rated outputs only; personal taste separate |
| Hook first; swipes and winners before blank prompts | V16 hook session; V18 brief packets with teardown / winner / market-intel / coverage evidence |
| Proof discipline ("never invent a client result") | Claim validation + locked product spec (functional design §3.4, §3.5) |
| Reddit / review data for the niche | A18 Reddit VOC (S2) |
| Long-form scripts as the first format | Long-form ports in S1 (V3/V10); the instinct layer over it is S3 |
3. Not in the plan before — added to the cards
- Losers as first-class memory — a durable dead-angle record with the rule it taught, written from confirmed losers the way winners are. → S15, S9.
- A swipe library by mechanism — a curated store (mechanism, source, "the one thing to steal") that feeds brief packets. → V18.
- The promotion ritual as a product step — end-of-session "what is worth keeping, where" and mid-session "remember this", both proposing captures the operator ratifies. → V17, V19.
- An eval loop for prompt/skill edits — a benchmark set per layer, run before a layer or skill version is published. → V14, V20.
4. Learned (adopted as rules)
- Clear wins/losses are promoted immediately; the second repeat becomes a rule/play; the third repeat becomes a workflow. Written into S15; V17/V19 carry the ritual moments that propose those captures.
- "Proof we are allowed to use", "what died / dead angles" and "what is working right now" are explicit product/client fields. Written into A15; V14 is the layer system that reads them, not the card that defines them.
- Mathew's local workspace is seed content, not throwaway: his Script, voice corrections, winners/losers and pairs become the first overlay and the first memory when S3 starts. Keep the folder intact.
5. Not usable as-is
Obsidian [[wikilinks]] as navigation; .claude/skills/ as runtime packaging; folders as canonical state; per-person daily/ notes as session continuity; skills that assume Claude writes straight into the operator's filesystem. All one-operator-one-laptop conventions. In Skynet these become entities with owners, permissions, evidence packets and lineage; the rituals behind them survive as product steps.
6. Questions for Mathew (folded into D32)
- Which three artefacts do you touch every day: winners, losers, swipes, daily, inbox, client file?
- What crosses from personal instinct into shared team memory, and who approves it?
- Should losers / dead angles be first-class in the platform, or implied inside winners and coverage?
- Should the swipe bank be a shared library by mechanism, or stay personal?
- Is "third time becomes a workflow" worth a productised capture flow?
- Outside the archives but raised in Mathew's Loom: is "one format first, LFS, blood sugar" a pilot tactic or the general operating rule?
The videos are not required for the system; they answer rationale and daily practice, which these six questions cover.