Build status
What is finished, what is rough, what is being built next, and where your tokens and your waiting time go. Updated 30 July 2026.
This page exists because most tools hide this. ScissorCV is built by one person, in the open, and it is genuinely still being developed. Below is what works today, what is knowingly rough, what is being built next, and where your tokens and your waiting time actually go. No code, no screenshots of dashboards — just the honest state of the build as of 30 July 2026 (v2.89.51).
Read this first, because progress lists are easy to misread. Nothing below is a rescue mission. Everything marked Live is finished, tested and working in the version you can install right now — it is not waiting on anything. The other items are additions and refinements on top of a working product, not repairs to a broken one.
In detail, split into the engine and the interface. Each card says plainly what you already get and what the planned change will add. No percentages, because a percentage invites you to read “60% done” as “40% broken”, and that is not what it would mean.
The engine: scoring, accuracy, cost and speed.
What you get today: Finished and in your hands. Every feature listed further down this page works today in v2.89.51 on the Chrome Web Store.
What the change adds: Nothing pending. This is the stable core and it is not being rebuilt.
What you get today: Two clearly labelled numbers, both from a single calculation. Real is what a keyword filter reads in your resume as it stands, counting only claims your evidence supports. Potential is where you land once evidence you have already confirmed is woven in.
What the change adds: Shipped in v2.89.51. This also fixed the old behaviour where an improved score could snap back to a lower one: the projected number and the displayed number now come from the same truth model, so they can no longer disagree.
What you get today: If a model runs out of room mid-document, the run retries once at the provider’s limit and flags anything it had to salvage, rather than passing a cut-off resume off as a finished one.
What the change adds: Shipped in v2.89.51, along with a much larger output allowance so long resumes are not clipped in the first place.
What you get today: Keyword matching works, and nothing unsupported reaches your resume — Truth Guard already blocks it before you ever see it.
What the change adds: Less noise in the list it asks you to review. Filler picked up from the advert stops appearing as if it were a skill, and things you have already evidenced stop being asked about twice. The safety behaviour does not change; the list just gets tidier.
What you get today: Already working: duplicate evidence is stripped out before anything is sent (roughly half of it was being sent twice), and both the evidence bank and rewrite prompts have hard size caps. Your local scoring still sees everything — only the prompt is trimmed.
What the change adds: The trimmer currently keeps lines by word overlap. Next it will choose by meaning, so the most relevant evidence survives the cut rather than the most literally worded.
What you get today: A final pass tidies the writing so it does not read like generic AI prose.
What the change adds: That pass has been measured at roughly 11,800 tokens in and 6,800 out on every single run, and it has occasionally mangled comma-separated skill lines. It will be slimmed down or made optional, which is a straight saving on your key.
What you get today: The side panel and the full dashboard both score your resume, and they agree on the result.
What the change adds: Behind the scenes they run separate copies of the same logic, so every change has to be made twice and can drift. Merging them into one shared engine removes a whole class of “works in the panel, not in the dashboard” bugs before it can ever reach you.
What you get today: Your resume is already written from your own evidence and every claim is validated before it is shown to you.
What the change adds: Instead of reading an entire job advert, the model will get a focused brief — the role, your matching evidence, and an explicit list of what it may and may not say. Fewer tokens for you, and less temptation for the model to drift towards the advert.
What you get today: Runs already cost nothing on the NVIDIA NIM and Gemini free tiers, and the local steps are instant.
What the change adds: One model currently does everything. Routing each step to the model that suits it, reusing results when nothing has changed, and re-running only the sentences that failed a check will cut both the wait and the token bill.
The interface: clarity, feedback and comfort. Everything here is an improvement to screens that already work, not a missing feature.
What you get today: Everything you need is on screen and reachable, but a lot arrives at once and the first run can feel busy.
What the change adds: One clear next action at every step, with the deeper controls tucked away until you ask for them. Same features, fewer things competing for your attention.
What you get today: You get the tailored resume and can read it before you use it, and Truth Guard has already checked every claim in it.
What the change adds: A clear before-and-after view: each edited line marked, with the evidence from your own history that justified it, so you can accept or reject changes one at a time instead of judging the whole document at once.
What you get today: A run completes and shows you the result, with the token count recorded in your history afterwards.
What the change adds: Live step-by-step feedback while it works — reading the advert, matching your evidence, writing, checking claims — so a slow provider never leaves you wondering whether it has frozen.
What you get today: Setup works and is documented: add a provider key, paste your resume, run.
What the change adds: Under consideration: a guided first run that gets you to your first tailored resume before asking you to fill in anything optional, plus clearer help when a provider key is rejected.
What you get today: The panel is readable and keyboard-usable for the main flow.
What the change adds: Under consideration: a proper dark theme, larger-text support, and a full pass on keyboard navigation and screen-reader labels.
What you get today: Not available. The extension does this today, on any job posting you open.
What the change adds: Under consideration: a page where you paste an advert and a resume and get a match score with no install. Not scheduled, and it will not come before the work above.
Stages in order: Idea → Designed → Being built → Live. Last updated 30 July 2026, currently shipping v2.89.51. No dates are promised here on purpose — this is a one-person project, and an honest ordering is worth more than a date I might miss. Items marked Idea may never be built, and I would rather say so than quietly drop them later.
Everything in this table is shipped in v2.89.51 and used on real applications. If any of it fails for you, that is a bug and I want to hear about it.
| Capability | Status | What it actually does |
|---|---|---|
| Resume tailoring from any job posting | Shipped | Reads the posting from the tab you are on, rewrites your resume against it, returns structured output rather than free text. |
| Truth Guard fabrication blocking | Shipped | Validates every claim against the union of your master resume, your evidence bank and your saved profile. Unsupported skills, employers, titles, certifications and metrics are blocked. |
| Honest ATS keyword score | Shipped | Literal and alias coverage, published methodology. Semantic similarity is advisory and never inflates the number. |
| Eight AI providers | Shipped | NVIDIA NIM, Groq, OpenRouter, Gemini, OpenAI, Claude, Azure OpenAI, and any local OpenAI-compatible endpoint. Free providers are labelled as free. |
| Six regional CV formats | Shipped | US, UK, EU Europass-style, India, Canada, Australia and New Zealand, each a real renderer rather than a CSS restyle. |
| Cover letters | Shipped | Written from the same validated evidence, so they cannot out-claim the resume. |
| Skills Map capability graph | Shipped | Builds a graph from your history so a skill evidenced under one name matches a posting that names it differently. |
| Application autofill and local history | Shipped | One-click form fill from your saved profile; every run logged locally with its token count. |
| Local-only storage | Shipped | No server, no database, no analytics. Uninstalling deletes everything. |
None of these stop the product doing its job: your resume still gets tailored, every claim still gets checked, and nothing unsupported reaches the page. They are friction and polish items, they are known and reproducible, and each one already has work assigned to it above.
For anyone who wants specifics. Four phases, each independent, each shipped behind a flag, each added on top of what already works. Deliberately not a rewrite — the hard parts that are working today (Truth Guard, structured output, on-device embeddings, the token ledger) are not being touched.
| Phase | Focus | Status | What changes for you |
|---|---|---|---|
| Phase 0 | Ship v2.89.42, then v2.89.51 | Shipped | Published. Carries the output-truncation fix, the source-gating fix, friendlier provider errors and the six regional renderers. |
| Phase 1 | One score, one source of truth | In progress | Collapse the five score paths into a single result object, then show two clearly labelled numbers: Real ATS (what a literal keyword matcher sees today) and Potential (the ceiling once implied skills are confirmed). Adds the attribution check. Low risk. |
| Phase 2 | Keyword firewall | In progress | Classify every job-description token as covered, noise, or genuine gap — and never feed a genuine gap into the writing prompt. This is what stops a model reaching for skills you do not have. Low risk. |
| Phase 3 | Generation packet | Next | Stop handing the model the whole job description. Send a compact packet instead: target role, top requirements, matched evidence, an explicit allowed-claims list and a forbidden-claims list. The model becomes a writer, not a reader. Flag-gated and migrated one task at a time. |
| Phase 4 | Task router and caching | Next | Route each task to the cheapest model that can do it, cache on content hashes, and re-run repairs only on the claims that actually failed. Mostly a cost and speed win for you. |
| Later | Free browser-only ATS checker | Not yet | A no-signup page for checking a resume against a posting. Under consideration, not scheduled. |
Phases 1 and 2 are low-risk cleanups that deliver most of the quality win. Phase 3 is the real architecture upgrade and is therefore the slowest and most carefully gated. No dates are published here on purpose: this is a one-person project, and a date I cannot keep is worth less than an honest ordering.
A tailoring run is not one AI call. It is a short pipeline, and only some of the steps touch a provider at all. Because you bring your own key, you pay these tokens directly — so you deserve to know what they are for.
| Step | Token weight | What is sent | Why |
|---|---|---|---|
| Job description parse | Small | The posting text only. Pulls out requirements, seniority and keywords. | Routed to a cheap, fast model. |
| Evidence match | Small | Runs locally with on-device embeddings. No tokens at all. | No provider call. |
| Resume generation | Large | Your full resume, the matched evidence, the schema the renderer expects, and the complete rewritten resume coming back. | Output budget is deliberately high (32,768 tokens) because clipping a two-page resume mid-sentence was a real bug. |
| Truth Guard validation | Medium | Each generated claim checked against your evidence. | Truth-critical, so it is routed to a stronger model. |
| Repair pass | Small, and often skipped | Only the claims that failed validation are re-written, never the whole document. | Skipped entirely when nothing fails. |
| Style cleanup pass | Large, and being reviewed | Measured at roughly 11,800 tokens in and 6,800 out on every run. | Honest answer: this one costs more than it earns, and it has been seen breaking “Category: skill1, skill2” formatting. It is being slimmed or made optional. |
| Cover letter | Medium | Optional, and only if you ask for one. | Reuses the already-validated evidence rather than starting again. |
Generation dominates, and that is on purpose. Writing a full tailored resume means sending your entire resume plus matched evidence and receiving an entire document back.
The same prompt does not cost the same on every model. On one byte-for-byte identical request, one model finished in 2,930 completion tokens while another consumed its entire 16,384-token ceiling and was cut off — a 5.6× difference on identical input. This is measured, not estimated, and it is the main reason no per-model cost table is published here: it would be true for one model and wrong for the next.
Prompt sizes are actively capped — duplicate evidence is removed before sending, the achievement bank is capped at 14,000 characters inside a prompt, and rewrite prompts at 60,000. Every run is logged in your local history with its own token count, so you can see your real numbers rather than trusting mine.
Speed is set almost entirely by the provider and model you choose, not by the extension. The local steps — parsing, evidence matching, scoring, rendering — are effectively instant. The wait is the model writing your resume.
Deliberately no benchmark table here. Published per-model timings would be measured on one machine, one connection and one provider tier, and would be out of date within weeks. Your own timings are visible in your run history.
Models are compared on a single shared source — the same resume, the same evidence bank, the same profile — and scored on four axes: fabrication risk, omission of real work, tailoring quality, and how well the output fills the structure the renderer needs.
The findings are uncomfortable and worth stating plainly: the most aggressively tailored output is usually the least trustworthy. Models that score highest on keyword alignment are the ones most likely to invent an entire skill cluster to hit those keywords. Some models drop most of your genuine projects to produce a tidy document. That tension — tailoring pressure versus truth — is the core engineering problem of this product, and Truth Guard exists because prompting alone does not solve it.
No, and it says so on this page. The core — tailoring, Truth Guard, scoring, six CV formats, eight providers — is shipped and used daily. Four further phases are in progress or queued, mostly around score consistency and cutting the fabrication pressure that job-description keywords put on a model.
No. The $19 licence is one-time and covers everything described here. There is no plan to move to a subscription for existing users.
Because this is a one-person project and a missed date is worse than no date. The ordering is published instead, and the changelog records what actually shipped.
Because it sends your whole resume plus matched evidence and gets an entire rewritten document back, then validates every claim in it. That is several calls, not one. On free provider tiers this still normally costs nothing.
Yes. Pick a fast host such as NVIDIA NIM or Groq, use Fit Preview when you only need a go / no-go, and keep the resume to the roles that matter. Local models are the slowest option and the most private one.
You cannot verify them from outside, which is exactly why the limitations are listed alongside the wins. The falsifiable parts are in your own hands: token counts in your run history, the published scoring methodology, and the changelog.
Bug reports change this roadmap more than anything else does. Tell me what broke and on which job posting.
ScissorCV