clembot

clembot.wanessalabs.com, index

An agent harness
you can audit.

Clembot is the agent system this studio works with: a roster that drafts, reviews, schedules and logs, and writes down everything it did while doing it. This site is built from that record. Not a case study about the record: the record.

See the receiptsTry the approval gate

Drifting behind this page: the approval threshold, with no threshold drawn. These carriers are a sample, gated at the roster's real rate of 15 in 115. A gated one halts wherever it happens to be and stays there until the review line sweeps down past it. The rest cross on their own.

evidence chain · headd8b5854933025037818d0f09d45ebe4dbfd7ea1cd23c714e96cfb93baf3912fbanchored 2026-07-22

the roster

Every agent, and what it is allowed to do

115 nodes and 45 relationships, exported from the system's own definitions rather than drawn by hand. Colour is the department. The ring is the tier: whether a thing can act on its own, or stops at the approval gate first.

Constellation of 115 units across eight departments. Each node links to its full entry in the list below this diagram.
bip-batch, command, human-triggered, defined not proven/bootstrap-new-project, command, human-triggered, defined not proven/brand-check, command, human-triggered, defined not provenBrand Guardian, agent, approval-gated, defined not proven/campaign, command, human-triggered, defined not provenCampaign Architect, agent, human-triggered, defined not provenCarousel Drafter, agent, human-triggered, defined not provenCode Reviewer, agent, approval-gated, defined not proven/competitor-scan, command, human-triggered, live/content-repurpose, command, human-triggered, defined not proven/content-review, command, human-triggered, defined not provenContent Writer, agent, human-triggered, liveDebugger, agent, approval-gated, defined not proven/deploy, command, human-triggered, defined not proven/design-asset, command, human-triggered, defined not provenDesign Director, agent, human-triggered, live/distribute, command, human-triggered, defined not provendistributor, agent, approval-gated, liveDoc Writer, agent, human-triggered, defined not provendoorman, agent, read-only, defined not proven/dream, command, human-triggered, defined not proven/dream-dry-run, command, human-triggered, defined not provendream-keeper, agent, approval-gated, defined not provenfeed-fetcher, agent, read-only, defined not proven/fix-issue, command, human-triggered, defined not provenfollow-resolver, agent, read-only, defined not provenfollows-digester, agent, read-only, defined not provenideate-leadership, command, human-triggered, defined not proven/launch-network, command, human-triggered, defined not provenleadership-persona, agent, approval-gated, defined not provenLinkedIn Drafter, agent, human-triggered, liveMarket Analyst, agent, human-triggered, liveMarketing Dispatcher, agent, human-triggered, livemid-session-save, command, human-triggered, defined not provenClembot-Brain, persona, read-only, liveClembot-Dev, persona, read-only, liveClembot-Marketing, persona, read-only, liveClembot-Publisher, persona, read-only, liveClembot-Research, persona, read-only, liveClembot-Sales, persona, read-only, livepost-summarizer, agent, read-only, defined not proven/pr-review, command, human-triggered, defined not proven/project, command, human-triggered, defined not provenpublish-thought, command, human-triggered, defined not provenPublisher, agent, approval-gated, defined not proven/recall-search <query>, command, human-triggered, defined not provenRefactorer, agent, approval-gated, defined not provenAmbient conversations, routine, read-only, liveAnatomy case-study ingest, routine, human-triggered, live/bip-batch, routine, human-triggered, liveBuild brief export, routine, read-only, liveContent-generation commands, routine, human-triggered, livecontent sync-postiz, routine, read-only, liveDaily Product Pulse, routine, read-only, liveFollow / Podcast poller, routine, read-only, liveGmail Poller, routine, read-only, liveGmail Scout, routine, read-only, liveList30 Content Loop, routine, approval-gated, liveNewsroom standup, routine, approval-gated, liveNotion Calendar Watcher, routine, read-only, livePOW Debt Scanner, routine, read-only, livePublic site feed export, routine, read-only, livePublic site map export, routine, read-only, liveSabbatical Weekly Digest, routine, read-only, liveStack Supervisor, routine, read-only, liveStale Deploy Detector, routine, read-only, liveStale Draft Checker, routine, read-only, liveSTATE.md Daily Sync, routine, read-only, liveUpcoming Drafts Digest, routine, read-only, liveWeekly LinkedIn build log, routine, approval-gated, liveWeekly LinkedIn Draft — RETIRED 2026-07-26, routine, read-only, supersededWeekly metrics pull, routine, read-only, liveSabbatical Tracker, agent, human-triggered, defined not provenSecurity Auditor, agent, approval-gated, defined not proven/seo-pulse, command, human-triggered, defined not provenSEO Specialist, agent, human-triggered, defined not proven/session-end, command, human-triggered, defined not proven/setup-client-routines, command, human-triggered, defined not provenshare-poster, agent, approval-gated, defined not provenarticle-writing, skill, not executable, livecampaign, skill, not executable, defined not provencompetitor-scan, skill, not executable, defined not provencontent-engine, skill, not executable, livecontinuous-learning, skill, not executable, livelaunch-network, skill, not executable, livemarket-research, skill, not executable, livemarketingskills, skill, not executable, liverebrand-asset, skill, not executable, defined not provenseo, skill, not executable, liveseo-pulse, skill, not executable, defined not provensuperpowers, skill, not executable, livetweet-thread, skill, not executable, liveweekly-report, skill, not executable, defined not provenwrite-blog, skill, not executable, liveSocial Strategist, agent, human-triggered, live/status, command, human-triggered, defined not proven/sync-state, command, human-triggered, defined not provenTest Writer, agent, approval-gated, defined not proventhought-leadership, command, human-triggered, defined not proventhoughts-publisher, agent, approval-gated, defined not proven/tweet-thread, command, human-triggered, defined not provenurl-unfurler, agent, read-only, defined not provenAsset librarian, vacancy, not executable, vacantIncident response, vacancy, not executable, vacantCost accounting, vacancy, not executable, vacantInvoicing, vacancy, not executable, vacantRoadmap / spec, vacancy, not executable, vacantUser research, vacancy, not executable, vacantCRM sync, vacancy, not executable, vacantLead qualification, vacancy, not executable, vacantOutreach sequencer, vacancy, not executable, vacantProposal / SOW writer, vacancy, not executable, vacant/vet, command, human-triggered, defined not proven/weekly-report, command, human-triggered, defined not proven/write-blog, command, human-triggered, defined not proven
  1. Marketing DispatcherCampaign ArchitectA campaign request arrives at the dispatcher, which hands it to the architect.
  2. Campaign ArchitectContent WriterThe architect breaks the campaign into pieces. A writer drafts one of them.
  3. Content WriterBrand GuardianThe draft reaches the gate, where the rules stop it. Nothing past this point is defined to happen on its own.
  4. Brand GuardianPublisherApproved, and the publisher is defined to distribute it and append a record to the chain. It has never run under that name.

This is the path the system declares, not a recording of one run. Of the 5 units on it, 2 have receipts in the evidence chain: Marketing Dispatcher and Content Writer. The other 3 have none, so Campaign Architect, Brand Guardian and Publisher have never run under their own names here. The five task groups that looked like recorded runs were written in one batch against sample data, so they are excluded. See what is actually recorded.

Most of the roster is read-only, which is what keeps the gated tier small enough for one person to review by hand. Asubagent holding no write tool cannot write, whatever a prompt talks it into. Open the full map.

By the numbers

commits in the ledger
138
2026-05-04 → 2026-09-14
evidence records
169
15 agents · 19 commands
agent messages · 30d
202
across 6 personas
shipped or in production
6
listed below, with links where they exist

the arc

A 13-week plan, 21 weeks in, still shipping

Started 2026-04-24, planned through 2026-07-23. Every phase below shows what the plan said and what the record shows: commits from the git ledger, and projects that reached the portfolio inside that phase's own dates.

The config still calls two phases "planned", months after the work in them landed. That is left as written. A plan is a claim about the future and the ledger is a record of the past, so where they disagree the ledger is the evidence. The config says week 15. The calendar says 21.

  1. 01

    Foundation

    weeks 1-4 · 2026-04-24 to 2026-05-21plan: complete

    36 platform commits

    • marketing-bootstrap scaffold complete
    • clembot-vault structured with project template
    • clembot/sabbatical and clembot/linkedin scaffolded
    • distributor v1 specced (Instagram adaptor)

    + 1 goal withheld, named client work

  2. 02

    Build

    weeks 5-8 · 2026-05-22 to 2026-06-18plan: planned

    15 platform commits

    • Notion pages fully structured and populated for each active project
    • Dashboard reading live vault data (not mock)

    + 4 goals withheld, named client work

  3. 03

    Launch

    weeks 9-12 · 2026-06-19 to 2026-07-16plan: planned

    24 platform commits

    • First paying Clembot client signed (upsell from Slack community)
    • Clembot install package documented and repeatable (onboarding playbook)
    • Public launch post — "here's what I built and what it cost"
    • Pricing confirmed and live: $100/month retainer (client pays own Claude Code usage)
  4. →

    After the plan

    8 weeks · 2026-07-17 to 2026-09-14not planned, happened anyway

    63 platform commits · 18 projects shipped across the portfolio: Clembot Doorman, CertScroll, Reddit Dishwasher Daily, Reddit Fibbers Dice Daily, Reddit Poker Daily, Lemon Coach, and 12 more

    The plan ended 2026-07-23. The ledger ran 8 weeks past it, and this page was built in that stretch.

The two figures in each row come from different places and do not measure the same thing. Commits are the platform repository's git ledger, which is private, so it is published as metadata only: date, type, subject, and how many lines moved, with 3 subjects withheld in full and counted anyway. That ledger starts 2026-05-04, ten days into phase 1, so phase 1's count is short by whatever happened first. Every one of its 138 commits falls inside the windows above. Shipped projects are read from the portfolio's own source and span every repository, not just that one.

how it works

7 rules doing most of the work

01

A human approves everything that leaves the system

No timer publishes. Every outbound post passes a person first, and the switch that would change that ships defaulted off.

02

One writer per table

A single command-line tool owns the content lifecycle. The publishing tool and the client-facing workspace are projections of it, never sources of truth.

03

An unverified number gets a label, not a guess

If a figure isn't in the config, it ships marked [VERIFY]. A plausible number is worse than a missing one, because nobody checks it.

04

Redact by allowlist, and fail closed

Everything published here is built field-by-field from a literal, scanned before it is written, and the scanner aborts the whole export on a single hit.

05

The log is the product

Every agent action writes an evidence record. The records are hash-chained, so editing history is detectable rather than merely discouraged.

06

Code and client data live in different repositories

The engine is one repo. Every install's data is a separate vault the engine reads through an environment variable. They are never mixed.

07

A tool's description is untrusted input

Everything an external tool publishes about itself is loaded into the model's context before a single call is made. It is data that arrives looking like instructions, so it gets read and graded before adoption, by something that is not the model.

output

What actually shipped

That list is exported from the project's own handoff file, which is why some entries have no link, not everything that ships is a public URL. The maintained index of completed work lives at wanessalabs.com/projects ↗.

Point it at your own build

Clembot is not for sale. It is how the work gets done, published so you can audit the method rather than take the output on trust. What you can install is Doorman: it reads the prompts you have already typed, says what your build keeps reaching for and does not have, and refuses a tool that should not be there. Free, offline, and nothing about your build leaves it.

What Doorman does