Business Plan Skills

Business Plan Skills is a set of 137 skills that help Claude Code or Codex write, test and release business plans. It covers bankable plans for lenders, investor-grade plans, standalone marketing plans, pitch decks, and the strategy, finance and policy documents around them.

It is for founders, consultants, boards, lenders and investors, mainly in East and Central Africa. Use it when a plan must keep assumptions apart from evidence, tie its narrative to its model, and claim no more than the evidence reviewed.

At a glance

• 137 active skills in 16 categories: 134 under skills/ and 3 country-context skills for Uganda (the default), Kenya and Tanzania.
• Routing: on 31 October 2026, 80 of 80 test prompts reached the right skill in the top three, and 64 of 80 (80.0%) reached it first.
• Checks: 13 of the 14 checks listed under Validate exit 0. The evidence-register check exits 1 because 4 of 24 source rows are overdue.
• Install: a project-scope install writes 903 files, including 137 SKILL.md files, and doctor reports 0 missing.
• Audit score: 50.0 out of 100 on 31 October 2026, against a target of 95.
• Not a lender model: the seven shipped Excel workbooks are starting structures. The SaaS projection workbook has no balance sheet and no debt schedule.

Start here

1. Install in Claude Code:

   /plugin marketplace add https://github.com/peterbamuhigire/business-plan-skills
   /plugin install business-plan@chwezi-business-plan
   

   The installer script, Codex and manual routes are under Install and update.

2. Check the engine from a clone. This needs Python 3.11 or later and PyYAML:

   git clone https://github.com/peterbamuhigire/business-plan-skills
   cd business-plan-skills
   python3 -X utf8 scripts/validate_skill_engine.py --baseline docs/quality/skill-quality-baseline.json
   

   Expect active skills: 137 and fully compliant: 137, with exit 0.

3. Ask for work in plain words, for example "Run the client intake and list the missing assumptions." The business-plan orchestrator runs a whole engagement, and Find the right skill lists other starting points.

Contents

• When to use it
• Capabilities
• Find the right skill
• How it works
• What is inside
• Install and update
• Validate
• Quality and improvement
• What the October 2026 kaizen improved
  • Wrong numbers and spreadsheet checks
  • Evidence, country context and packaging
  • Routing and documentation
  • How it was measured, and what was reverted
• What is still open
• Cited sources
• Contributing, author and licence

When to use it

Use this engine when the work is a business plan, or a document that feeds or tests one:

• a full plan, from client intake and evidence design to assembly and release;
• a bankable plan for a bank or development-finance institution, or an investor memorandum and pitch deck;
• a financial model, bankability score, stress test, valuation or investment-committee rehearsal;
• a standalone marketing plan, a grant proposal or a quarterly execution plan;
• a finance, procurement, HR, governance, grants, controls or monitoring-and-evaluation manual.

Other Chwezi engines are consulted alongside this one, never instead of it:

• the Chwezi Accounting Doctrine engine (chwezi-accounting-doctrine) for IFRS and IAS, tax and controls;
• chwezi-design-engine for the look of any deliverable;
• the Digital Research Engine (GitHub repository chwezi-research-engine; catalogue id digital-research-skills) for current facts;
• proposal-skills for tenders and PPDA submissions.

Resolve each through your engine-routing table. No path to a sibling engine is stored here.

Capabilities

The engine holds 137 active skills in 16 categories, grouped under What is inside. The skill catalogue lists each one. The table under Find the right skill is the capability map: common tasks and the skill that handles each.

Find the right skill

Each row is a task that the routing fixtures in tests/routing-fixtures.json already use. The skill catalogue lists all 137 skills.

How it works

AGENTS.md is the router. It works with any runner. CLAUDE.md imports it for Claude Code. Always-on principles live in rules/.

A skill is a directory that holds a SKILL.md file and, for deeper material, a references/ folder. The router names the right SKILL.md, and the assistant reads it from disk.

Full plans start at the business-plan orchestrator. It runs an engagement through these stages:

• client intake and evidence design;
• sixteen numbered plan sections, with a grant-proposal variant;
• the financial model and the investment-committee challenge;
• assembly and release.

Standalone marketing plans start at the marketing-plan orchestrator.

Three rules hold across every skill:

• Assumptions and evidence are kept apart.
• The narrative and the model must reconcile.
• A release claim goes no further than the evidence actually reviewed.

Multi-phase work follows the runtime-agnostic orchestration contract. It sets scoped work packages, evidence checkpoints, context hygiene, least agency, and sanitised handling of external content.

Country context for Uganda (the default), Kenya and Tanzania supplies currency, tax, regulatory, labour and banking facts. Verify each one at the time of writing.

The engine works to standards that it cites directly:

• IFRS 15, IAS 20, IAS 21, IFRS 13 and IFRS for SMEs for revenue, grants, foreign currency and fair value in projections. Accounting doctrine itself is deferred to the Chwezi Accounting Doctrine engine.
• COSO and ISO 31000 for internal control and risk.
• The IFC Performance Standards (2012) for environmental and social management in development-finance submissions.
• ISO/IEC 27001 and ISO/IEC 42001 for security and AI governance in SaaS due diligence.
• The ICC/ESOMAR International Code for market research.
• The PPDA framework for public procurement in Uganda.

What is inside

The engine has 137 active skills: 134 under skills/ and 3 under country-context/. Templates in templates/ are not counted, and neither are sector guide files that are references rather than skills.

The full list, with a one-line statement of what each skill does, is in the skill catalogue.

Beyond the plan pipeline, the skills cover:

• business-model design, market validation and competitive intelligence;
• pricing, valuation, bankability scoring, stress testing and due diligence;
• SaaS, AI and agent-product economics, ICT and e-commerce businesses, and hospitality;
• pitch preparation and organisational policy manuals.

The engine produces:

• complete business plans and plan sections, bankable plans for banks and development-finance institutions, investor memoranda and pitch decks;
• Excel financial models, built by the scripts in scripts/build-financial-models/;
• bankability scores, stress tests and valuations;
• marketing plans, competitive-intelligence reports, grant proposals and quarterly execution plans;
• finance, procurement, HR, governance, grants, controls and monitoring-and-evaluation manuals.

The shipped workbooks are not lender models

The seven shipped Excel workbooks are starting structures, not finished lender models. The SaaS projection workbook has no balance sheet and no debt schedule. Its Adjusted EBITDA adds back depreciation and share-based pay that its operating expenses never charge.

Section 0 of its reference lists each gap with cell references and gives the procedure for a lender case. The open fix is recorded under What is still open.

Install and update

Prerequisites

• Claude Code or Codex.
• Node.js 18 or later, for the installer scripts and the hook tests. The installer and both hook test files ran on Node.js 22 on 31 October 2026. Version 18 is the minimum the installer states, and it was not tested.
• Python 3.11 or later, for the validators. CI uses Python 3.12 with PyYAML==6.0.2. The 31 October 2026 run used Python 3.14.8 with PyYAML 6.0.3. The validators and the test suite need only PyYAML.
• Optional: LibreOffice for workbook recalculation, xlsxwriter for the model builders, and the packages in tools/requirements.txt for tools/scrape_metrics.py. What each check needs says which check uses which.

Claude Code plugin

The repository ships a marketplace in .claude-plugin/:

/plugin marketplace add https://github.com/peterbamuhigire/business-plan-skills
/plugin install business-plan@chwezi-business-plan

Installer scripts

install.sh and install.ps1 call scripts/install-engine.js. It copies the engine into ~/.claude (--scope user, the default) or into .claude under the current directory (--scope project). Add --dry-run to see the plan without writing anything, or --json for machine-readable output.

git clone https://github.com/peterbamuhigire/business-plan-skills
cd business-plan-skills
bash install.sh --scope project --dry-run
bash install.sh --scope project

Run the script with bash. install.sh is committed without its executable bit, so ./install.sh fails with "Permission denied" on Linux and macOS. On Windows PowerShell, use .\install.ps1 --scope project.

A project-scope install into an empty directory on 31 October 2026 wrote 903 files. They include 137 SKILL.md files, the hooks and the rules. Then node scripts/install-engine.js doctor --scope project reported 0 missing and 0 locally modified.

Update

For an installer-script install, pull the repository and run the installer again with the same scope. On 31 October 2026 a second run into the same directory succeeded, and doctor still reported 0 missing and 0 locally modified. That test used an unchanged checkout. It shows that the command runs, not that a changed tree upgrades cleanly. A plugin install updates through Claude Code's plugin manager, which was not tested.

Codex

Codex reads AGENTS.md as the router. Before substantive work, run the model-policy check described in .codex/README.md: python <engine-root>/.codex/ensure_model_policy.py --runtime codex --check. Claude skips that section.

The shared compatible plugin manifest and the repo-local marketplace expose the skills through Codex's local marketplace path. From a supported Codex CLI or client, add the repository marketplace and install the plugin:

codex plugin marketplace add .
codex plugin add business-plan@business-plan-local

These local installation checks do not establish public-directory acceptance. The Claude marketplace and the Codex repo marketplace are distinct catalogues.

Manual use

Clone the repository, read CLAUDE.md (Claude Code) or AGENTS.md (any runner), then open the SKILL.md that the router names. Skills are read directly from disk. No registration is required.

Validate

Checks

Run these from the repository root. The commands use python3 (use python on Windows) and change no tracked file. Python bytecode caches are git-ignored, and -B avoids them. What each check needs says which check uses PyYAML, Node.js, LibreOffice or xlsxwriter.

python3 -X utf8 scripts/validate_skill_engine.py --baseline docs/quality/skill-quality-baseline.json
python3 -X utf8 scripts/routing_smoke_test.py --threshold 1.0 --report --min-precision-at-1 0.80
python3 -X utf8 scripts/routing_link_check.py
python3 -X utf8 scripts/check_text_integrity.py
python3 -X utf8 scripts/check_package_file_budget.py
python3 -X utf8 scripts/source_ingestion_guardrail.py
python3 -X utf8 tools/evidence-register/refresh_evidence_register.py --check
python3 -X utf8 tools/sector-gates/validate_sector_gates.py
python3 -X utf8 tools/exemplar-packs/validate_exemplar_packs.py
python3 -X utf8 scripts/build-financial-models/verify_workbooks.py
for bundle in examples/full-plan-packages/*/release-bundle.json; do
  python3 -X utf8 tools/release-gate/validate_release_bundle.py "$bundle"
done
python3 -X utf8 -m unittest discover -s tests -p "test_*.py"
python3 -X utf8 skills/meta-utility/skill-writing/scripts/quick_validate.py <skill-directory>
node hooks/test-destructive-bash-gate.js
node hooks/test-plugin-hook-config.js

CI runs the catalogue, routing, evidence, sector-gate, workbook, exemplar-pack, release-bundle and unit-test checks, plus a compileall. The workflow is skill-quality.yml.

source-ingestion-guardrail.yml runs the guardrail. The text-integrity and file-budget checks also run inside validate_skill_engine.py. routing_link_check.py, quick_validate.py and the hook tests are not in CI.

Results on 31 October 2026

These runs used Python 3.14.8 and Node.js 22.

What each check needs

The table comes from the imports in the code and from running the checks with and without each tool. The PyYAML row comes from the import yaml lines only, because no check was run without PyYAML.

Notes on the table:

• Without LibreOffice, recalc_check.py exits 2 (NOT_ASSESSED) and the 7 tests are skipped, not passed.
• On a Python without xlsxwriter, each model builder stops with ModuleNotFoundError. The checks and the unit tests do not import it.
• No check imports openpyxl. It was absent from both interpreters used, and every check above still passed.
• The pytest run used python3 -m pytest -p no:cacheprovider tests.

Workbook checks

For a delivered XLSX model, run both workbook tools and keep both reports. formula_map.py reads the file's structure and stored values. recalc_check.py recalculates a copy in LibreOffice and tests tie-outs on the recalculated values, so two formula-driven sides can be compared.

python3 -X utf8 tools/workbook-audit/formula_map.py <workbook.xlsx> --require-scenarios --require-balance-checks --require-wired-inputs --json-out <formula-map.json>
python3 -X utf8 tools/workbook-audit/recalc_check.py <workbook.xlsx> --checks <contract.json> --require-checks --require-fresh-cache --json-out <recalc.json>

The model builders overwrite the shipped .xlsx templates in place. Run them in a copy of the repository, never in the checkout you intend to commit.

All seven builders ran in a copy on 31 October 2026, inside a scratch virtual environment with xlsxwriter 3.3.1. verify_workbooks.py then passed all seven workbooks. The system Python on that workstation has no xlsxwriter, so there each builder stops with ModuleNotFoundError.

Text integrity

scripts/check_text_integrity.py scans Markdown outside docs/ for damage that a structural validator cannot see:

• control or replacement characters;
• glued numeric ranges (35% where the source said 3–5%, or two decimals fused into one number), glued year ranges and glued thousands groups;
• host-specific paths (drive letters, a Windows web-server folder, home directories);
• any line whose hash matches one of the 298 stripped copies in tests/fixtures/stripped-operator-line-hashes.txt.

It exists because a bulk rewrite in April 2026 removed non-ASCII characters from 47 files and turned numeric ranges into wrong numbers. The Uganda inflation row read 35% until the original 3–5% was restored on 31 October 2026.

Run against the tree as it stood before the fix (commit 0d1eb51, with the hash file supplied by --hashes), it reported 306 findings and exited 1. Run against the current tree, it exits 0. It also runs inside validate_skill_engine.py. Do not weaken the check or its fixture to make an edit pass. Keep the dash in a numeric range.

File budget

The file-budget checker applies a fixed 5,000-entry ceiling to both the full checkout and a package-oriented count. It guards against the ZIP entry limit documented for OpenAI plugin submissions (submission errors). The counts are not a built-archive measurement.

Quality and improvement

Kaizen evidence is kept in dated folders. Each holds the same seven files: scope and evidence, baseline scorecard, backlog, experiment log, validation record, final report and next cycle. The documentation index lists the audits, kaizen records, source registers and quality gates under docs/.

• docs/kaizen/2026-10-engine-kaizen/: the 29 October 2026 cycle.
• docs/kaizen/2026-10-engine-kaizen-pass-2/: pass 2, 31 October 2026, with the open work and decisions and the documentation follow-ups.
• docs/kaizen/2026-10-readme-kaizen/: the 31 October 2026 rewrite of this README. It is a scoped cycle and publishes no score.

Latest published score: 50.0 out of 100 (pass 2, 31 October 2026). The raw mean and the published value, min(raw, 65), are both 50.0, because the cap of 65 does not bind. The target of 95 is not met, and no dimension reaches 65. The lowest are accessibility and inclusion at 25 (not worked on) and standards currency at 40. The scores are practitioner judgements, not validator pass rates.

• The 29 October cycle first froze a baseline of 60.0, then recalibrated it to 42.5 because the original ratings overstated structural evidence. Its final report keeps both, and 42.5 is the comparable starting point.
• Pass 2 measured a fresh baseline of 44.0 before it changed anything. That is lower than the 29 October final because pass 2 assessed things that cycle left not assessed. Engine content did not change between the two dates.
• In pass 2, hygiene rose from 40 to 60, applied proof from 40 to 50, and standards currency from 30 to 40. Doctrine, routing, production and handoff, and safety and integrity each rose from 50 to 55.
• Skill depth (50), output readiness (55) and accessibility and inclusion (25) did not move.
• All scores are taken from the evidence files. None was re-measured.

What the October 2026 kaizen improved

Two cycles ran in October 2026.

• The 29 October cycle: final report, experiment log and validation record.
• Pass 2, on 31 October: final report, backlog, experiment log and validation record.

A documentation wave followed pass 2 on the same day. This section and the four that follow list only changes that were committed and kept. Each short hash was checked with git show --stat and is an ancestor of main. Real defects come first, ordered by severity.

"Re-measured" means the command was run again on 31 October 2026, on the tree from before the change (extracted from git into a scratch directory) and on the current tree. "From the evidence" means the figure is copied from the file named and was not re-run. Everything that needs LibreOffice ran with LibreOffice 26.2.6.3. None of the re-measurements ran on Windows, on a CI runner or with a paid model call.

Kept changes: wrong numbers and spreadsheet checks

• Wrong numbers in plan content. A bulk rewrite in April 2026 (e1a9aa1) deleted every non-ASCII character from 47 files, so numeric ranges became other numbers. The Uganda inflation row read "35% (2025)" where the source said 3–5%. The bankability checklist read "DSCR 1.25x" with the ≥ sign gone.
  • Fix: pass 2 restored the numeric operators on 256 lines in 33 files, each proven to be an exact strip of a pre-damage line (e762dbc). It added scripts/check_text_integrity.py with 298 known-line hashes, run inside validate_skill_engine.py (423214c).
  • Re-measured: the current check run on the pre-fix tree reports 306 findings and exits 1 (254 known stripped lines, 25 glued thousands groups, 13 glued year ranges, 4 glued decimal ranges, 10 host paths). On the current tree it exits 0.
  • Not restored: 1,370 lines that lost an em dash, arrow, section sign or box-drawing character. This is an open decision, below.
• A mismatched workbook passed the spreadsheet check. A synthetic workbook with cash 100 against balance-sheet cash 80 passed formula_map.py with zero findings.
  • Fix (4b2def5, 29 October): an opt-in reconciliation contract (literal inputs, a declared subtraction, a tolerance, a numeric stored result) now fails it. validate_release_bundle.py recomputes the check from the workbook bytes instead of trusting a supplied PASS report.
  • Re-measured: the check as it stood before the change exits 0 with no findings on that workbook. The current tool with the contract exits 1 with 2 findings.
  • The forged-report and altered-hash cases are unit tests in tests/test_release_bundle.py.
  • Limit: verified against synthetic fixtures only. No real client model was reviewed, and the four example release bundles still report decision=blocked (re-measured).
• A formula-driven model could never pass the gate. In a real three-statement model both sides of a tie-out are formulas, which the contract above rejects. The workbook generators also write formulas without a stored value, so the file stores 0.
  • Fix (1350d4a): tools/workbook-audit/recalc_check.py recalculates a copy in LibreOffice (throwaway profile, original untouched). It reports stale stored values and error cells, and tests tie-outs on recalculated values. With no engine it exits 2 (NOT_ASSESSED), never a pass.
  • Re-measured on the built-to-fail fixtures in the pass 2 labs/: the earlier structural check returns reconciliation-source-not-literal for the mismatched and the matched workbook alike. recalc_check.py fails the mismatch (sides differ by 25.0) and passes the matched one.
  • On the shipped projection workbook it reports 1,772 of 1,852 formula cells whose stored value differs from a recalculation.
  • Limits: seven of its 19 tests need LibreOffice and are skipped without it. The release gate does not yet require it.
• Dead inputs and prose-satisfied balance checks. --require-balance-checks passed the projection workbook because the words "Ending ARR" matched, although the workbook has no balance sheet. Also, 10 of its 104 numeric inputs fed no formula.
  • Fix (9c044f2): formula_map.py now reports unreferenced inputs (--require-wired-inputs) and requires balance evidence on a formula row. 7 new tests.
  • Re-measured: on the shipped projection workbook --require-wired-inputs exits 1 with 10 findings. An independent hand scan in pass 2 found the same count (from the evidence: 00-scope-and-evidence.md).

Kept changes: evidence, country context and packaging

• Overdue evidence rows read as current. Ten of 23 register rows were past their recheck date while marked verified-current. The Markdown table still said so after the JSON said stale.
  • Fix, 29 October (c2adec7): an overdue verified-current row is now a validation error, and the ten rows were marked stale.
  • Fix, pass 2 (d4f9943): each issuer page was read for the claim it supports. Six rows were renewed with narrowed claims. Four stayed stale because the page could not be read (certificate error, HTTP 403, no body text). A primary-source Bank of Uganda row was added. The Markdown table is now generated from the JSON and checked for drift.
  • Re-measured: overdue rows 10 of 23 to 4 of 24. An agent did the review, not a human reviewer, and --check still exits 1.
• Country context stated undated or unverified figures as current. A 2025 exchange rate was labelled current, and the policy rate was a superseded 2025 range. All three country skills restated tax rates as plain facts, while the accounting engine's register holds them as draft.
  • Fix: the Uganda rows are now dated. The Central Bank Rate reads 12.0% from 14 October 2026 with its Bank of Uganda source, and the planning rate is labelled a 2025 assumption (c392a40). Each taxation section carries a dated callout that no rate is verified (24bd64b).
  • Limits: no tax rate was changed or verified. The Bank of Uganda statement was read on 31 October 2026, first two PDF pages only. The register lookups behind the callout are from the evidence (04-validation-record.md). Kenya and Tanzania rates were not refreshed.
• The financial-projection reference promised more than the workbook models. It called the workbook mandatory and listed a balance sheet and debt flows that the workbook lacks. The workbook's Adjusted EBITDA adds back depreciation and share-based pay that its cost lines never deducted.
  • Documented, not fixed (01997bd): a scope table in the reference, a decision row and an anti-pattern in the financial-projections skill, an instruction to quote operating income, and recalculation routed from the orchestrator, delivery contract and router.
  • Skill validators and the unit tests passed at that commit (from the evidence). Both pass on the current tree (re-measured). The workbook is still not a lender model.
• The plugin manifest exposed 134 of 137 skills. The three country-context skills were missing from the Claude-compatible manifest, and no registered validator checked the package-size limit.
  • Fix: 22a99c6 added the three paths and ran a fixed 5,000-entry cap inside validate_skill_engine.py (checker added in eb8d462, test runner order fixed in 58817d0).
  • Re-measured: plugin.json lists 134 skill paths before 22a99c6 and 137 after. node scripts/generate-plugin-manifest.js --engine . --root . --check exits 0 on the current tree. On 31 October 2026 the cap counted 1,681 files in the checkout and 1,169 in a package-oriented subset, not a built archive. The checkout count includes .git, so it rises with every commit.
  • Isolated Claude and Codex marketplace installs in containers ran on 29 October (from the evidence) and were not repeated. Public-directory acceptance is not assessed.
• Paths and bytes that resolved nowhere. The router (8 lines), one skill and the model README named Windows host paths (10 lines in Markdown, plus one in .gitignore). All seven workbook generators wrote to folders that no longer exist after the skills moved under skills/pipeline/. One skill pointed at a wrong sibling-engine path. One retail reference held a stray control byte.
  • Fix: host paths were removed and a host-path signature was added to the text check. The generator paths were repointed and a static test added (5fc69b9). The sibling path was corrected (02395c1). The byte was removed with a focused test (970d5ad).
  • Re-measured: host-path findings 10 to 0. tests/test_generator_paths.py on the pre-fix tree fails for all 7 generators and passes on the current tree. Control bytes in the retail file 1 to 0.
  • Limit: the generator test reads target paths and does not run the builders, which need xlsxwriter.

Kept changes: routing and documentation

Routing and discoverability

Routing is measured at first place, and an unreadable skill can no longer be skipped silently. Only top-three recall was measured. A skill whose frontmatter failed to parse was dropped without a message. This was found when a description containing a colon and a space broke the YAML and the score fell to 78 of 80.

• Fix: routing_smoke_test.py --report now prints precision@1, fixture coverage and owned negatives. An unreadable skill fails and names the file. Two descriptions were made specific (00-plan-assembly, 10-financial-projections). 19 fixtures were added (c7d956a, 33617d4). The CI workflow passes a precision@1 floor of 0.80 (015c68b; not observed on a runner).
• Re-measured, before (at 50d50f9) and now:

• The 1.3-point rise is not a like-for-like gain: the 61 original fixtures run against the current skills still score 48 of 61.
• The measure is a lexical scorer, not live-agent routing. The floor equals the measured value, so there is no headroom. 46 of 137 skills still share a template description.

Documentation and portability

• The documented install command failed and the check list was incomplete. The README told Linux users to run ./install.sh, which fails with "Permission denied" because the file is committed without its executable bit (mode 100644, re-measured). It listed four checks.
  • Fix (8ac52d2, ee81598): the README now uses bash install.sh. It lists every check with its 31 October result and the dependency it needs (PyYAML, LibreOffice, xlsxwriter, Node.js). AGENTS.md now uses python3 in place of python, states the CI routing floor, adds the link and text-integrity checks, and carries a known-limits note for the projection workbook and the overdue rows.
  • Re-run for this section: every command in the Checks block exits 0 except the evidence-register check, which exits 1 (four rows overdue). The PowerShell form and the plugin install commands were not run.
• Contributor and template documents had stale commands and paths. CONTRIBUTING.md had 8 lines with backslash paths or PowerShell cmdlets, and the release-gate document had 1. Both now have none (re-measured by pattern search).
  • The model README listed 17 repository paths in backticks, of which 9 did not exist. It now lists 18 and all exist (ad hoc check, not a registered validator).
  • Its claim that the workbooks were usable for lender presentations was replaced by a pointer to the scope table and the Adjusted EBITDA defect (e69cff2, 9756062).
• No index of the records under docs/. docs/README.md now lists the kaizen records, registers, contracts and dated history. It says that the formula-map PASS is structural only (8c446ee). 07-docs-followups.md records the configuration, CI and skill-text items that the documentation wave found but did not change (0a2360e).

How the kaizen was measured, and what was reverted

• Each cycle froze its baseline before changing anything (1b5fb14, 0d1eb51) and recorded Five Whys for each critical defect (9a4730a, 00afcec, f05f1b3). Each new check was shown to fail on a built-to-fail input before it was kept (the table in the pass 2 validation record).
• Unit tests: 66 at d3d5013 (before the 29 October cycle), 80 at 50d50f9, 130 now. All were re-measured with python3 -m unittest discover.
• Two attempts at the 00-plan-assembly description were reverted before commit. The first lost the words "plan assembly" and regressed a fixture. The second broke the YAML, which exposed the silent-skip defect above.
• A Codex-specific plugin override added in eb8d462 was removed in 22a99c6, because it still exposed 134 skills. The shared Claude-compatible manifest exposes all 137.

What is still open

Open findings that affect use

• The shipped SaaS projection workbook is defective for lender use (P0). It has no balance sheet and no debt schedule. Its Adjusted EBITDA adds back items that were never deducted.
  • Measured on 31 October 2026: formula_map.py --require-wired-inputs reports 10 of 104 numeric inputs that no formula uses. recalc_check.py --require-fresh-cache finds that 1,772 of 1,852 formula cells store a value that differs from a recalculation.
  • The fix is to rebuild scripts/build-financial-models/build_financial_projection.py, with an accounting review of tax, depreciation and share-based pay. It waits on a maintainer decision to install xlsxwriter and on that review.
  • Until then, quote operating income rather than Adjusted EBITDA, and follow the lender-case procedure in the workbook's reference.
• The release gate does not yet require the recalculation report (P0). validate_release_bundle.py can mark a spreadsheet handoff passing on the literal-input contract alone. The four example packs stay blocked.
• Four register rows are stale. They are UBOS, KNBS, UNBS and the Bank of Uganda supervision page, whose issuer pages could not be read. So refresh_evidence_register.py --check exits 1. Kenya and Tanzania policy and exchange rates were not refreshed.
  • No tax rate is verified. The accounting engine's register holds the Uganda and Kenya VAT, PAYE, income-tax and withholding entries as draft or skeleton entries, and has none for Tanzania.
  • Treat every rate in country-context/ as a dated figure to confirm before use.
• Routing is coarse for 46 of 137 skills. They share one template description, so routing between them rests on the skill name. Precision@1 is 80.0%, and a CI ratchet holds it there. 75 SKILL.md files repeat a heading.
• No complete plan has been rendered, reviewed for accessibility, or reviewed by a person other than the author.

Decisions for the maintainer

Three decisions wait on the maintainer:

1. Whether xlsxwriter may be installed.
2. Whether the em dashes and symbols stripped in April 2026 should be restored.
3. Whether the Uganda context keeps a fixed UGX planning rate.

Date-triggered gates

tools/evidence-register/refresh_evidence_register.py --check fails when a register row's recheck_due date is earlier than today. It already exits 1, and it fails on more rows on the dates below.

The script has no --as-of option. The counts were produced by calling its validate() function with each future date as today, which is the logic the command runs. A row is renewed by reading the issuer page and recording what it shows. Retrieving a URL renews nothing.

The six rows renewed on 31 October 2026 are TZ-NBS-STATISTICAL-ABSTRACT, UG-URA-TAX-GUIDANCE, UG-URSB-BUSINESS-REGISTRATION, UG-NEMA-REGULATIONS, UG-UCC-LICENSING and GLOBAL-WORLD-BANK-DATA.

The planned re-audit is 30 November 2026, the day the Bank of Uganda policy-rate row falls due. No other check in the repository reads the date (searched scripts/, tools/, tests/ and .github/). tools/scrape_metrics.py only stamps its output.

Cited sources

The sources this repository cites as the basis for its skills are listed in docs/cited-sources.md. The list holds 200 books, 11 repositories, 14 groups of standards and official sources, and 26 website and article entries.

Books that appear only in reading or wish lists (for example 05-reading-list.md), and titles supplied without content, are not listed. These are citations only. No book content is reproduced.

Contributing, author and licence

• Contributing: read CONTRIBUTING.md before you add or change a skill. It covers authoring a skill and validating a change.
• Author: Peter Bamuhigire.
• Licence: MIT. See LICENSE.