Weekly builds
72 builds across 10 weeks. Each tile is one piece of work, with what it produced and where it stands now.
Read this first
These tiles come straight from my own weekly build records. Work marked Blocked or Stopped is left in on purpose, because a record that only shows the wins is not a record. Process reviews and personal projects are not shown here.
Index
The flags on each tile
| Flag | What it means | |
|---|---|---|
| Status flags, on the right of each tile | ||
| Live | Running now and in use. | |
| Shipped | Finished and delivered. A one-time piece of work. | |
| In Progress | Started, with open items still to do. | |
| Blocked | Could not go forward. The tile says what stopped it. | |
| Stopped | Ended on purpose or retracted. The tile says why. | |
| Category flags, on the left of each tile | ||
| Website | Work on thebaddiestack.tech: pages, the blog engine and layout. | |
| Infrastructure | The systems underneath: AI employees, automations and file structure. | |
| Brand Content | Posts, newsletters, the logo and the brand databases. | |
| Analytics | Dashboards, build ledgers and audience analysis. | |
| Grants | Finding, scoring and applying for grants. | |
| Job Search | Resumes, applications and sourcing runs. | |
| Client Onboarding | Materials built to bring a client into a working system. | |
| Offers | Paid and free offers: sales pages, workshops and curricula. | |
| Admin | Errands handled end to end in the browser. | |
The weeks
| Week | Dates | Open |
|---|---|---|
| Week 10 | Sep 26–30 | Open week 10 |
| Week 9 | Sep 19–25 | Open week 9 |
| Week 8 | Sep 12–18 | Open week 8 |
| Week 7 | Sep 5–11 | Open week 7 |
| Week 6 | Aug 29–Sep 4 | Open week 6 |
| Week 5 | Aug 22–28 | Open week 5 |
| Week 4 | Aug 15–21 | Open week 4 |
| Week 3 | Aug 8–14 | Open week 3 |
| Week 2 | Aug 1–7 | Open week 2 |
| Week 1 | Jul 25–31 | Open week 1 |
Sep 29
Latest scheduled refresh: 68 threads, 24 projects, 12 reusable AI employees across Jul 25–Sep 29. Traffic numbers are still carried over from Aug 23 — no analytics connector reachable.
What it produced
Sep 29
The original eye-and-circuit mark kept as-is and recolored at the pixel level (caramel circuitry, espresso eye, cream background), with the name set in title case: The Baddie Stack.
What it produced
Sep 26–Sep 27
Ten tailored resumes against ten live postings: OneReach.ai, Clarasight, forgd, RunPod, Pencil, Spruce, Mistral, FAR.AI, Whatnot and Tailscale. Nine scored 100 and Pencil scored 98, but the score was deliberately redefined first. Batches 4–5 scored raw posting nouns, so honest domain gaps read as weak resumes and “clear 85” quietly became an incentive to fabricate. Batch 6 scores only defensible terms and lists every removed term with its reason in score6.py::EXCLUSIONS. The exclusion list, not the number, is the fit signal — and a 100 sitting next to Batch 5's honest 50 without that caveat would be worse than no number at all.
What it produced
Sep 27
Zero of 10 submitted, and nothing is in flight. The Tailscale form was fully mapped in the live DOM, about 15 form values were known, and the first fill was attempted — which is exactly the moment the rule about writing values to disk first had to be honoured, because a mapped form in a live browser tab is not a record. It wasn't. Every state-changing browser action failed: 3 fill attempts, 0 values written, timeout then browser_operation_failed twice. Diagnosis was extended to a second platform and a click on an Ashby board failed identically. Two platforms, two action types, one result.
What it produced
Sep 27
The batch also produced its own corrections, and those are the interesting part. Batch 6 went out reporting SQL, MCP and sub-agents as honest gaps; CO then stated SQL experience, a year of MCP builds and agent operations. Three gaps were stale and forgd's technical fit had been mis-scored. Later the same day VPN/networking and Docker/Kubernetes facts emptied the Tailscale exclusion list entirely, turning it from “strong responsibilities, weak domain” into the best match in the batch. Stale gap claims in brain/06, Andromeda Cluster (79) and Prima (74) are flagged for retraction rather than quietly rewritten.
What it produced
Sep 29
A 12-day thread closed out with thirteen numbered fixes across three areas. The recurring hero bug turned out to be a nested max-w-3xl/max-w-2xl cap inside an already-correctly-wide container — removing the inner caps was the fix, not touching the container again. The shared blog template was capped at max-w-2xl on both header and article wrapper while the rest of the site ran wider; one change fixes every current and future post. Homepage and blog-template work is confirmed live at commit e805283. The Services page round is final in source and was approved — but no push confirmation followed.
What it produced
Sep 29
The vault held to-do lists, Baddie Stack business work, social content ideas, podcast scripts and AI-employee skills in one undifferentiated pile. It was rebuilt as a numbered PARA-style structure — 00 Inbox, 01 Dashboard, 10 The Baddie Stack, 30 AI Employees, 90 Meta, 99 Archive — across 73 notes, with the video scripts pulled out into their own folder and back-links and hashtags so connections render visually. Then every one of the 34 installed plugins got a written note, reviewed against each plugin's live data.json rather than marketplace blurbs.
What it produced
Sep 28
Both unattended database runs fired on Sep 28 and both did the right thing up to the point of writing. Each correctly re-resolved the folder that had moved on Sep 24, refused to ask questions, found the prior dated output, and read it plus the base file to avoid duplicate brands. Then each stopped mid-run. No natural-clothing and no black-owned output file was produced for Sep 28, so the newest artifacts remain Sep 24 and Sep 25. Nothing was corrupted and nothing was overwritten — the failure is absence, not damage.
What it produced
Sep 19
Grocery Essentials (WF-01) now writes into the real Obsidian vault, confirmed by a live pipeline run. A real GitLab Runner was registered through the API (no menu clicking). The Amazon Returns workflow got its own signed-in browser profile, and order matching was built and dry-run tested — not yet run on a real order.
What it produced
Sep 19–Sep 20
The Black-owned database reached 10 brands, including Brylamar and Bask and Lather Co. New brands now wait in a review queue for CO's Add or Not Interested. Both databases and their 4 scheduled tasks were pointed at one canonical folder after a stale-path bug.
What it produced
Sep 19–Sep 20
HARVEST-LOG.md updated to close out Batch 4. Batch 6 built: 4 resumes, not yet applied. The JetBrains Senior TPM form was staged but not submitted — it needs CO's answer on a required location field.
What it produced
Sep 22–Sep 25
Two scheduled 3-day sourcing runs. Sep 22 surfaced 5 roles that cleared every hard filter; Sep 25 was the thinnest run yet with 1 match. All are waiting on CO's review; none are built into resumes yet.
What it produced
Sep 21
Four days after the last sweep, cadence holding. No new eligible cash opportunities; 3 discarded or scored out.
What it produced
Sep 21, Sep 25
Two scheduled dashboard refreshes. No analytics connector is installed, so traffic figures are still the Aug 23 pull. A wider harvest-counting blind spot was flagged, not fixed.
What it produced
Sep 13
“status-build” became a scheduled task every 4 days at 9pm PT that runs the documented build-status routine. Its first run came back empty; two bugs were found and fixed (trailing spaces iCloud adds to folder names, and folders that had moved). Coverage went from 0 folders resolved to 12 of 15. The Sep 1–12 snapshot page was built the same day.
What it produced
Sep 13–Sep 17
The Sep 13 sweep found four August deadlines had passed with nothing submitted (Hey Helen, Amber, HerRise, Women Founders) — logged as misses. Two were re-staged for the next cycle and two more were drafted. The Sep 17 sweep found nothing new that fit.
What it produced
Sep 13–Sep 14
A failure-brief page (built through v6) added to the site at /failure_brief with its own share image. The site's forms moved off Formspree onto Supabase through one shared intake endpoint. CO pasted the secret values into Vercel herself; a Name field was added after a submission was hard to find.
What it produced
Sep 14–Sep 17
All 7 Batch 4 forms filled (stopped short of submit; resume attach still blocked). CO submitted 3 on Sep 14 and confirmed all 7 by Sep 17. A suspected prompt injection inside a tool result was spotted, not acted on, and disclosed. After the Sep 16 sourcing run, CO made 6 corrections to search practice, including removing Workday from every query.
What it produced
Sep 17
Blog index rebuilt as a 2×2 grid with month headers and reading-time/pillar pills. 13 backlog posts redated to 2 a day with the Founder and Famous Baddie posts interleaved. The image-placement rule is now enforced in code across 62 posts — verified, but held uncommitted until CO says go. A failure brief and Architect Challenge review closed the thread.
What it produced
Sep 17–Sep 18
Homepage and services page reworked: chocolate nav, sections widened to match “Who I Support,” centered headings with full-width body text, card grids with the site-wide shadow standard, and the redundant contact section removed. The blog post template was widened for every post.
What it produced
Sep 4–Sep 17
The trigger reference grew into a new trigger: cowork defense saves a verbatim transcript of a thread plus copies of its outputs. The Trigger Terms dashboard was kept accurate as new triggers were added elsewhere.
What it produced
Sep 5–Sep 7
5 new tailored resumes (Prima, Heidi Health, Portcast, Sophos, Peratera), honestly ATS-scored (91 / 84 / 74 / 68 / 50 — 3 below the usual 85 floor, reported as-is).
What it produced
Sep 5
Working plan for ElevenLabs Startup Grant usage: what's capped, the HeyGen + ElevenLabs pairing, and a weekly build-first voiceover/video/podcast content pipeline. Brain-layer only so far (grant caps, credit rates, voice-slot limits) — not yet a full AI-employee build.
What it produced
Sep 8
Batch 4 documentation layer: which builds map to which target roles, plus case-study copy for LinkedIn. Sourcing queue and target-role/search-strategy brain files updated alongside it.
What it produced
Sep 8–Sep 11
Onboarding video and slide deck for a client on CO's own ops system — Proton Mail + Proton Calendar + Obsidian as the three-layer “Founder Foundation” (communication, commitments, context/memory).
What it produced
Sep 11
All 22 blog drafts expanded from 4–5 min to a new 6-min standing length (1,100–1,250 words). /blog/drafts rebuilt full-width, two-column. Mid-thread, a stale duplicate project folder silently overwrote a day's work (the expansion, the redesign, the image pipeline) — diagnosed and recovered via the raw session log and git plumbing commands, pushed clean.
What it produced
Sep 11
Turned 5 of CO's own screen-recorded work sessions into job-search portfolio material: a timestamp analysis and ranked brief per video, plus a shareable DNS-migration checklist page. The checklist was rebuilt on the canonical brand palette and a corrected voice framework — one voice everywhere, no restrained-vs-warm register split, with lifestyle-metaphor domains (skincare, haircare, makeup, fashion, horticulture, fragrance) written into CLAUDE.md as canonical.
What it produced
Sep 11–Sep 12
Voiceover scripts for the same 5 portfolio videos. The Email Signature script was approved as the working framework. The other 4 (DNS Debugging, File Rescue, Auto MD Sorting, Intro Hermes Install) went through several rounds — voice approximated from the style rules instead of CO's actual writing, then a cosmetic edit presented as a rewrite — and were still not usable as of Sep 12.
What it produced
Sep 11–Sep 12
New self-updating database of Black-owned lifestyle brands — fragrance, beauty, wellness, home. Tap a tile to mark “not a fit” or favorite/star it, with a free-text reason meant to shape future suggestions. Revised twice on CO's color-scheme feedback.
What it produced
Sep 11–Sep 12
A database of natural-fiber clothing brands, synthetics excluded, each verified against the brand's own materials page plus one independent source. 16 brands on launch (4 confirmed Black-owned), with a transparent “checked, not added” appendix. Then corrected onto the canonical brand palette after CO flagged the color scheme as hard to read.
What it produced
Sep 11
Companion Baddie Stack audience resource page for everyday grocery essentials.
What it produced
Sep 3–Sep 4
The search pivoted to international roles only — no US roles. A recurring 3-day sourcing digest was set up; its first run became Batch 4 and its second queued 6 more roles.
What it produced
Sep 4
Page redesign and local dev environment setup for thebaddiestack.tech, done in the same window the site's git repo was recovered from the 09-03/04 corruption (a clean backup of the corrupted state was preserved before repair).
What it produced
Sep 4
Published a reference dashboard of every active trigger phrase — account-wide (folder harvest, build status, sync the rules, execute) and project-specific (plan the blog; one project's daily-logging triggers) — pulled straight from the canonical CLAUDE.md.
What it produced
Sep 2
The client work and the reusable asset underneath it were deliberately separated, so the same vault could be handed to the next person unchanged. Twenty-nine notes turn the founder brain into something shippable: twelve numbered folders, nine templates, and a seeded set of example notes so a new user opens the thing and sees the structure already in use instead of an empty grid and a page of instructions. The onboarding copy was written to lower the barrier on purpose — you do not need to already know how Obsidian works to get value out of it. That separation is the difference between a bespoke build that cannot be repeated and an asset that can be sold more than once.
What it produced
Sep 2–Sep 3
The data layer the blog runs on, built before any of the visible page work. Three pieces, and the third is the one that matters: the posts file itself, the review-state file that tracks what has been read and approved, and a staging path that pulls a batch of new posts together with their matching review-state entries straight out of the inbox instead of hand-editing either file. This is the part that decides whether adding a post stays a two-minute job or turns into a fiddly one six months from now. Building it first meant every later change to rotation, layout and image placement had something stable underneath instead of a fragile convention.
What it produced
Aug 23–Aug 24
Dashboard refreshed with the last real Vercel pull (80 visitors / 262 pageviews, Jul 25–Aug 23). A weekly session-ledger task now measures threads so the uncounted-thread gap closes on its own.
What it produced
Aug 23–Aug 28
ElevenLabs grant accepted (2FA on, “Accept Your Grant” clicked, confirmed by CO). Sweeps ran Aug 24, 25, 26 and 28. The Hey Helen application was drafted, and Women Founders Grant and The Hustle Grant were verified and logged.
What it produced
Aug 24–Aug 27
The first recruiter screen of the whole search, from a Batch 2 application (~19 days to contact). The interview brief was built five times because questions were written into it instead of asked — closed with a failure brief. Scale figures corrected (40,000 = GitHub, 10,000-engineer build = Pluralsight).
What it produced
Aug 22–Aug 27
A lightweight proof of concept: check whether priority grocery items are in stock at two DoorDash storefronts, on GitLab's shared runners about every 2 days, keyed off the page's machine-readable stock field. The “failure brief” trigger was misread as a request for new infrastructure and corrected.
What it produced
Aug 26
Local git commit-back and a vault sync script for the grocery automation. An unverified folder path broke the setup commands and then shipped a second time inside the sync script. Closed with a failure brief; the vault-write gap was fully fixed on Sep 19.
What it produced
Aug 15–Aug 16
Grant discovery put on a daily schedule. A separate track opened for personal stipends to CO (judged on her own background), kept apart from business grants to The Baddie Stack LLC.
What it produced
Aug 16
The dashboard moved from a one-off build to a script that regenerates it from projects.json and metrics.json.
What it produced
Aug 17–Aug 18
A weekly newsletter employee: building-in-public updates, a Baddie Stack review, a Rose/Bud/Thorn protocol, podcast and blog pulls, and a CTA. Its intake “truth gate” sources every claim from commits or CO's answers. Five setup items are still open (tagline, repo access, podcast feed, blog feed, email platform).
What it produced
Aug 19–Aug 20
8 URLs in, 6 resumes built (ATS 87–98). CO's laptop crashed and all six filled forms were refilled from scratch. A scorer bug (keyphrases split across line breaks read as missing) was fixed, and all 8 resumes were regenerated with the full 10-tool AI list. CO confirmed all six submitted on Aug 20.
What it produced
Aug 20
A styled diagnostic of why 20 applications had produced 0 recruiter screens. CO challenged it and both headline causes were wrong — one was read off a stale LinkedIn export, the other turned an accessibility note into a claim about her. Retracted and replaced with a failure brief.
What it produced
Aug 20
The first grant win: ElevenLabs Startup Grants, about $4k in credits. The required logo and link-back went into the site footer the same day.
What it produced
Aug 20
A demo inbox was built for the Aug 20 workshop so live screen-shares wouldn't expose a real inbox. The session started with a refusal that CO's own CLAUDE.md already answered, and ended with Google disabling the demo account.
What it produced
Aug 8
First inventory of everything built since Jul 25: 31 threads, 16 projects, 7 reusable AI employees, 4,777 of CO's messages and 3,845 tool actions. Vercel showed 14 visitors / 97 pageviews for Jul 25–Aug 8 from 40+ deploys. Main finding: build velocity is far ahead of distribution.
What it produced
Aug 8
The 299 recovered links turned into a tap-through dashboard that works on any device, with search, category chips, a Focus swipe view, and add/remove without a Claude session. 125 generic titles (map pins, short links) resolved to what each link actually is.
What it produced
Aug 8
Ranked ~150 of CO's posts (Jan 1–Aug 6) by reactions, pulled reactors from 14 top posts, then reconciled the list against her real connections export so she knows who she can DM now.
What it produced
Aug 7–Aug 9
The ai-employee-builder skill codifies the account's employee pattern, and can build a new employee or finish a half-built one. The LinkedIn Engagement Posts employee was built in the same thread.
What it produced
Aug 11
The ATS agent became a complete employee: README, job description, CLAUDE.md, brain (identity, master resume source, target filters, ATS scoring method, known tool limitation) and a pipeline. Batch 2 marked submitted; GoodData logged as the rejection.
What it produced
Aug 11–Aug 12
The employee that reads across the other employees. Priority framework, a register of every project and agent, standing routines, and an explicit sensitive-data boundary.
What it produced
Aug 11
The original coaching guide reorganised into a client onboarding coach employee — the full three-phase method for walking a client into their own Website Employee.
What it produced
Aug 11
The build ledger became a repeatable, dated status dashboard.
What it produced
Aug 11
CO submitted the Galaxy of Stars application after Jax reviewed it as a good fit.
What it produced
Aug 1–Aug 5
Site 2.0 finished and went live: blog engine with a local draft-review inbox, Stripe-gated paid PDF delivery, cart, discovery form, portfolio and services pages. The 5 failed deploys of the whole window all came from Aug 3 (GitLab web-IDE truncated commits); every deploy since Aug 6 succeeded.
What it produced
Aug 1
Two years of links dumped into a WhatsApp self-chat, parsed, deduped, resolved and categorised. Turned into a live dashboard the next week.
What it produced
Aug 3
First discovery run, plus the rules that keep grant work safe: hard filters, fit scoring, a compliance wall, and a three-tier autonomy line because grant applications are signed documents.
What it produced
Aug 3–Aug 7
The folder connected to Cowork for months was the desktop app's own data directory. Everything was extracted, verified and rehomed. On Aug 7 the four drifting rulebooks were collapsed into one canonical core, synced into every project CLAUDE.md so the rules actually auto-load.
What it produced
Aug 4–Aug 6
The Website Employee generalised into a reusable starter — brain files, coaching guide, three-phase onboarding, prompt library. The pattern that other employees were cut from.
What it produced
Aug 4–Aug 6
Was a six-agent spec with nothing running. Now a full employee — job description, brain, harvest log — plus a produced content series: “Things I Wish AI Founders Knew,” eight episodes each with a podcast script and a video script.
What it produced
Aug 5
Batch 1: 10 companies evaluated, 8 applied, 2 discarded on hard filters. Batch 2: 8 evaluated, 6 resumes built and ATS-scored (87–100), 2 discarded. Batch 2 forms were filled but the tool couldn't attach the PDFs — CO submitted all 6 herself (confirmed Aug 11).
What it produced
Aug 5–Aug 6
The 3-day weekend intensive and its free front-door workshop, The Receipts Audit. Sales page, curriculum, facilitator run-of-show, workbooks, promo schedule and social cards — built and deployed inside two days.
What it produced
Aug 5–Aug 7
“Behind the Founder — Hotel Room” went out Aug 5 and became the best-performing non-newsletter post in months. A weekday engagement ritual was scheduled, and on Aug 7 the skill grew into a full employee folder with a voice brain and a performance log.
What it produced
Aug 3, Aug 5
A hotel refund pursued end-to-end in the browser in 50 minutes — 249 tool calls, written channels only, no phone. Two days later, an Amazon return QR code retrieved by direct browser control in 6 minutes.
What it produced
Aug 6–Aug 7
Full connection list classified by inferred gender (first-name library plus title fallback) to size the actual reachable audience for a product aimed at Black women in tech.
What it produced
Aug 7
Four case-study editions scheduled as LinkedIn Articles under the existing newsletter, one per Tuesday 9am (Aug 11, 18, 25, Sep 1), with image placeholders left for CO's screenshots. All 5 drafts then archived to disk and removed from the site's draft queue.
What it produced
Jul 27–Jul 31
The Baddie Site Revampt thread (Jul 27–29) reworked thebaddiestack.tech, then Site 2.0 started on Jul 30 — the Next.js rebuild that became the live site. Site 2.0 ran into the next week and finished Aug 5.
What it produced
Jul 29–Jul 30
Built the social-post-from-notes skill: rough notes or a voice memo in, an on-voice post out. It later became the base of the LinkedIn Voice employee.
What it produced
Jul 31
First day of the grant-seeking employee. The core brain files and LLC source documents went in, setting up the eligibility and scoring work that followed on Aug 3.
What it produced
