Home › Learn › The technical anchor › Project review
The technical anchor · an open-source Claude skill
Project review: what is your data project missing?
Give Claude this skill and your project documents. It tells you what is missing at the stage you are in, quotes where your documents answer each question, finds where two documents disagree, and gives you the cheapest test to run this week instead of finding out at go-live. It works for a new data platform or a project on an existing one, from discovery to sign-off, on any technology.
Get it on GitHub → Free and open source: code under Apache 2.0, checks and templates under CC BY 4.0.
Why would a team need this?
Why it exists
Data projects rarely fail because a developer couldn't write the code. They fail because something nobody owned slipped through: a file that arrives twice, a history load nobody estimated, a definition of revenue nobody signed, a plan that assumed source access in week one when approval takes six weeks. These failures are the same on every project, even though every project is different. So they can be written down once as a checklist, and a reviewer can look for them in any project's documents.
This skill is that reviewer. It comes as four skills: one each for business analysis, project management and technical work, which a BA, PM or engineer can run alone, and one full review that runs all three and checks the seams between them, where most projects fail. It is the technical anchor's habit of looking at the whole picture, and the fail-fast thinking of discovery, packaged so any team can run it on their own documents at any stage.
How does it work without knowing my project?
How it works
- Three layers keep it generic and still specific. Common checks ask questions every data project faces and never name a product. Each check says when it applies: the kind of work, the components (ingestion, transformation, modelling, serving, orchestration, platform, governance), the patterns (batch, streaming, change data capture, history load) and the lifecycle stage. A binding per technology says how each check is done there, and its traps. Today Databricks; another technology is one more file.
- Judged at your stage. Each check says by which of the ten lifecycle stages a plan, a design, evidence and proof are due. A team in discovery is asked for plans; a team in test is asked for evidence. Anything not yet due shows as "coming up", not as a gap.
- Three ways to spot a gap. Signals in the documents (an estimate that is one number with no assumptions, "near real-time" with no number), disagreements between documents (the plan and the risk log give different dates for source access), and silence, which becomes a question rather than an assumption.
- Scripts where judgement isn't needed. Picking the checks, scanning for mechanical signals and building the report are done by small scripts, so the same project always gets the same checklist. Claude does the reading, quoting and comparing.
What does it look for?
What it checks
| Group | Covers |
|---|---|
| Business analysis (43 checks) | the current process and the number the business trusts today, data requirements (sources, grain, history, freshness, late data), metric definitions, business rules, data ownership, consumers, policy needs, scope, acceptance and user testing |
| Project management (64 checks) | plan and dependencies, budget and spend, resourcing, vendors and source teams, governance and stage gates, risks and decisions, stakeholders, communication, history load and cut-over on the plan, go-live readiness, benefits |
| Technical (91 checks) | duplicates and reruns, first load and history load, late and corrected data, deletes, history over time, schema change, data quality, reconciliation, performance, cost, personal data, environments, release and git, orchestration, operations, testing, design, serving, the effort estimate |
| Databricks (12 checks + 57 how-tos) | how each common check is done on Databricks, its traps, and Databricks-only checks such as Unity Catalog ownership, compute policies, Auto Loader schema evolution and safe job repairs |
See every check, by category →
Every check lives in the pack of the discipline that maintains it, and has an owner role (technical anchor, architect, business analyst, project manager, developer, test lead, release, operations, security and others), so the report shows each person their part. Ten templates help fill what is missing: discovery document, estimate, scope, risk log, decision record, test strategy, release and cut-over, stakeholder map, status report, and how to run a pre-mortem.
What do I get back?
What you get back
One report, with the most important part first: a fail-fast plan (each risk with the cheapest test now and the result that would change the plan), a pre-mortem (the most likely ways this project fails, from the evidence), the places where your documents disagree, the 15 most serious gaps, then everything else by concern and by role, and what you are doing well. Three rows from a real test run:
| Risk | Cheapest test now | Result that would change the plan |
|---|---|---|
| Revenue on the dashboard won't match the board figure | Compute last week's revenue both ways and put both next to the finance pack | Off by more than finance's tolerance: fix the gold model before sprint 4 |
| A resent file doubles a day's online sales | Drop the same file twice in the development environment and compare totals | Totals change: replace by day instead of appending |
| Three years of history has no load plan | Time one month of history over the real connection | Scaled up, it doesn't fit the source owner's window: re-scope now |
From the review of the invented sample project described below.
How do I run it?
How to use it
- Claude Code: install the plugin with
/plugin marketplace add dataagentsai/data-project-reviewthen/plugin install data-project-review@data-project-review, and ask: "Review our project documents in ./docs with data-project-review", or data-review-ba, -pm or -tech for one discipline. - claude.ai: upload one of the skill zips built by
scripts/package.py, then attach your documents and ask the same. - Your documents can be Markdown, Word or PDF. If you don't write a short project profile, Claude drafts one from your documents and asks only what they don't say, usually the stage you are in.
Review documents, not people: the report says "not mentioned" when a document is silent, which is a question to ask, not a verdict.
How do we know it finds real gaps?
How we know it works
We wrote an invented retailer's project, twelve realistic documents, and hid 27 gaps in them, some by leaving things out, some by vague words, some only visible when two documents disagree, and two never mentioned at all. We also put in five things done well. A fresh Claude session reviewed it using only the skill, without seeing the answer key.
| Measure | Result |
|---|---|
| Planted gaps found | 27 of 27 (19 clearly marked missing or contradicting, 8 marked partly covered or not mentioned) |
| Gaps hidden by leaving them out | 19 of 19 |
| Gaps only visible when two documents disagree | 4 of 4 |
| Vague wording | 2 of 2 |
| Never mentioned anywhere | 2 of 2 |
| Good work wrongly flagged as a gap | 0 of 5 |
It also raised 14 serious gaps we hadn't planted, such as "never re-estimated against actual progress", which read as real holes in the invented documents. After the run we merged three overlapping checks, shortened the report and fixed two detectors. The full results are in the repository. Next: running it on real projects.
Next
Where to go next
The thinking behind it: the technical anchor and discovery, step by step. The data problems it checks for, recorded on real Spark and Delta: the Problem Lab.